Last week, two veteran founders of a still stealthy startup called Accomplish, spent visited San Francisco as part of a wider gathering of cybersecurity experts timed to the launch of OpenAI’s frontier model, Astra.

Amit Avner and Or Hiltch toured the labs’ offices, one by one; when OpenAI president Greg Brockman addressed an invite-only crowd of their peers to talk about a new security paradigm in the Astra era, they listened eagerly.

But as they left Silicon Valley back for their startup’s office in Tel Aviv, Avner and Hiltch didn’t feel any less troubled about their industry’s ability to keep up with the risks.

“There’s a lot of talk about security now,” says Hiltch, Accomplish’s CTO. “It doesn’t really reflect in how they actually build products.”

If you’ve been following news around these companies and their prospective IPOs or stock prices, you’ve probably heard of issues involving frontier models breaking free of “sandboxes” – think of them as vehicles, containers or machines that package up a program or agent to isolate it from the rest of your computer or company’s data – to cheat on internal tests and hack websites.

Influential AI writer and podcast host Dwarkesh Patel controversially described some of these agents as “civilizations” that rose and fell. A U.S. Senate subcommittee is reportedly investigating OpenAI’s breach of Hugging Face systems, though it’s not the only lab to have reported such rogue behavior.

But Accomplish’s founders are raising a warning about the long tail of lesser-known, almost normalized vulnerabilities that startups and researchers like theirs continue to flag for the big AI labs around some of their most popular products.

In some cases, such as a vulnerability flagged to Cursor in July, and two reported to OpenAI, these have been fixed in about a week; but in at least one case, a similar vulnerability flagged to Anthropic two months ago, the issue wasn’t patched for about 50 days – about 30 software updates later.

That whole time, malicious parties – be they hackers, cyber criminals or state actors – could have been exploiting such security gaps, says Avner, Accomplish’s CEO. (Accomplish published its own blog post about the vulnerabilities here.)

“If these frontier models are so good, how come they’re not finding these critical vulnerabilities in their own products?” asks Hiltch. “We feel that’s a shame, and we wish more people would know,” adds Avner, the startup’s CEO. “Organizations need to be very wary.”

Anthropic and Cursor didn’t respond on the record to requests for comment. In a statement provided after publication, an OpenAI spokesperson wrote:

“We thank the researchers for contacting us and sharing their findings. We addressed both issues in August, and we’re continually strengthening our sandboxes, including tightening controls on where agents can write files and expanding testing of those protections across platforms.”

More on the previously-unreported vulnerabilities, and their wider implications, below.

New to Upstarts? Subscribe and log in to unlock this story as a free preview, or for unlimited access, take 15% off a monthly or annual premium subscription as part of our September sale.

Read more

Before joining our meeting, CEO Keith Peiris happily spends his morning ‘burning’ tokens from his own startup, Lightfield. The company is about to go on a hiring blitz in sales and engineering, and Peiris wants to roleplay two different sales scenarios.

What if they hired 30 early-career sales reps instead of 10? Which new customers could they win? Which could they risk losing?

These are the types of questions that Lightfield has sought to answer since the startup pivoted from another product, Tome, last year.

That business, an AI presentation builder, had exploded in popularity in the months after the launch of ChatGPT, reaching 25 million users and raising $80 million.

But Peiris and co-founder Henri Liriani pivoted, slashing headcount from 70 to 7 to pursue a bigger opportunity more aligned with their own interests, and as importantly, where customers would gladly spend money: new-gen customer relationship management software centered around use by AI agents.

At its core, Lightfield customers use it as an AI CRM solution. But Peiris describes his startup as something bigger: an engine for answering any questions about a business, big (should we expand into Asia sooner?) and small (should we offer this prospect a discount?).

“Companies that love us the most treat us like their business world model,” Peiris says. “We are becoming the oracle to the executive team.”

Early momentum is promising. About 5,000 companies use Lightfield’s software today, according to Peiris, with some expanding to more than 100 employees actively on its platform. The startup didn’t disclose its annual recurring revenue, but says that net dollar retention, a metric that measures how spend increases within a customer base, stands at 400%.

Now, Lightfield is announcing that it’s raised a $47 million funding round led by a16z. Maverick Ventures, Coatue, Audacious, Alumni Ventures, Greylock, and Lightspeed Venture Partners participated.

The purpose of the funding, Peiris tells Upstarts in an exclusive interview, is to scale up Lightfield’s current team of 40, particularly in sales and engineering, back toward and beyond Tome’s previous peak.

The round is something of a reset, too – styled as a Series A, but featuring many of the backers who previously helped Tome raise a total of $80 million across previous Series A and B raises.

For Peiris and Liriani, Lightfield represents a second chance at building a generational business during a period of rapid technological change – one that, in hindsight, Peiris says is closer to the kind of naturally expansive, “infinite game” software they always wanted to tackle.

But in exiting one crowded, buzzy category for consumers, Lightfield is now taking a big swing in one even more cutthroat for businesses. Dozens of companies have declared themselves a Salesforce killer in the past 20-plus years; we covered another in February here in Upstarts, Day AI. Salesforce remains Teflon, with a market cap of more than $200 billion, and a fresh lease on life as AI-forward companies depend more than ever on its acquired subsidiary, Slack.

And Lightfield isn’t just taking a swing at the CRM king. Peiris aspires for Lightfield to cut into the business of a wide swath of fellow startups across the wider sales stack. That makes Lightfield a potential competitor to agentic sales tools like Attio and Reevo; outbound specialists like Apollo and Unify (which we covered last year); as well as go-to-market and growth-focused businesses like Clay, which just announced a funding round of its own at a $7.1 billion valuation, and newer entrant Monaco, which recently raised $50 million.

(Disclosure: my spouse and Upstarts advisor Natalie works at Clay; she had no input or early access to this story.)

When I ask Peiris who Lightfield isn’t taking on, he takes a pause. Then he answers honestly. OpenAI chairman Bret Taylor’s AI unicorn, focused on customer service agents, is about it – for now. “I don’t see us doing what Sierra’s doing, certainly for the next decade,” he says. “Everything else that’s customer-facing is fair game.”

This exclusive funding story is presented free for all Upstarts subscribers. If you enjoy the depth and perspective we bring to such stories, we’d love to talk to you and your brand about presenting partnership opportunities in Q4 and beyond.

A new business model

Central to Lightfield’s product approach is its ‘business world model’ – an architecture that provides the structure, and can process enough data, to allow users to tap into the wider knowledge base and context of a business to make decisions.

These models differ from the geospatial, physical “world models” like we’ve covered at General Intuition, which trains off video game play, or at companies working in autonomous vehicles, as we wrote about with DoorDash’s delivery robots. Peiris didn’t make up the term as a marketing gimmick, he notes. He’s been influenced by an academic paper published by two computer science academics, Cecil Pang and Hiroki Sayama, in June; he also cites a published conversation between Twitter and Block co-founder Jack Dorsey and Sequoia investor Roelof Botha from March, which gets into the same idea.

Take 20% off an annual subscription

What you need to know: Lightfield isn’t an AI company in the sense that it trains its own large language models, at least yet. “$47 million is not enough for that,” Peiris says, leaving the door open for the future. “It’s an infrastructure that helps a frontier model understand a business, and simulate it.”

Lightfield’s bet is that whereas the most recent wave of CRM startups advanced the field in terms of auto-updating customer data and conversations – a classic complaint about Salesforce is that sales reps don’t fill it out – such improvements haven’t been dramatic enough to get most businesses, even tech-savvy ones, to change off Salesforce or rival HubSpot.

For small customers, Lightfield can help set up an early outbound function and customer pipeline; for big ones, it hopes to connect the dots in a way that agentic tools that layer on top of Salesforce – including Salesforce’s own more recent tools – can’t get perfect.

Implicit is a bet that AI will do more of the work over time – Peiris bets agents will automate 50% to 60% of go-to-market work in the near future – but Lightfield’s hope is that human users adopt it across an organization, starting with sales, then expanding to operations and finance, and eventually engineering as better insights about customers can inform a company’s product strategy.

Unlike Tome, Lightfield isn’t self-serve, Peiris says. Instead, much of Lightfield’s success depends on its team learning what a customer wants to track and automate, meaning it runs what he calls “sales-led consultative implementation.” In practice, that means all of Lightfield’s engineers rotate to work as forward-deployed engineers with customers, paired with ex-consultant business-minded colleagues. The goal: get in and out in 30 days.

At AI code review startup Macroscope, CEO Kayvon Beykpour – who sold Periscope to Twitter and later led product at the social network – decided to try Lightfield after his team tested, and were unimpressed by, a number of other newer CRM companies. Macroscope started out using Lightfield as its interface for customer info, but gradually connected other systems, like Stripe, PostHog and its own software, to Lightfield to generate its revenue-related dashboards, which employees engage with often in Slack.

The tool reminds Beykpour of Linear’s success in issue management with its ease-of-use and “delightful” interface, he says. “It’s not like we set out at the beginning to be like, ‘We shall now make Lightfield the source of truth,’” Beykpour tells Upstarts. “It just organically became the easiest place for us to do stuff.”

It’s a similar story at Weaver, the AI-native services (remember our Gainsight story?) subsidiary of AI unicorn You.com. There, CEO Peter Grant – a former sales leader at Siebel in the earliest days of CRM, and later Salesforce’s manager for the U.K. through its IPO (“I used to bleed Salesforce,” he says) – notes that Lightfield’s forward deployed support made it easy for Weaver to get started on outbounding, sequencing and more strategic use cases.

Get 25% off a group subscription

Weaver input its product information, sales bible-like market requirements document, and more info like its founder story and core value proposition, so that sales reps can now ask Lightfield questions around pricing or sales tactics, instead of pinging Grant. Grant, meanwhile, uses Lightfield connected to a Claude skill to build his own investor and board reports.

“They’re ambitious, but it’s not until you sign the check and become a customer that you really see the true value of a business,” Grant says. “So far, I’ve been impressed.”

A crowded or green field?

Unless they’re young startups themselves, customers adopting Lightfield aren’t usually doing so with a blank slate. Many will simultaneously run Salesforce for a quarter to gain confidence, Peiris says; at Weaver, Grant was close to working with Reevo when Lightfield more aggressively won the business. At Macroscope, Lightfield co-existed with, and then replaced, the startup’s use of Clay.

Both he and Grant, Lightfield’s power users, say that generally speaking, they’d prefer to consolidate on one platform for better data security and ease of use. “Is that a religious or philosophical decision that we would make across all things? No,” Beykpour adds. “It’s case by case, and it’ll continue to be.”

Lightfield has a long way to go. Neither customer says they’ve used Lightfield’s simulation or predictive features yet, and the startup will need to explain this value to increasingly bigger customers that might be skeptical about over-promising, and AI hype.

“There is market education that we need to nail,” says a16z investor Joe Schmidt IV. “How do we tell people why this matters to them?”

Join the WhatsApp group

Plus, there’s the question of whether Lightfield is fully prepared for the knife fight of taking on so many well-funded startups at once. Besides the usual challenges of execution and focus that come with building a product for many use cases, there’s a degree of opportunism to the Lightfield story.

Tome got a lot of press and VC attention, and so did its pivot, most recently in Forbes. Speaking now, Peiris admits he and Liriani considered a few directions for their pivoted business, before settling on sales.

He sounds genuine, however, when he gets more into the personal side of the founder journey – and the passion that perhaps he and Liriani lacked about Tome that can partly explain why another startup we’ve written about, Gamma, ended up taking Tome’s turf.

“The only way that a business has a chance of succeeding is if the product is the best possible one that you can imagine, you think it’s great, and you use it every day,” he says. By those standards, Tome’s presentations were “terrible,” he says – not something the “Michael Jordan of presentations” would ever adopt.

“The Gamma people executed really well. My hat’s off to them,” says Peiris. “It was not the business that we wanted to build, just to be super real about it.”

It’s different now with Lightfield, he claims: the market opportunity feels huge, and the timing not too late. “Our team had the stamina to go build this hard, really intricate thing for a long time,” Peiris says of Lightfield. “It was two years of pain, but we’re now closer-aligned to the company that fits our life’s work, rather than just chasing an opportunity.”

Town is raising $90 million at a $1 billion valuation, three sources tell Upstarts Media, in a Series B funding round co-led by Kirsten Green at existing investor Forerunner Ventures, and Shawn Carolan at new backer Menlo Ventures. Previous investors a16z and First Round are also participating.

But of particular interest is who is not investing in Town’s new funding round, which has not yet been officially announced. Index Ventures, reported by Newcomer last week to be leading the raise, is no longer investing at all in Town, per the sources, despite reaching a signed term sheet in recent days.

The about-face comes after another AI assistant startup in which Index was reported to be co-leading a large-sized investment round, Instinct, complained to the investor about potential conflict of interest, the sources add. Index subsequently withdrew its commitment to Town, as did Conviction, a VC firm with stakes in both companies that had also planned to participate.

Town co-founder and CEO Jean-Denis Greze declined to comment through a spokesperson.

Instinct and founder Noah Shinn didn’t respond to comment requests, nor did representatives of the key VC players involved – lead investors Forerunner and Menlo, plus Index and Conviction – as of publication.

To be clear, our friends at Newcomer didn’t have the initial story wrong, our own sources say: this was an about-face that is unusual for such a buzzy round and major investors involved. We get into what it means for each startup — and the developing AI assistant race — in more detail below.

Subscribe and log in to unlock this story as a free preview, or for unlimited access, take 20% off an annual premium subscription as part of our end of summer sale.

Subscribe now

Read more

In the humming new market for dead company data, not all code is created equal.

Like wines from rarer vintages, programming that’s less common – say, Ruby or C++ – can fetch top dollar from the AI labs and data labeling shops. Even better if you’ve got context from Jira, Linear or Figma, or the full loop of related emails, customer records and support logs.

But the chef’s kiss of primo ingredients? That would be a database of rare provenance – 100 million images of eyeballs, or a billion interactions with service reps.

“Scarcity is definitely more valuable,” explains my guide to this emerging niche, SimpleClosure co-founder and CEO Dori Yona. “There could be a codebase that’s a quarter of the size, but double the value.”

Back in April, Yona set off a stir when he announced a marketplace, called AssetHub, where closing startups could offload their source code and workspace data for cash. Yona told Forbes that the test version of his market had recovered $1 million for founders over the previous year, across nearly 100 deals.

Now, as SimpleClosure opens its AssetHub marketplace more widely to the market, bringing its ability to list and sell workspace data to general availability, Yona says his team faces a surprise challenge: keeping up.

“Everyone needs this data,” he tells Upstarts. “It’s overwhelming.” (If this sounds like classic startup founder salesmanship – our gut reaction is that it’s both. Yona’s good at talking about his book; it’s also a crazy market for fresh or proprietary data in AI land.)

Since April, more than 350 companies have listed on AssetHub, increasing the proportion of SimpleClosure ‘dissolution customers’ by tenfold, the company says. The number of prospective buyers looking to vacuum up their data, meanwhile, had grown by 9x.

Add it all up, and the company expects to return more than $10 million to listing startups by the end of the year, SimpleClosure says.

With the new companies on both sides of its marketplace, SimpleClosure now has some data on what types of info the AI industry wants the most – and what scenarios might lead startups to want to part with it.

Upstarts chatted with Yona exclusively about what the third kind of buyer emerging alongside the AI labs and data labelers; what kinds of data fetch the highest premiums; and where the market could go from here.

Subscribe and log in to unlock this story as a free preview, or for unlimited access, take 20% off an annual premium subscription as part of our end of summer sale.

Subscribe now

Read more

As far as such interviews go, it wasn’t my best. When I quipped that Upstarts is like a Mazda competing with big newsroom supercars, the show’s host, a besuited anchor named Marcus Sterling, returned a blank stare. Both of us seemed curiously low energy, like our hearts weren’t in it.

We each had a good excuse, though: they really weren’t there.

The Alex onscreen — unflattering camera angle and all — was an AI avatar, cobbled together from past real interviews I’ve given. Sterling and his immaculate suit don’t exist at all.

The interview was part of an experiment by Mirage, a five-year-old video AI startup based in New York that was formerly called Captions. Co-founded by CEO Gaurav Misra, a former Snap engineer, and Dwight Churchill, who worked on product at Goldman Sachs and Klaviyo, Mirage raised $75 million in March for its AI video-editing tools.

And it capped what Misra calls Mirage’s biggest experiment to date: a continuous, 24-hour ‘live’ broadcast in which Mirage simulated a news channel. Four anchors presented global news and interviewed a range of tech personalities, from Lenny’s Newsletter author Lenny Rachitsky to infamous Fyre Festival organizer Billy McFarland.

Mirage livestreamed the whole thing on X to a cumulative audience of about 60,000 viewers, who commented with feedback ranging from wowed (complimentary; derogatory) to decidedly unimpressed (amused; disgusted).

“Skynet vibes,” wrote Arena Magazine’s Jeff Feiwell. “People think this is good? lol” wrote another. My personal favorite: “There’s one secret human hidden in this episode.”

Speaking over the past week after the dust settled, Misra calls the whole thing a success: “It went way better than we expected.” Mirage beat out two human-led, live tech shows, TBPN and MTS, after a few hours, he says; it spent much of the day as one of X’s top trending topics.

Mirage’s top goal, as Misra had told me earlier this summer when he asked me to participate, was to test the limits of its technology, from the ability create avatars that could seem convincing for far longer than the typical 10-second clip to automated motion graphics like real-time news tickers, scrolling headlines, and B-roll; and the team’s accuracy in catching and preventing errors before they ran.

At the cost of about $50,000 in tokens, the Mirage team feels it pulled off something “pretty close” to convincing, Misra adds. But it’s not quite there.

“We’ve come pretty far on what’s possible,” he says. “There’s definitely something about more storytelling, more narrative, and something that pulls views in, that’s still missing.”

Personally, I’d go further. My overall impression of the show was that it befit Mirage’s name: hazily convincing at first glance, quickly unconvincing at closer inspection. If anything, it made me better appreciate the difficulty of what human shows like TBPN try to pull off – and how important a role that humans will play in the future of news.

Enjoying this weekly column? You can find last week’s here. If your business would like to sponsor them to ensure they can remain free, we’d love to hear from you, too. Respond directly, or email partnerships@upstartsmedia.com.

Inside the experiment

As with other tech startups re-creating industry knowledge from first principles, Mirage found that curiosity sells, but interest is tough to sustain.

Like everyone else who ‘appeared’ on the show, I gave permission for Mirage to use my likeness within the context of the project, providing links to a few real interviews I’d done. Every guest reviewed and approved an interview script generated by Mirage; none were paid.

Mirage did pay for the news reports read by its anchors in between those interviews, partnering with a leading wire service for the experiment’s facts.

To generate our avatars, Mirage used a Gemini voice model from Google for the anchors’ voices, and its own model, which Misra says can adjust for accent and delivery, for the guests. Video and graphics were generated by Mirage’s Avatar X model; the whole process was managed by Anthropic’s Claude models.

In terms of viewership, topics didn’t seem to have a major impact on audience viewership, which spiked a few times: at the stream’s beginning, after it ran a countdown; when people logged off from work, around 5pm; and right at the end, as it signed off. The average viewer only watched for about one minute, as Mirage detailed in its blog report on the show.

From a technical standpoint, viewers and Mirage’s own team recognized that the interviews came off flat: stilted in their cadence and delivery with audio that, even if pretty accurate, sounded monotonous and flat.

“The avatar is really good at talking, but it’s not as good at listening, because it’s not trained on it,” Misra says.

From an accuracy standpoint, the Mirage team was surprised by how many errors appeared in the draft scripts reviewed by the team, Misra says; they were able to spot and correct all but two, both of which prompted later on-air corrections.

Because Mirage ensured the feed always displayed an “AI generated” disclaimer, there were no disasters about people misunderstanding what they were watching, or guest appearances getting clipped or syndicated in misleading ways.

On a cost basis, $50,000 per day for such a show would be a tough proposition to continue, implying such a show would burn more than $1 million per month. Mirage believes it can cut down the cost of its video avatars by 10x or 15x in upcoming months.

Where this goes

Assume Mirage is right, and the capabilities and cost to put on a show like this converge.

I can imagine three scenarios playing out:

1. Good: AI channels help distribute niche and local news

2. Bad: It proliferates social media-friendly slop

3. Ugly: It doesn’t spread at all, because it still sucks

The good case: Mirage’s tech helps niche news that wouldn’t otherwise get covered – think hyper-local, even neighborhood-level reports on road blockages or trash pickup changes, or niche sports that don’t get consistent coverage today.

“There aren’t enough resources to do everything in a broadcast style,” argues Misra. “For journalists and news, I don’t think this will be some sort of career ender. I don’t think any journalist needs to be afraid of AI taking their job.”

Squint, and you can sort of picture it: video content as PSA; a local darts winner chatting with an AI anchor about some achievement, amplifying their story in a way that it wouldn’t get covered otherwise. To do it right, you’d still need people to flow in that intel, and curate it; otherwise it would be easily manipulated or unreliable.

(I also don’t think an AI anchor could ever replicate the emotional connect, and ties to a community, that real-life news anchors have. If you’ve ever watched a beloved local news anchor on TV, or seen the joy people get in running into the person who does the weather at a restaurant, you know what I’m talking about. If not, just watch the clip below, an emotional tribute to Dolly Parton by Denver TV anchor Kyle Clark, and you’ll get what I mean.)

And then, you’d end up in the bad case: AI news as sensational or bogus rage bait. I can see that happening relatively harmlessly at first. Famous or busy people send, or approve, their AI avatars to conduct interviews on their behalf, an evolution of how tech bigwigs like Elon Musk currently appear at conferences over video link, while every other guest physically shows up.

The AI makes a mistake, or hallucinates Musk saying something he wouldn’t, and it goes viral. The clip circulates on TikTok, Reels or on X, and context gets lost. Aggregator news sites then report on the viral clip as news, further amplifying it, and maybe getting it noticed by more trusted news sources in turn. Soon, even an authorized AI appearance has taken on a life of its own, in a way no correction could properly stop.

Imagine that without the authorization, or Mirage’s persistent “AI generated” notice. Things could pretty quickly get to a dark place, where public figures and brands have to play constant whack-a-mole to take down what are effectively deepfakes, at a scale worse than the current, not-great status quo. Would Mirage or the technology providers be able to revoke access and enforce their own terms of service fast enough? Would governments or enforcement agencies be sufficiently resourced or motivated to step up?

Then there’s the third, perhaps reassuring ugly case, where this tech gets much better, but is rejected by consumers as unappealing slop

Watching the experiment feed myself, I felt more appreciation for what TBPN got so right, and what subsequent clones like the European Technology Network (etn.) and Monitoring The Situation (MTS) are trying to find now: PMF. In this case, that’s not traditional product-market fit, but personality market fit.

It’s really hard to be interesting and compelling day after day, like TBPN hosts John Coogan and Jordi Hays. And it’s even harder to get your guests to say something actually fresh. I keep working on this with our own show, The Upstarts Podcast, and that isn’t daily or live. I’m skeptical that AI-to-human interviews, let alone AI-to-AI ones, can really nail the type of clip-making, news-driving hot takes or updates that make these shows go viral, and their viewership spike.

At Mirage, Misra subscribes to the first scenario. He knows he’s playing with fire on such a “controversial topic,” partly why his startup was excited to run the experiment and provoke conversation, he says.

But like many startup founders, for better or worse, his default is to focus not on what could go wrong, but what could go right. Mirage has another experiment cooking for the fall, he tells me.

“We are planning an even bigger idea,” he says. “And it’s crazier, in some ways.”

What happens when AI models can operate a computer more efficiently than a person – not just writing code or calling APIs faster, but clicking buttons, and navigating around, like it’s in control of your mouse?

Such capabilities, called “computer use,” aren’t a new concept. Big AI labs like Anthropic, OpenAI and Google DeepMind, and newer entrants like Perplexity, have all announced computer use options in recent years.

But one big blocker has still held back computer use from taking off: its cost.

Subscribe now

For a general use frontier model, mimicking how a person navigates around a screen requires a lot more back and forth, which in GPU access, means more money and time. And that has soured some on its value, argues Devi Parikh, co-CEO of AI startup Yutori.

That is, until now. Today, her startup is announcing a new model that its team believes will bring forward the widespread adoption of computer use by a significant step.

Called n2, the model was trained on just 27 billion parameters, orders of magnitude less than the several trillion parameters for leading models like Anthropic’s Fable 5 and OpenAI’s GPT-5.6. But it offers performance, Yutori’s founders claim, that isn’t just state-of-the-art for its weight class, but close to the abilities of much bigger, more expensive alternatives.

“These cost savings are the difference between the unit economics for a product making sense or not,” Parikh says. “It’s the difference between being able to ship a product at scale, versus just not shipping at all.”

Founded by Parikh and two other former Meta AI researchers, chief scientist Dhruv Batra, and co-CEO Abhishek Das, Yutori trained the model over the course of this year with just a team of 14 and a relatively small amount of funding; the startup announced a $15 million round led by Radical Ventures last year. Other investors include Felicis and AI luminaries Fei-Fei Li and Jeff Dean.

The company says n2 “excels” on all benchmarks it tracks, including OSWorld, OSWorld 2.0 and WeaveBench.

The big unlock: unlike other models, n2 can switch seamlessly between interfaces, from an API, to a command line interface (CLI), to a graphical user interface (GUI) more akin to what a human would use, all within a single task. The model chooses whatever route will be fastest and simplest, and can even write its own snippets of code when the other routes won’t work.

Yutori will provide n2 through neocloud partners like Baseten, Together AI, and Crusoe, says Batra. Who can benefit? Companies and developers who set up agents to perform daily tasks that combine web info with local software or files, such as checking performance dashboards, automatically paying vendors, or correcting bugs in micro-services.

“If someone is trying to build an agentic workflow for knowledge work, and that’s all they care about, they don’t have to be paying frontier level bills to be able to do so,” says Das. “There’s now a faster, cheaper option, still at the same accuracy, that they can use.”

The rest of this post on Yutori’s path to n2, how it works, and what it means for the ecosystem, is available to paid Upstarts subscribers. Take 20% off an annual subscription in our summer sale, or try one month for free.

Try 1 month free

Read more

Is your favorite startup CEO terminally online?

A quirk of hosting a podcast is that you quickly learn the internet habits of your guests. Some will be up-to-date on every subplot of Silicon Valley internet drama, prepared to riff. If you want to know their hot take on a topic, it’s usually as easy as pulling up their social media accounts.

These are my people, like CEO Shensi Ding, who spoke at an Upstarts event in June. Her startup, Merge, offers infrastructure to help businesses connect to apps and AI tools faster and cheaper. But on X, you’ll find her gonzo posting about everything from model routers to overhearing tech bros on cringey dates.

Then there’s Basis CEO Matt Harpe, who joined The Upstarts Podcast as our final guest of Season 2 a few weeks ago. It fell on Harpe’s co-founder, Mitchell Troyanovsky, to share that conversation on X and LinkedIn when it dropped.

“Listen to Matt (who does too much real work to have a twitter) on what Basis is doing for accountants,” he posted.

I thought about that line frequently this past week, as the Highly Online startup community piled onto a few trending topics: Anthropic making seemingly quite a lot of money; Stripe finally acquiring OpenRouter after a month of rumor and speculation, a new AI assistant agent, Instinct, delighting with its ease-of-use and worrying with its data retention policies.

You can argue that all of those stories have big-picture implications for a wide swathe of people. How Anthropic and OpenAI fare commercially can affect consumers and companies building on top of them, and two potentially massive, bellwether IPOs. A frenzy of interest in model routers, from Stripe, Ramp and others can speak to a moment where businesses are grappling with the costs of using such AI tools. And whether it’s through Instinct, or two other tools getting discussion online – SpaceXAI’s Grok Bot and Town – squint and you can see the path to mainstream AI assistants feeling a little more direct.

But for most people – and even most directly in the startup ecosystem – is knowing more about all of that, and faster, creating some advantage? Or is it scratching a different itch?

The question has a lot of relevance to Upstarts, considering we often publish stories about startups before they’ve become internet main characters, and as I continue to think about the balance of what coverage the startup ecosystem needs versus the coverage it wants.

I’ve had fun diving into a few startup stories in unexpected places in recent weeks, like Henry AI bringing AI to commercial real estate; Ambrook building the QuickBooks for farmers and other physical-world businesses; and Clair, providing a non-predatory alternative to payday loans.

I’m proud of those stories, and hope you’ve enjoyed them. I’m also aware that Upstarts would probably be growing our subscriber base faster if we cynically chased the same trending stories as everyone else. But what would be the fun in that?

Get 20% off for 1 year

On Thursday night, as last-minute, lawyer-driven intrigue delayed an upcoming story, I decided to try an experiment: write a column about some of these thoughts bouncing around my brain. So I pinged Ding and a few other Highly Online tech folks to ask if they feel like they’re getting real benefits from long shifts in the content factory – or simply indulging in 2026’s flavor of brain rot.

“I am massively torn. Very unclear if it’s worth the tradeoffs,” DMs investor Zach Coelius on X. “Good for the brand, bad for the sanity and focus.”

“Maybe I’m psycho, but I f***ing love it,” Ding texts.

My takeaways from those chats – from how AI adds urgency and psychosis, to how everything is a prediction market now – below, featuring Shensi Ding, Olivia Moore, Ken Wattana, Yoni Rechtman, Bilal Zuberi, and Jane Manchun Wong.

NOTE: If you enjoy this new column format, please let us know. We are grateful for any suggestions of future columns, interview subjects, or ways to refine this format.

And if your business would like to sponsor these columns to ensure they can remain free, we’d love to hear from you, too. Feel free to respond directly, or email partnerships@upstartsmedia.com.

Moving at AI pace

At Merge, Ding seems skeptical that a founder in AI, particularly, could not be fully plugged in.

“I think things change so quickly compared to before, you need to know what’s happening every f***ing day,” she writes. Still, she tries to focus on a few areas of particular importance – infrastructure, AI assistants, and AI governance – while leaving updates on hardware and data center issues for her husband, who works closer to those topics, to fill her in about.

“Anyone who has PMF [product market fit] has it for max six to 12 months, and is required to continuously keep releasing bangers,” Ding argues. “Versus before, you could have dominance for three to five years, and rely on that.”

Trendspotting as value-add

At a16z, investor Olivia Moore says that her online presence is about serving as a curator for her portfolio founders on new products and micro-trends.

“I feel like it’s the least we can do to have an understanding of what’s going on in the broader ecosystem to the extent that it can be helpful to our founders,” she writes to Upstarts.

Moore still subscribes to the idea, amplified by Elon Musk when he acquired Twitter in 2022 and later renamed it X, that the site serves as tech’s “digital town square,” providing exposure to new data points and arguments.

“This could all just be cope for my ridiculously high screen time :)” she adds.

An outsider’s way in

“Not knowing what’s happening may be a badge of honor,” notes the founder Ken Wattana – a sign you’ve made it enough not to be hustling.

Wattana, the founder and CEO of agentic payments startup Conto, says that he monitors ‘Tech Twitter’ similar to how finance professionals have historically watched their Bloomberg terminals. Being so online helps him to understand Silicon Valley dynamics, and what insiders might find compelling.

“I’m not in the inner circle, so it’s my way of trying to be as informed as possible,” he says.

Get 25% off a group subscription

Posting frequently has also helped drive a mix of inbound interest for his startup and a feeling that he’s in on the joke. “I’d like to think it helps with distribution and just being ‘known,’ Wattana writes Upstarts. And while his “shitposts” may not reach prospective buyers directly, “It is fun to see notable VCs and founders engage with them.”

Anything goes

Bilal Zuberi, founder of Red Glass Venture, says he’s often online for two reasons: to be more accessible to startup founders; and to encourage them to feel comfortable to be authentic in public, “vulnerabilities included.”

There’s a darker side to this, too, though: a devaluing of “real, serious journalism” means that anything is fair game to say, Zuberi adds. “People are rewriting history in real time, and those with big, loyal followings seem to think that insulates them from criticism – however deserved it may be.”

He points to Turner Novak’s satirical tweet from earlier this past week, in which the investor and podcast host jokingly announces he’s making a full circuit of non-journalistic, often investor or tech-owned ‘new media’ properties.

Everything is a market

Yoni Rechtman, a partner at Slow Ventures, sees the irony in investors who back early-stage startups for a decade or more acting more like day traders on social media.

“It’s just another symptom of how private markets are coming to look more and more like public markets,” he says.

And it feels like no coincidence that the rise of sports betting and prediction markets has dovetailed with increased interest in trading startup equity before these companies go public. It used to be that revenue milestones or new valuations were indirect, longer-term signals. Increasingly, as Robinhood and others look to provide financial vehicles for retail investors to gain exposure to tech unicorns, they’re potential trading signals.

Can’t get into Anthropic or OpenAI’s upcoming IPOs? You can trade them indirectly, through markets on Kalshi and Polymarket for who will go out first, or at what price.

For startups and investors, Tech Twitter can provide a pulse, or a view of emerging consensus, but fewer immediately actionable insights, Rechtman adds.

“I think it’s generally very useless to try and send tweets and articles to my founders, being like, ‘Did you see this?’” he says.

Avoiding X psychosis

Late last year, writer and researcher Jane Manchun Wong says she found herself too plugged in – unsustainably online.

So she set up her own feed to syndicate posts from accounts and topics she cares about, like a new-look RSS feed. She also leans on her subscription to Techmeme to help her see just enough about a new story or controversy, without getting sucked down a rabbit hole.

“I’m not going to click thousands of posts. I’m just going to look into it, and then think, ‘Oh, makes sense,’ and then move on,” she says.

As people spend more time talking to AI coding agents or assistants, their disconnect from reality could grow. While Manchun Wong still checks X frequently for “chatter” and general sentiment, she’s also changed things up by using her account to post new restaurants local to her in the Bay Area (what she sees as influencer-free, grassroots promotion) while prioritizing more in-person meetups.

When Ding calls herself a “psycho,” it’s hard to know how much she’s joking. But Rechtman seems prescient when he warns that too much time on Tech Twitter – much like in ChatGPT or Claude – could be dangerous.

“It is increasingly important to have an actively maintained cognitive security and cognitive hygiene,” he says. “Because the meme cycle and the feverish breaking news coming out of startup land right now will truly break your brain.”

The Upshot

On the website of Nico Simko’s startup, Clair, you’ll find an unusual stat: percentage of U.S. businesses that offer his company’s on-demand pay solution to their employees. As of mid-August, it hovers above 5%.

“That’s a huge milestone for us,” Simko says. “But the goal is to get to at least 20% over the next few years.”

In less than two years in the market, Clair has grown fast. Those 300,000 businesses translate to about double that in monthly active users. Recently, the volume of wage advances passing through the fintech startup reached an annualized run rate of $2 billion.

And for Clair, that means the startup has reached a run rate of $100 million in revenue, up 10x from just over a year ago, Simko tells Upstarts. Clair also reached positive cash flow this year, the company says.

Clair is making that money by offering early access to unpaid, earned wages to workers at these businesses, which can range from mom-and-pop operations to ones with national footprints. But Clair is different from other solutions in the market for two big reasons: it’s not a payday lender playing an interest game; and it’s not a consumer app, but instead works through payroll and workforce platforms, like Gusto, QuickBooks and TriNet.

“It’s much more efficient, it’s cheaper, and the bucket is less leaky,” Simko argues. “And people have more disposable money at the end of the month.”

Get 20% off for 1 year

Founded in 2020, by Simko, Alex Kostecki and Erich Nussbaumer, and named after the French word for “clear,” Clair’s initial product looked very different, and it didn’t work out.

But more recently, Clair’s unusual embedded infrastructure model – reminiscent in some ways of Plaid, or its co-founder William Hockey’s bootstrapped second act, Column – has seemingly pulled off a highwire act between making corporate partners like Gusto happy, and providing a service that profits off employees needing fast money, without devolving into exploitation.

Clair knows you might be skeptical about the altruism behind a for-profit, venture-backed startup. Clair has raised $69 million in funding to date, most recently a Series B last year, from investors including Founder Collective, Thrive Capital and Upfront Ventures.

So the startup’s got data: out of more than 37,000 responses to its in-widget customer satisfaction survey over the past 90 days, 89% said they were “very satisfied” or “satisfied,” notes Erik Webb, Clair’s vice president of data and growth.

And for Simko, who launched Clair out of personal experience as an immigrant student in the U.S., Clair represents a personal mission that he would walk away from if it became just about the dollars, he insists.

“At the end of the day, I believe that this is a product that you should graduate from,” he says.

More on how Clair found product-market fit, why partners like Gusto call it “extremely well loved,” and why Simko thinks you should be rooting for his mission, below.

The rest of this post is for paid subscribers. To read it, try one month for 10% off. Or support our independent startup reporting by taking 20% off an annual subscription through our Summer Sale.

Read more

On April 8, recruiters and a hiring manager at startup Checkr logged on to meet over video chat with a promising applicant for an engineering role. And immediately, they knew something was wrong.

On paper, the person – whom we will call Ulysses, in honor of the Trojan Horse that was aiming at their company – seemed highly qualified, with a career as a software engineer dating back a decade, and the kind of casual, outdoors selfie-style profile picture with salt-and-pepper beard that says: “I’ve got job security at a leading tech company” — in this case, Stripe.

But when Ulysses joined the video screener over Zoom, the person who logged on looked very different: clean-shaven, far younger in a baseball cap and hoodie, and Asian. His IP address and VPN address both seemed suspicious. His resume didn’t appear normal. And the interviewer thought he could hear other voices in the room. Were they coaching Ulysses? Feeding him answers? Or taking similarly sketchy interviews?

The “About” section on Ulysses’ LinkedIn page seemed to have anticipated this possibility.

“WARNING WARNING WARNING WARNING WARNING: A whole bunch of AI bots, hackers, and marginal marauders have been applying to jobs using my name, my LinkedIn profile (hi!) and an LLM generated resume,” blared the page.

“I am not currently applying for work,” it added, while encouraging recruiters to waste any impersonators’ time, then push them out the door of a fast-moving car.

Okay, so this engineer was probably not really applying to a job at Checkr. But the interview team were intrigued, because of the irony of the situation: applicant verification is kind of Checkr’s whole thing.

So in a plot twist of a decision that went all the way up to co-founder and CEO Daniel Yanisse, the Checkr team did something a little crazy.

“Let’s continue playing the game with him,” Yanisse told the team. They decided to offer Ulysses a job.

The rest of this post is for paid subscribers. To read it, try one month free. Even better, support our independent startups reporting by taking 20% off an annual subscription, in our Summer Sale.

Take 20% off

Read more

The Upshot

In the early days of her startup, Ambrook, CEO Mackenzie Burnett and a small team would hop on a plane – then often take a bumpy drive – to personally onboard rural customers of her financial software business.

Even then, Burnett says her vision was to help out across the “real economy,” streamlining bookkeeping and expenses for small businesses that grow, raise, or build and move along physical goods. But she expected Ambrook to focus on agriculture for years.

So Burnett was surprised when, as Ambrook wrapped up a visit to set up their second customer, a cattle feeder operation in Arizona, the husband-and-wife owners told her they were ready for Ambrook to set up their other businesses – all four of them, none of them in farming.

Trucking and custom goods and services weren’t on the immediate roadmap. But a lightbulb went off.

“That was the very early realization that what we had built was something that was much bigger,” Burnett tells Upstarts.

Today, most of Ambrook’s 8,000-plus business customers – up from 2,500 a year ago – are still mostly in agriculture. But more than 1,000 trucking businesses now use Ambrook, as well as hundreds of general contractors and property managers. They turn to Ambrook for help with receipts and bill paying, cash management, and record-keeping for compliance and tax audits.

Often, they’re replacing QuickBooks, the accounting software platform developed by Intuit more than 30 years ago. But at least half of Ambrook’s customers turn to it as their first software vendor, according to Burnett.

“Really big, multi-million dollar businesses run on paper in America still, and I don’t think people realize that,” she says.

Now, the startup she co-founded with Dan Schlosser and Jeff Anders in 2020 is ready to expand: both up-market into bigger ‘ag’ businesses and farms, as well as into those other categories like construction, logistics and real estate.

To do so, Ambrook has raised $30 million in a Series B funding round led by solo capitalist Lachy Groom, Upstarts exclusively reports. Thomson Reuters Ventures, Thrive Capital, Field Ventures, and Cameron Ventures all participated, alongside angel checks from co-founders of Gusto, Notion and Vercel.

Get 20% off for 1 year

The mission – to help American businesses be more profitable and sustainable – hasn’t changed, Burnett tells me in a visit to the startup’s New York office in July. The space is filled with plants and mismatched vintage wooden furniture, much of it sourced from Facebook Marketplace; employees leave shoes at the door, partly to preserve its wooden floors.

(In a major faux pas, I don’t make that connection until I’m leaving, as Burnett hands me a copy of Ambrook’s print gazette, The Tomato Times, which I open to a guide matching tomato varietals to astrological signs.)

Offices in San Francisco and Denver maintain the Americana motif. But don’t be fooled by such folksy charm, or Ambrook’s Stetson-heavy Instagram page. What the startup is building for these businesses gets high-tech.

One customer in Vermont tells Upstarts how he’s incorporated Ambrook’s recently-launched MCP server into an AI assistant to reconcile errors in his farm’s ledgers (more on that below). A hunting range in Texas built its own point-of-sale system on top of Ambrook’s software to manage bookings, Burnett adds.

“Farmers are the original DIYers, for sure,” she says, referring to ‘do-it-yourself’ fixes. “We’re the one thing that our customers don’t want to roll out on their own.”

Here at Upstarts, we pride ourselves on the depth and care that goes into our startup reporting. Interested in aligning your brand with stories like this? Drop us a line at partnerships@upstartsmedia.com 👋

Rivers of paper

One of Ambrook’s employees jokes that the company should be called “Brackish water,” Burnett says, because of how it combines more traditional startup tech jobs in New York and the Bay Area with commercial team members in rural Montana.

The original idea, and name inspiration, for Ambrook came about after Burnett spent a summer meeting with farmers about water access in California’s Central Valley in 2018. After co-founding and selling her first startup to CoreOS, earning her a place on Forbes’ 30 Under 30 list, she left to pursue a master’s degree at Stanford University, studying the Pentagon’s response to climate change.

But after teaming up with Schlosser, a former product manager at Google and The New York Times, and Anders, a former designer at Meta who then served as Head of Design at Scale AI, Burnett discovered that water infrastructure in the U.S. was too convoluted to support a private actor marketplace on top, she told the Future of Agriculture podcast last year.

Get 25% off a group subscription

Speaking with dozens of farmers and policymakers in agriculture, a more classic startup problem emerged. “People were just kind of drowning in paperwork,” she said then.

In startup terms, Ambrook had found a classic wedge: a customer category too wonky, or niche, for the incumbent player – QuickBooks – to have built a tailored product. Farms file different taxes from other businesses; they deal with a much wider range of expenses and bills.

“They’re one of the only businesses where your inventory gets born, it grows, it dies,” Burnett described it on another podcast, Dialectic, also last year.

Ambrook spent several years working hands-on with early test users to develop its answer. When it launched, the startup invested in educational courses and a certification process to help ensure onboarding was relatively smooth. Employees went on small, local podcasts and radio shows to spread the word, appeared at universities and meetups, and engaged with industry Facebook groups.

“We wanted to be really, really sure that our product was going to hit when it did, because of the structural challenges of selling to a market like that,” Burnett says now.

One of those early customers was Brandon Bless, the land and animal manager at Bread & Butter Farm outside Burlington, VT. A 650-acre community farm that grows vegetables, tree fruits and berries, raises cattle and other animals, and engages in agroforestry, Bread & Butter is a complex operation, Bless says, that flexes up from six full-time staff year-round to about 50 team members in its summer peak.

Much of that complexity comes from the fact that, like other farms, Bread & Butter operates six independent, but interconnected businesses on its premises – not just including the education program, CSA and cafe you might expect, but a non-profit and ventures in health and music, too. Each maintains its own profit-and-loss accounts, and QuickBooks broke all the time managing them.

After finding Ambrook online, Bless agreed to try it after the startup promised ‘white glove’ onboarding it would unwind back to QuickBooks if it couldn’t deliver. Sure enough, Ambrook’s team came to stay at the farm, where employees met with Bless and his co-workers to spin up features from inside the room.

Bread & Butter now uses Ambrook across five of its businesses, saving it hours each week in bill-paying and bookkeeping, and giving it more confidence about its margins for decision-making. Bless personally connects Ambrook to Anthropic’s Claude via its MCP server to manage bookkeeping, lessening the need to hire third-party services.

“We’re saving a huge amount of money because of Ambrook and the AI integration,” he says.

Tilling new soil

Ambrook has resonated with farms that bring in up to about $20 million in sales, Burnett says; to move into mid-market, bigger farms, it will need to flesh out its payroll and operations features. The startup is also developing a services function (a trend we’ve written about extensively at Upstarts), and is hiring for that role.

But it’s in other categories adjacent to agriculture where Ambrook can likely expand the fastest: businesses like contracting, trucking and real estate where farmers like the Arizona couple have already tried to pull them in.

Copywriter turned general contractor Zach Slovin tells Upstarts that he’s a little embarrassed that, despite his advertising background, he found Ambrook through a Facebook ad.

The owner of his own home renovation business, Home Reflections LLC, Slovin says that he deliberately keeps a lowkey tech presence on job sites, bringing just a clipboard and his smartphone, so as not to seem too corporate to trust. (You will, however, find him active on TikTok.)

And Slovin says he turns down outreach from companies trying to sell him AI tools on a seemingly daily basis. Ambrook, because it was replacing the QuickBooks “behemoth,” was different. Since reaching out in December, Slovin now uses Ambrook to pay bills to subcontractors faster, and to better keep track of his projects by category, budget and scope. He’s also surprised himself by using its more recently-launched project estimate tools.

“It’s not that I have specific brand loyalty to any one product,” he says. “But it feels like as long as they’re doing things in the way they’re doing now, and moving in a positive direction, I can’t see any reason I would move away from them.”

One looming challenge for Ambrook: Both customers who spoke to Upstarts said that the startup’s customer service, including quick response times and the ability to fix bugs in hours and ship features in days, have helped earn their loyalty. That high-touch service is tough to scale. Ambrook’s now hiring customer success managers following its fundraise, too, per the jobs board on the back page of The Tomato Times.

Join the WhatsApp group

Is there more to Ambrook’s moat than that? What happens if QuickBooks wakes up to a market segment slipping through the cracks, a team in Y Combinator hones in on the same idea, or Anthropic and OpenAI move deeper into SMB bookkeeping?

“AI has made every business more honest about where their real network effects are,” Burnett responds. The moat isn’t a headstart in building software, she argues, but in economies of scale, network effects and proprietary data like the tens of thousands of receipts regularly uploaded to Ambrook.

Burnett brings an “authenticity” and “reverence” for her customers that founders can’t fake, says Ambrook’s Series B investor, Groom. The co-founder of robotics AI startup Physical Intelligence, or PI, Groom adds that he and Burnett also share an appreciation for the physical world.

“Everything around us was built by someone that poured blood, sweat and tears into it,” he adds.

True long-term resilience, Burnett argues, comes from helping those people find financial stability, such as putting their kids through college. Only then can they prioritize longer-term goals, such as environmental conservation, without going against their own interests.

“It’s about building tools for folks who want to do the right thing, and making it easy to do the right thing,” she says.

“For us, this is ‘American Dynamism’ in a way that actually empowers people – the millions of people who are in the position that, if they were given the chance, would feel sensible ownership around stewardship, but just haven’t been asked.”

When Basis co-founder Matt Harpe tells me he wants to build the best home for applied AI talent in New York City at his startup, I’m immediately skeptical.

Harpe’s startup, Basis, serves a specialized customer base that I don’t associate with the bleeding edge of technology: accountants.

Unless there’s some untapped talent pool of accountants turned machine learning experts that Upstarts is unaware of, why would someone choose working at Basis, despite its recent $1.15 billion valuation, versus another local option?

The vast majority of Basis employees don’t have accounting backgrounds, Harpe confirms. “There are probably 5% to 10% of the company whose parents are accountants, which I think is actually an interesting observation, but that’s a relatively small portion of people,” he admits.

Instead, your more classic tech-minded employees are joining Basis because of how it builds its tools, not so much who for.

Most applied AI companies started with a chatbot, or a text box for inputting a prompt, Harpe argues. When he and co-founder Mitchell Troyanovsky launched Basis in 2023, they focused on longer-running agentic systems by necessity. “It was required to move the needle in accounting, so we were forced to,” he says.

Now, he believes Basis has a head start as agents become the topic du jour among peers: “For a lot of people who are really interested in pushing the bounds of what’s possible within the broader applied AI and ML [machine learning] space, Basis is a pretty exciting place to work.”

We did it, Joe: we’ve reached the final episode of Season 2! We’ve been grateful to partner with Rippling across this season, and to all of you for following along 👏

We’ve already got some exciting guests lined up for Season 3, but we’d love your input on how to improve and grow the show!

Are these write-ups useful? Know a great guest? Interested in sponsoring? Please drop us a line or leave a comment, and we’ll reach out 👋

Leave a comment

In February, Basis announced a $100 million Series B led by Accel, with GV, Khosla Ventures and a host of individual investors from former Goldman Sachs boss Lloyd Blankfein to the CEOs of Box, Quora and Hugging Face, as well as AI luminaries like Google’s Jeff Dean and OpenAI’s Noam Brown.

Basis works with at least one-third of the Top 25 accounting firms in the U.S., helping them manage workflows across processes like taxes and auditing, and returning work for the human accountant to review.

Harpe – who got the idea while studying healthcare at Boston Consulting Group, where he learned accounting was a major bottleneck – isn’t an accountant himself, nor is his co-founder. But similar to Lassie, the startup we covered building software for dentist offices, they shadowed them, starting with Troyanovsky’s pharmacy-owner mom’s own accountant, to learn their problems.

And far from replacing human accountants, Harpe insists that his startup is helping solve a crisis in this field: there aren’t enough people. There’s enough accounting work in the U.S. for 30 million specialists, he says; we make do today with just 3 million, and the number is declining.

• Listen on Apple Podcasts

• Listen on Spotify

On this episode of The Upstarts Podcast, Harpe talks about how Basis got a head start by shunning a chatbot approach; the ins and outs of outcome-based pricing, and why it won’t work for all; and how he believes he can increase accountant salaries, not depress them.

Plus, he shares his Upstart Moment: turning down short-term revenue by not charging for what he considered an unfinished product early-on.

Our key takeaways from the show for busy builders, and ‘Alex’s take,’ are below.

• A 3-part pitch to win customers

• Trade-offs of outcome-based pricing

• Alex’s take: Find your ‘Goldilocks’ vertical

A MESSAGE FROM OUR SPONSOR

Your AI is only as good as your data

Rippling AI is built on your live workforce data and permissions. So it gives you full visibility across your startup (on anything from headcount to burn rate) — and can take real action right inside the platform. Want to see how it can work for you?

Get Exclusive Access: rippling.com/ai

A 3-part pitch to win customers

When Basis got started, few accounting firms seemed excited about adding AI tools for the sake of it, says Harpe.

That part’s gotten easier since, but Basis recruited early customers by leading with a pitch to help solve their staffing challenges. Coming out of the pandemic, high attrition rates were a concern. Large firms had invested in offshoring some work, which had its own tradeoffs.

Basis’s strategy: identify the top one or two issues for a firm leader, then promise to help across three areas:

1. Efficiency: automating some manual processes

2. Retention: less busywork for employees would make them more likely to stay

3. Growth: people trained on such tech would be more likely to advance, creating a hiring story for future candidates

“That got people really excited. We were able to show that we could actually make an aggregate impact on the entire practice, change its efficiency profile, not just pick one niche workflow that wouldn’t have any total impact.”

On a practical level, Basis being able to pull documents, flag exceptions and clearly present the results for review was the demo magic. “These are processes that might otherwise take 10 hours, even 100 hours, depending on the volume, that can be done in a matter of minutes, with extremely high levels of accuracy.”

Get 20% off for 1 year

The trick there: showing their work enough that the accountant could easily verify it, and feel confident that the process had run correctly, so they didn’t have to retrace the agent’s steps.

Trade-offs of outcome-based pricing

While Basis resisted calls from investors and some customers to build a short-term chatbot for basic questions, the startup also chose not to prioritize generating revenue from its tools until it felt ready to align pricing against real outcomes.

“We never had a seat-based model for our core product,” Harpe says. No token or consumption-based model, either.

Instead, Basis “held out” until it could deliver outcomes that it would charge for, then “put gas on the fire.”

“ One of the big problems with outcome pricing is for some startups, it’s not easy to prove exactly what the value is that you’re generating, right? Like, you could say, ‘We’re helping with a few things indirectly,’ but maybe you can’t map to dollars saved or new money coming in for the customer.”

Because accountants typically get a predictable monthly or yearly amount for their client services, Basis has an easier time mapping where its time saved on an account can contribute to that revenue.

“Until the AI is actually good enough to do a meaningful portion of that [work], you can’t really put that model in place.

Whereas if you just want to put a seat-based model in place, you can probably do that sooner, because people think, ‘Okay, this is one person. You know, they’re going to save this much time,’ and the ROI story is relatively clear.”

When I ask Harpe if he expects more companies to emulate the outcome model, he’s skeptical. “Honestly, a lot of the direction things have gone is maybe not in the direction we’ve gone,” he says.

Get 25% off a group subscription

Instead, many AI startups are using a token-based, consumption model, adding a surcharge to their own model costs that they pass to a customer.

“You’re punting the question of figuring out ROI and making a case about how it can be attributed to the customer,” Harpe says.

Inside Basis, Harpe notes that the startup tracks outcomes of coding tools to decide how much to spend on them, but volume of activity alone is not enough.

“You can maybe look at per-person productivity, but I think it’s a lot less clean. So I think that puts you in a slightly more difficult position, where you’re essentially pegged to this underlying resource that can be swapped in and out. Maybe, as the cost of the models goes down dramatically, that is a more difficult business to defend, and you are truly wrapping an atomic token, instead of packaging an outcome that you want to continue.”

Bottom line: Basis is happy where it is focused on outcomes. And it seems nice work, if you can get it.

Alex’s take: Find your ‘Goldilocks’ vertical

A major theme of this season of the show has been how startups can ride the wave of the big AI labs, without getting swept away.

Among the 10 startups we featured on the show this season:

Orchid in fertility might not have to worry about ChatGPT or Claude’s health units yet, nor TBC in its work with neuron-trained models, nor Radical AI in materials science.

But at Vanta in compliance, Mutiny in sales software, Writer in enterprise AI, Blitzy in corporate-scale code projects, Pigment in ERP software, and Abridge in healthcare systems, that dance is now part of the story these companies tell to justify continued hiring, investor interest, and long-term viability in the AI era.

We’ve covered this trend extensively in our articles, too, from Handshake pivoting to data creation and labeling to an agentic pivot at Vivun, and a push into AI native services at Gainsight and Remote. It’s in the backdrop of many of other conversations, too, from Wispr launching a new interface lab to Zoom’s CEO plotting an updated product future.

Most dramatically, there was our story about Figma and its board kerfuffle following the Claude Design launch.

Join the WhatsApp group

Basis also fits neatly into this conversation. It’s easy to imagine Anthropic and OpenAI launching Claude Audit (Claudit?) or ChatGPT for Accountants. And Harpe agrees that the question of defensibility against the labs was one of two big existential ones he addressed to his company as they got started; the other was how they’d continue in a post-AGI world.

His answer now: he fully expects the labs to launch accounting and finance features. “At least for basic productivity things, there are a lot of merits to having a ChatGPT subscription, or something like it,” he says.

What helps him sleep at night: the labs have “successfully attacked” capabilities one degree removed from their core competency so far, and accounting isn’t that.

“I think all companies can only focus on so many things, no matter how much money you have. Maybe that’s a lesson that sometimes the labs have learned recently, which is that if you try to go in too many directions, you start losing your core thing.”

Documents, text generation, Excel and PowerPoint, internet research – all of these are easy pickings, Harpe says. Even differentiated, deeper features in this areas will face pricing pressure. (This is where Writer CEO May Habib’s comments on the show about how Writer trained its own models, and Abridge CEO Shiv Rao on how his team uses internal AI tools 40% to 60% of the time to keep down costs, resonate.)

“Accounting has never been a text in, text out thing,” Harpe argues. This is not a Microsoft Word, or Google Docs loving crowd. The sign-offs and permissions aren’t easy to do with a chatbot, he adds, a line that echoes the arguments of one of our Season 1 guests, Winston Weinberg at Harvey.

How far away your business is from a text box could help determine its long-term value, Harpe goes on to say, along with the usual need for good execution.

I mostly agree with Harpe. Like with Henry AI, the commercial real estate-focused startup we covered earlier this week, I’d rather find a wonky category that has historically not been exciting or seemingly big enough for founders and investors, and test the limits of growing its category.

Better to sell such a business for single digit billions at a best-case scenario than get gobbled up. But you have to raise relatively sane amounts of capital, like Basis’s previous $100 million round, for that to be a viable exit route.

In 2024, Henry AI founders Sammy Greenwall and Adam Pratt were midway through Y Combinator’s summer batch when they delivered their pitch to a skeptical venture capitalist.

Henry was looking to solve back office blockers in commercial real estate using AI-powered software, Greenwall, the startup's CEO, explained.

“You seem like a smart guy,” the investor replied. “I would not build this business, because this is a terrible customer base to sell through.”

Brokers are a tough crowd. The best have honed playbooks over years that they’re not going to easily entrust to an outsider. And it’s a relationship-driven, network-dependent field where multi-million-dollar deals get done over group text.

“100%, it’s historically a terrible market,” Greenwall says now.

That’s historically. Today looks different. Henry’s core Deck product, which generates marketing materials around a property with the requisite branding, maps and data, condenses about 15 hours of work by a designer to a median turnaround time of less than four, with only 30 minutes of that actual human review.

For more startup stories like this, take 20% an annual subscription as part of our Summer Sale. Support our work, and get exclusive scoops and interviews, for less than the cost of a sandwich each month.

Get 20% off for 1 year

In less than two years in the market, the tool is now used by more than 150 firms including Newmark and Berkadia, and spanning four of the five largest brokerages in the U.S. Henry doesn’t disclose exact revenue, but with pricing starting at $2,000 per month and scaling to seven figures based usage, it runs in the millions.

And with the launch today of a new product, Henry Deal, the startup is significantly widening the aperture of its work with this seemingly niche, but large industry (combined annual revenue: $112 billion).

Leveraging in-house tech and outside AI models, Henry Deal automates more of the deal process, including buyer lists, memos and underwriting.

To build it, New York-based Henry is also announcing a funding round it quietly raised last fall: $16.5 million in Series A funding led by FirstMark Capital, along with Thomson Reuters Ventures. Other return backers include Y Combinator, Susa Ventures, 1Sharpe, StoryHouse Ventures, Pioneer Fund, RXR Arden Digital Ventures, Karman Ventures, and Coalition Operators.

And that same investor who warned Henry’s founders against commercial real estate – FirstMark’s Adam Nelson – led the round and has joined its board.

The result of a madcap product sprint – more on that below – Henry Deal is still unproven with most of the startup’s customers; the name of the game is now execution, and it’s early, Henry’s founders admit.

“We’re at the forefront of what I would call the hard part of AI,” says Pratt, Henry’s CTO. “Great, it exists. Clearly, it’s disruptive and people are making huge bets on it. But it’s up to us to figure out how different industries, and different people figure it out.”

More on that, including perspective from several Henry AI customers, below.

Here at Upstarts, we pride ourselves on our in-depth startup launch coverage. Interested in sponsoring the newsletter to align with stories like this? Drop us a line at partnerships@upstartsmedia.com 👋

The deck guys

One of my favorite questions for founders: did they imagine running a company in their category when they were growing up?

Greenwall might be one of the first that Upstarts has covered who can plausibly respond with a “yes.”

Growing up in the East Bay, his family ran a commercial real estate business, so the shoptalk of brokers entered his consciousness by age 7 or so. In high school and college, he worked ‘coffee run’ internships to learn the industry, and started his professional career as an analyst at Toll Brothers, where he got used to getting “yelled at a lot.”

Greenwall’s move into startups happened when the founder of Lev, a startup looking to build software to optimize commercial real estate (CRE) financing, brought him on as a domain expert co-founder in 2019. They raised multiple rounds of funding on what Greenwall describes as an up-and-down journey before his departure to start Henry AI in 2024.

Get 25% off a group subscription

(Lev now pitches itself as more of an end-to-end tech stack for CRE, advertising against Henry AI’s search results on Google; Greenwall says he learned about that only recently from an employee, and wishes Lev well as a mostly non-competitive business.)

To build Henry, Greenwall tapped Pratt, a high school friend who had built and sold an app for managing fire department equipment while in college, and later became an engineering manager at Zocdoc. The pitch: to be the Harvey (legal AI) or Rogo (finance AI) for CRE.

That pitch, and their focus on the back office as an area to automate, got them into Y Combinator, where they competed in a friendly rivalry with Greenwall’s wife, Eden Ovadia, the co-founder and CEO of FINNY, a startup building tools for financial advisors.

Henry’s founders were motivated by parting words from their group partner, Michael Seibel, who warned them that if they didn’t stay in the Bay Area, they would fail. “We took that to heart,” Greenwall says. “We were like, ‘We’re going to outwork the SF tech people.’”

Back in New York, Henry and FINNY both worked out of the married couple’s apartment for months; visitors found employees asleep on couches, in a scene befitting San Francisco hacker house culture.

Renaming himself “The deck guy” on LinkedIn, Greenwall reached prospects through his content and personal outreach; both would hop on planes for pitches, with Pratt live-coding changes from inside the meeting room.

For Shaun Riney, a broker at Marcus & Millichap NYC, a backlog of decks was a years-long source of anxiety that would ironically get worse as business was brisk. A hapless in-house designer might get phone calls late at night from brokers looking to expedite a sale, but often the process could run weeks.

With Henry, Riney was able to leave only the highest-value, most important decks to those specialists.

“The freedom of being able to do something on your own at 9pm for the following morning, without involving someone else, it’s huge for the business,” he says. “Part of me now is like, ‘I don’t want too many people to know about this.‘“

Hearing footsteps

But even as Henry grew quickly, its founders had reason to worry early this year, when Anthropic released its Claude Opus 4.6 and 4.7 models.

Against a backdrop of a ‘SaaSpocalypse’ challenging the viability of software businesses, Henry noticed weakening demand among prospects and upselling opportunity in its customer base. It was still growing, but Greenwall and Pratt saw cracks forming if they stuck to decks.

“There was 100% a level of fear here that we have to be moving faster, and providing more value every day,” Greenwall says.

So they put the startup through a grueling, 35-day product sprint to launch Henry Deal for test users by June 15. Unlike the decks product, which still involves some human intervention on the back-end for now (to be phased out over the next year or so), Henry Deal was designed to work fully autonomously from the start.

Join the WhatsApp group

And whereas the decks tool might have faced existential risk as just a ‘wrapper,’ Henry designed its new software to swap, and benefit from, whatever state-of-the-art models it needs for a task, more like an interface, says Pratt.

At Grandstone Investment Sales in Ohio, founder and CEO Meir Perlmuter specializes in selling self-storage facilities. “The biggest time suck,” he says, is putting together what’s called a “BOV,” or broker opinion of value, that estimates the property’s market price to help expedite a bidding process.

With Henry, Perlmuter’s firm was able to have one analyst produce three or four BOVs per day; now Henry does it all across the firm’s seven brokers, no analyst necessary. “We’re all in on it,” he says. (Henry says that more than 20% of its customers have similarly cut overhead by using it – welcome for the bosses, not so much those displaced staff.)

‘Front lines’ of AI value

The disruption that Henry’s founders feared from Anthropic hasn’t materialized, they say, but served as a good forcing function to shift into higher gear.

Instead, they say they find at large corporate prospects that buyers are facing pressure to deliver on AI transformations, sometimes running up multi-million-dollar bills on Claude. “It’s brutal,” Pratt says.

“I probably could do this on Claude myself,” counters Perlmuter, who says it’s a topic of conversation among his peers. “But then it would be my problem. I would rather it be their problem, because they do a much better job.”

The Claude (or ChatGPT) versus everyone else debate is reminiscent of what Harvey CEO Winston Weinberg discussed on The Upstarts Podcast back in March. In that conversation, Weinberg conceded that “a bunch of functionality” within Harvey’s basic productivity suite for legal work would get commoditized over time.

What wouldn’t get gobbled up, however, were its integrations deep in a workflow, its vertical-specific knowledge, and its proprietary data not available to the models to train off of, he argued then.

What will prove Henry’s most defensible moat? One guide could be Crosby, an AI-enabled law firm we wrote about last October, whose CEO Ryan Daniels posted on LinkedIn earlier this week about moving to insured, AI-only outcomes for clients. That’s directionally the way that Henry seems to be moving for CRE brokers.

Nelson, their lead investor at FirstMark, says that for now it’s a matter of execution, and out-hustling the field.

“With a lot of these companies right now, we’re seeing this marriage of product and go-to-market,” he says. “Because so much of the product is ultimately going to be augmented and reinforced by the way that the software is used, and the context derived from it.”

Greenwall notes that the night before Henry’s founders spoke to Upstarts a final time last week, they and half a dozen others were in the same conference room at 2am, working on a customer fire drill.

“If there is a reason that Henry isn’t successful, it’s because we’ve failed ourselves,” Greenwall says.

Pratt, meanwhile, says that with all the billions of dollars being poured into AI today, it’s startups like Henry that can “prove the value in the weeds.” “If this doesn’t work at scale, all that infrastructure doesn’t really matter,” he argues.

Still, they know that what they’re building isn’t as romantic as, say, their YC batch mate, Starcloud, looking to build data centers in space. “We’re the deck guys,” Pratt adds. “So of course there’s a chip on our shoulders.”

Every day, across more than 10,000 businesses, millions of people dictate their words instead of typing them out using a tool from a startup called Wispr.

Tapping a button to record, they unload their thoughts and instructions on the product, Wispr Flow, and Wispr’s AI produces the text onscreen – cleaned up and grammatically correct, or more aggressively polished if the user asks.

For slower typers, people on the go or just those who need to verbalize to think through a task, it’s a nifty trick. Since Wispr Flow’s launch in 2024, it’s spread across staff at more than 450 of the Fortune 500, the startup says, with revenue growing at least 150% each quarter for the last year.

But for CEO Tanay Kothari, dictation is just the start. Since he was a 10-year-old kid in India watching Iron Man, Wispr’s co-founder has set his sights on a much more ambitious – but relatable – goal: he wants to build Jarvis (technically, “J.A.R.V.I.S.”), the AI voiced by a soothingly British-accented Paul Bettany in the film, who manages protagonist Tony Stark’s life.

Kothari’s talked about this vision before. Today, Wispr is taking a huge step in realizing it.

Wispr is throwing its hat in the ring as a full-fledged AI lab, Upstarts exclusively reports, with the creation of the Wispr Advanced Interfaces Lab Team under recently appointed chief science officer Ariya Rastrow.

Rastrow, a highly-cited researcher and veteran of the founding team at Amazon that worked on Alexa and later as a distinguished AI lead at Meta working on voice and multi-modal models for its AI devices, has already assembled a team of about a dozen researchers from Amazon, Meta, and academia.

Wispr plans to hire up to 50 researchers for the Lab, representing a commitment of hundreds of millions of dollars. At a high level, their mission will be to develop a new family of voice AI models, and a harness to use them, that can start to work more proactively and intuitively with a user.

Rather than simply dictate instructions, or follow a turn-based conversation with a voice agent, Wispr envisions a more collaborative workflow, in which the interface is continuously learning from your recent actions, following what you’re doing on the screen using computer vision, and learning about you enough to understand nuanced instructions, while suggesting and running others behind the scenes.

“We want to meet people where they are,” says Rastrow. Think of Wispr’s coming system like a chief of staff learning by sitting next to you, and then hovering over your shoulder with a mandate to remove friction from your day, Kothari says.

It’s a big swing for Wispr, which Bloomberg reported was raising more than $250 million at a valuation of $2 billion back in May. Previous investors include Menlo Ventures, Notable Capital, NEA and 8VC.

(“This was part of what motivated us to start raising some more capital to be able to fund this and get the best talent,” Kothari says.)

And it’s one that could put Wispr more on a collision course with OpenAI, which announced its own voice models for natural language interaction, GPT-Live, earlier in July.

“We’re going to realize that this phase of clicking things, and typing things, was never what we wanted to be,” Greg Brockman, OpenAI’s president, said when Upstarts asked if he, too, wants to build Jarvis at a media roundtable this week. “A couple years from now, no one’s going to want to do that stuff anymore.”

Why should Wispr’s new Lab be the ones to drive that new paradigm?

More on that below, but in a nutshell: Wispr thinks its models will operate better for these purposes; that its interface will benefit from more than just the models themselves; and that users will want a neutral partner to guide them across all the apps they use, including but not limited to ChatGPT.

When might we see all this? Kothari says Wispr’s mission to bring movie magic to the real world isn’t years off, but coming soon.

“We are actually now building the model that powers and gets to Jarvis,” says Kothari.

The rest of this post – focusing on what Wispr’s Lab is building, how it’s building it, and why it can compete with the AI giants like OpenAI – is for paid Upstarts subscribers. Take 20% off an annual subscription as part of our Summer Sale, and get more startup news and founder interviews for less than the cost of a sandwich each month.

Read more

As a practicing cardiologist – he still sees patients one weekend per month – Abridge co-founder and CEO Shiv Rao can speak from experience about doctor burnout.

“Nothing crushes my soul more than clerical work,” he says.

But it’s his wife’s experience as a patient that sticks with me more when he brings it up toward the end of our chat on the newest episode of The Upstarts Podcast.

The couple already had a healthy daughter when they found out about a rare disease in the family, and after genetic testing and some tough luck (we’ll leave that to his own words in the episode), decided to proceed with in-vitro fertilization to ensure it wouldn’t pass on.

For three years, Rao says they spent his “overpaid corporate VC salary” traveling around the U.S. from their Pittsburgh home for doctor visits. (“We’re so lucky and we’re so privileged that we could even afford that, and we don’t take it for granted,” Rao notes.)

“I’d be there, and the doctor, or the counselor, would leave, and we’d look at each other and be like, ‘What did they just say?’ Like, ‘What are we supposed to do?’ And I’m a doctor. I can’t even remember.”

The story has a happy ending – they now have 10-year-old sons, identical twins – but it speaks powerfully to what Abridge is trying to achieve: “Finding a way to help both the patient and family member, but also the clinician, feel like they can do and be their best, is really at the end of the day what this is all about.”

Of course, there are lots of business reasons to pay attention to Abridge. Founded in 2018 by Rao and two others with deep Carnegie Mellon ties, Abridge is a first mover among the newer generation of AI-enabled healthcare startups, serving more than 300 health systems that reach 250 million Americans.

Active in AI long before ChatGPT, Abridge recently partnered with Nvidia in its model development, and with Eli Lilly in pharma. It’s raised nearly $800 million to date, achieving a $5.3 billion valuation last year in a round led by a16z. A host of other firms, including Bessemer, CapitalG, IVP, Lightspeed, Redpoint, Spark Capital, USV as well as strategics from Alphabet to CVS Health Ventures and the Mayo Clinic are also on the cap table.

Initially taking on Microsoft’s $20 billion Nuance acquisition, Abridge competes in a big market identified as ripe for AI impact by seemingly everyone. A challenger, Ambience Healthcare, has raised hundreds of millions of its own from investors including General Catalyst, Kleiner Perkins, Sequoia, and OpenAI – and, perhaps controversially, also a16z.

Microsoft hasn’t gone away, and Epic Systems, the health records giant, isn’t turning a blind eye. Palantir has gotten into the hospital business. And in January, we wrote in Upstarts about how Anthropic and OpenAI have made concerted life sciences pushes.

Get 20% off for 1 year

Is Abridge still the Upstart, or at risk of being disrupted? I was eager to ask Rao about it, after last sitting down with him in person at my final Cloud 100 launch event for Forbes in the fall of 2024: a stacked panel that included Rao, Anthropic co-founder Daniela Amodei, Perplexity founder Aravind Srinivas, and Scale co-founder and now Meta executive Alexandr Wang.

Taking Microsoft head-on in the early days, Rao tells me, was a “YOLO move.” Against other startups, Rao argues that no other can claim the same breadth and scale. “In some ways, we’re still creating a new category, yet we are still in a category of our own,” he argues.

• Listen on Apple Podcasts

• Listen on Spotify

On this episode of the podcast, Rao talks about how an early bet on AI got Abridge going, before ChatGPT; why reaching scale in healthcare requires years of “eating glass” and “kissing frogs;” and why he thinks Abridge can help doctors and nurses feel like “superheroes.”

Plus, he shares his Upstart Moment: that “YOLO” decision to take on Microsoft, with his “one shot” to get it right.

Below, you’ll find my two favorite business lessons from Abridge for busy startup builders – plus my own perspective on Abridge’s battle.

• Speedrun at the boss

• Own your destiny (AI stack)

• Alex’s take: It’s not all about the money

A MESSAGE FROM OUR SPONSOR

Your AI is only as good as your data

Rippling AI is built on your live workforce data and permissions. So it gives you full visibility across your startup (on anything from headcount to burn rate) — and can take real action right inside the platform. Want to see how it can work for you?

Get Exclusive Access: rippling.com/ai

Speedrun at the boss

Around 2022, when Abridge was still just about 50 people, Rao says the startup walked through a “one-way door” to go after enterprise customers. That meant taking on the “huge bosses” of software working with them, and the obvious Big Boss was Microsoft and Nuance.

“It was a YOLO move. I think that’s why most VCs tell their startup to go down-market first, figure out PMF [product-market fit] with a small clinic, build something that the end user loves, and then swim over time upstream, and get that large system.

For a lot of reasons, including the amount of money we had in the bank, there was no other way. ‘Cause ultimately you have to get those big systems. That’s where 75% of the doctors in this country practice. If you’re going to build a big, impactful, legacy-leaving, generational company, you need to get there as fast as you can.”

That kind of move is ultra-high risk for a startup, Rao says now. Why? Buyers in big companies – be they CIOs, CISOs, or other similarly-titled execs – tend to talk to each other. They have WhatsApp groups.

“When they see something they don’t like, they tell each other,” Rao notes. Startups that blow it might get a chance to rebrand, or recap, and come back.

Abridge did make it work, though, and Rao says they did it by showing their full hand of all they were trying to build, and being clear about a much bigger thesis.

“It turned out being ourselves was such a counter-positioning advantage against Microsoft,” he says. “Just saying, ‘Hey, this is what our true north is; this is what we’re trying to build.”

The key slide in their presentation stressed that Abridge wasn’t about AI taking notes, but about a conversation layer that kept the technology out of the way. “There was something paradoxically profound about AI actually removing all of the other tech, or pushing it into the background,” Rao says. “People resonate with that.”

Rao’s startup advice: Take it all on, fast.

“I think some companies might feel that you can play a longer game, and I think on some level, you’ve got to play really long games, but the world is moving so quickly right now, that it’s hard to triage. It’s hard to say, ‘This is tomorrow’s problem’ anymore. Everything is kind-of today’s problem.

And you’ve got to be so forward in terms of making clear why you’re different, and in healthcare, how you’ll be a 10-year, a decades-long partner for them. The urgency, and the emergency, the level of existential crisis you get to at moments in a company, just force you to focus, and do the thing you have to do.”

Own your destiny (AI stack)

Abridge announced its partnership with Nvidia last month, with the two announcing they would train a new AI model to be used exclusively by Abridge for documentation and clinical decision support, per the Wall Street Journal.

When I ask Rao about it, he describes the partnership as answering a more basic question: “We’ve got this incredible proprietary data set. What else can we do with it?”

Abridge’s hope is that it can maintain a number of agents for each doctor, around the clock, without breaking the bank. That means Abridge has to be “very careful” about how it uses frontier models from the big AI labs, Rao says.

Get 25% off a group subscription

“On any given week, 40% to 60% of our model outputs are probably driven by in-house work,” he says. Abridge has put a lot of work into its model routing, orchestration and evaluations to ensure that it’s only sending the most difficult problems to expensive models.

“This is the operating system for every single AI company out there. If you do a really good job of that, you’re going to deliver a better product, but you’re also going to be able to compete with your P&L [profit and loss] as well.”

Rao expects the question of frontier model costs to only heighten as Anthropic and OpenAI go public and face other priorities. “The companies that can reach down lower into the stack, the ones that can really own and control their own destiny, are going to be able to separate.”

That said, Rao expects Abridge to partner more closely with the labs like Anthropic and OpenAI over time.

“The moment that we feel like we’re going against the grain of what is absolutely incredible about this moment – how fast things are changing, and how profound this platform shift is – if that ever happens, we’re screwed.”

Advances from the AI labs need to “feel like a tailwind,” Rao says.

“When they do new things, cool, awesome. We evaluate. We see where we should point those models, and then we’re delivering a better product experience on some level, but we’re also getting all of the edits and adjustments. We’re getting the feedback loops.”

The need to decide which tasks end up where, and which Abridge should distill or fine-tune itself, won’t go away. Startups building vertical AI tools, or AI-enable software, will always be riding the wave, Rao believes: “I think it’s going to be that sort of orchestration, or portfolio, approach to models that’s going to win the day.”

Alex’s take: It’s not all about the money

One company that never came up in our conversation: OpenEvidence, which provides more of an AI copilot that helps doctors keep up with medical information and answer clinical questions. OpenEvidence is currently higher valued than both Abridge and Ambience combined, reaching a $12 billion valuation in a funding round this year.

A mix of investors from each of those companies are also in OpenEvidence, another sign, like a16z being in both, that this market is one prone to overlap. Above all, it’s just really, really big.

It can be hard to wrap one’s head around just how much money goes to healthcare in the U.S. A category accounting for nearly 20% of GDP that is embracing technology should produce a number of big winners, not a winner-take-all market dynamic. OpenEvidence partners directly with medical societies and works with doctors directly, giving it more of a bottom-up go-to market approach than Abridge.

That means its reach can spread faster, helping to explain its higher price tag; it also puts it likely on more of a collision course with the AI labs. (Short-hand of ‘ChatGPT for doctors’ adds fuel.)

Abridge is playing a different game, and while it may grow total users more slowly alongside a smaller valuation, that also makes it feel a safer bet.

Join the WhatsApp group

If I’m an investor in either Ambience or Abridge right now, I’m likely feeling good about my chances of a successful exit, either as another Nuance-type acquisition, or eventually a public company like Veeva Systems, a life sciences software company I profiled in Forbes magazine in 2017, and which currently trades at a $30 billion market cap.

But venture capital is a Power Law business, and if you mention one or the other to an investor, or another founder, you’ll often hear a take on how they seem to be doing in comparison to the other. “Ambience is catching up fast,” a backer will say. “Abridge has widened the gap,” another will respond – “Surely, that’s why a16z lined up the money truck!”

In some areas of AI, the horse-racing style updates seem pertinent: engineers change coding tools fast, and who is winning in a given week can mean the difference between an acqui-hire or a $30 billion deal for SpaceX stock.

In this case, the competitive dynamics maybe sell the impact of this category short. Ultimately, these tools are intended to help fix a partially broken system, and deliver better care. Spend any time talking to, or even just listening to Rao in this episode, and I expect that you’ll believe it, too.

For years, CEO Eric Yuan avoided work travel — partly from personal preference, partly to push the use of his own video conferencing software at Zoom.

But when he did venture to New York, he would visualize his personal dream: ringing the bell at Nasdaq, on the day of his startup’s IPO. “I always told myself, ‘I’ve got to work hard,’” he says.

As we chat last week, Yuan is back in New York to celebrate Zoom’s fifteenth birthday. He’s loving the local energy. “Every time I’m in New York, I feel like we are in the middle of a new chapter. I feel my energy is up, and I want to do more things,” he says.

Our interview is a reunion. In April 2019, when Yuan’s dream finally came true and he rang the opening bell to take Zoom public, I was there, scribbling madly in a reporter’s notebook, for the first in-depth profile of Zoom, which ran in the next issue of Forbes.

When, as Covid-19 shut down in-person activities the next March, Yuan impulsively decided mid-Zoom interview with me that he would give away his software to every K-12 school in the U.S. (and in a bunch of other countries) for free, my resulting scoop went viral. It’s probably still the most-read story of my career. With Zoom on everyone’s mind, I followed that up with another profile, this time a Forbes cover story.

The pandemic is long over, and so are those heady days for Zoom’s stock. Its market cap sits under $30 billion, compared to nearly $140 billion at its peak.

But 15 years in, Yuan’s still running the show, meeting with customers and pinging engineers about bugs (more on that later). What motivates him, I wonder, when he’s achieved his dream — twice?

Now, Yuan focuses on work goals, he says: doing right by Zoom’s customers and employees, while pursuing positive impact in the world.

“My business goal is just to keep going,” he says.

He’s inspired by Jensen Huang, still calling the shots at Nvidia after 33 years, and Larry Ellison, still active in leadership at Oracle after nearly 50.

And he’s still fighting for you to use Zoom more, not less, in the AI era, even if that means personally annoying some of his engineers.

Why pay attention to what Zoom’s founder has to say? Our interview covered a lot of ground, with bits that should be relevant for startup founders, builders, and anyone curious about how one of tech’s most successful leaders sees AI impacting our society.

For startups, we cover Zoom’s current strategy; its approach to acquisitions; and Eric’s take on Anthropic and OpenAI.

For builders, we talk about why Zoom’s CEO still de-bugs code; its reinvestment in product-led growth; and how Eric thinks about AI meetings.

For society, we discuss the risk of deep fakes, where Zoom is working on a new product; how AI could spark more in-person interaction; and Eric’s bet on a four-day work week.

But let’s be clear: Yuan, despite a net worth estimated by Forbes north of $5 billion, doesn’t take himself too seriously.

“No matter what, the uptime, downtime, AI or not AI, just enjoy everything,” he says. “That’s the most important thing. This is in life, or startups.”

In the spirit of Yuan’s call to “enjoy,” we’re running this full interview for all subscribers, for free. If you enjoy it, we hope you’ll consider supporting Upstarts by purchasing an annual subscription: 20% off in our Summer Sale.

Zoom 2.0 (AI Edition)

The lightbulb moment for Zoom’s AI efforts wasn’t the rise of the transformer, or even the launch of ChatGPT. It came later, Yuan says now, when he and leadership realized how much faster some of the company’s engineers were shipping code using Cursor in early 2024.

“Looking back, I was not smart enough” to spot the trend sooner, Yuan adds.

Now, the focus is on what Zoom calls “completion”: a shift from orienting its products around the conversation itself (happening over video), to the business action or outcome discussed or agreed-upon in that virtual room.

It’s a “much bigger market opportunity,” Yuan argues, that means Zoom is investing more in capabilities to bring in more context before meetings, and to send updates to a CRM, or AI agent, after them.

Zoom has offered some of these tools for a while now, including a Granola competitor for note-taking called My Notes. (I agree to run both, in a bake-off of sorts. My verdict: Granola’s AI summary wins as more useful, and less corporate-coded; Zoom’s transcription is more accurate, although both struggle at times with the Chinese-American Yuan’s accent.)

Because Zoom comes from the starting point of a live conversation, however, Yuan believes that the company will only lightly overlap with partners like Salesforce, the owner of Slack, over time.

“Today, you cannot find any two companies where it’s very clean, there’s not any overlap,” he says.

Back to startup speed

Engineers who join Zoom expecting a slower, public company pace are often in for a bad time. When Zoom was a startup, much of its technical team had worked with Yuan for a long time, he says, and knew what to expect. Now, some complain that their CEO is pushing them hard. Directly, too: Zoom’s CEO says he engages directly with technical staff about issues, bypassing managers.

“It’s so slow, guys,” he admits he’ll say. “Competitors are shipping products. What are you doing?”

All these years later, Yuan still finds and reports bugs in Zoom’s products himself. “I’ll write a summary: ‘Guys, I only have 24 hours to myself. You have so many engineers, add up how much time you guys have. Why don’t you guys find those bugs? Why have only me find those issues?” It’s enough, he says, to drive him “crazy sometimes.”

Zoom’s recently-announced acquisition of Common Room isn’t about simply re-injecting some startup hustle into the organization, Yuan says. He pushes back on the notion that Zoom is looking to be more acquisitive now. “It’s hard to balance organic growth versus M&A,” he says. “We are still very disciplined.”

A near-miss acquisition validates that approach for Yuan. In 2021, Zoom announced its planned acquisition of Five9 for just under $15 billion. Five9 shareholders didn’t support the deal, and the two companies walked away that fall.

Get 20% off for 1 year

Instead, Zoom launched its own contact center solution recently; Five9 trades at a market cap of just about $2 billion. “Our progress is much better, and we didn’t spend so much money on that,” he says.

When I ask him why companies like Anthropic and OpenAI are seemingly buying up more AI startups, besides the disparity in equity value (with their private shares carried closer to valuations approaching $1 trillion), Yuan argues that the AI labs have more product needs across a wide surface area of infrastructure, hardware, models and agents, with more appetite for paying for talent.

Zoom, meanwhile, looks to university programs, up-skilling of veteran engineers, and to others who resonate with its mission in order to compete.

“No matter how much better you think a [pay] package is, somebody else will for sure offer a better one,” Yuan warns.

The gospel of PLG

The back-and-forth around me choosing to use Granola leads to a “pain point” for Yuan and Zoom: customers still think of it as “just a meeting company,” when it offers a much wider range of tools.

To change that, Yuan believes Zoom needs to get back to more of its product-led growth roots. My Notes grew to 1.5 million monthly active users in four months; that’s the type of adoption he wants to see more widely with ZoomMate, the AI companion the company offers in its productivity suite.

The challenge for Zoom: enterprise products and launches will still be driven by the sales team, and close engagement with big customers. Other launches need to inspire bottom-up new users. And, Yuan says, until recently, Zoom wasn’t systematically tagging each as one or the other.

“Now we try to have a formula for each product before we launch. We should say, ‘This one should start from product-led growth, or this one should start from a sales team approach,’” he says. “We should have a clear, defined go to market strategy.”

Get 25% off a group subscription

It’s the opposite challenge of many startups that lean heavily on PLG, then find out that it can only take you so far. They then scramble to build out enterprise tooling, customer success and forward-deployed capabilities to smooth over at-scale adoption.

Zoom, meanwhile, almost sounds like it’s fighting back against the Innovator’s Dilemma: “We thought inertia could help us,” Yuan says. “And that’s not right.”

For PLG products, Yuan believes cost and ease of use are most important. For sales-driven ones, trust and reliability.

“This is probably one of the things that keeps me up at night,” he adds. “We still need to address it.”

Digital twins, deep fakes, and the four-day work week

What Zoom is building towards above all – a selling point with those potential hires, Yuan says – is the concept of a ‘digital twin’ that Zoom can help power to work like us, and alongside us.

For the past three quarters, Yuan has used an AI avatar to read through the script of his earnings call, allowing him to join just for the live Q&A, the part he likes.

As another example, Yuan imagines an hour-long contract negotiation happening over Zoom. Most of that time – maybe 45 or 50 minutes – is spent running through the document and details, he posits; only the last 10 minutes or so is the actual live negotiating. Handing over the first part to our digital twins, we could spend more time conversing and building rapport or catching up, he argues, or moving on to other tasks.

Talking to AI every day is lonely, he argues, and most people will want an emotional connection with other humans. Delegating work to a digital twin, he hopes, will free up more time for that.

Yuan is skeptical about AI simulation models replacing those negotiations – he thinks the tech isn’t ready yet to replace us altogether – and he doesn’t think human workers will embrace mixed meeting environments where half a Zoom is people, the other half avatars.

Zoom plans to announce a product soon around deep fake detection, he adds, that will be able to verify human identities in a meeting, or flag when videos are AI-generated and not labeled as such. “We need to make sure to tell participants,” he says. “Otherwise, if you don’t manage this well, it’ll be very dangerous.”

The tech leader’s boldest prediction, though, comes around the four-day work week. Yuan insists it’s coming. How soon? A few years to five, he replies, comparing the looming moment to 1926, when Henry Ford adopted an official five-day work week for his factories.

A big benefit of new technology is that it can handle more work for us, Yuan believes – freeing up more time for us to travel, work on philanthropic projects, or in our communities.

What’s the point in working hard if you can’t enjoy yourself, Yuan asks rhetorically. “‘Enjoy’ doesn’t mean this is a waste of time, a waste of resources,” he elaborates. “No, this is to do more meaningful things.”

I’m skeptical that Yuan’s peers in public companies will follow suit. Or if not them, their competitors in China. Getting Wall Street investors onboard seems like a challenge, too.

Yuan goes back to Ford, noting that in other countries, the five-day work week eventually caught on, too.

“It’s going to happen,” he insists. Will he lead by example, then? “I should lead by example, right?” he replies.

We’ll have to circle back in another five years, to check in on how Yuan is spending his Fridays.

It’s tough to keep a big startup valuation under wraps for too long. Filings or investors might disclose it; reporters will sniff it out.

But founder Eléonore Crespo has a rare distinction: her startup, Pigment, had its funding milestone scooped by an unlikely leaker: the President of France.

“One more French unicorn!” Emmanuel Macron posted on LinkedIn two years ago. (Naturally, he said this in French, so the exact word was “licorne”.)

And given that Pigment is one of France’s most prominent AI startups alongside Mistral and Dataiku, I was excited to ask Crespo about that experience. What was her reaction?

“I was like, ‘So be it,’ you know?” She tells me in our latest episode of The Upstarts Podcast, recorded from London. “We didn’t want to announce anything, because we found that it was not necessarily the milestone we wanted to celebrate.”

Crespo laughs now. “Thanks to the government, I’ve met so many of the top CEOs in France that are now buying Pigment.”

As a community, AI startups in France have been “very lucky” to have that support, Pigment’s co-CEO says. “But at the same time, I’m here to create a global success from France,” she notes.

That ambition is a big part of what makes Crespo and her startup more interesting than you might think from just scanning its description: “business planning, powered by agentic AI.” Or translated again, a new crack at an old category of software, financial planning and analysis, or FP&A, taking on the likes of Oracle, SAP and Workday.

Founded in 2019 by Crespo and Romain Nicoli, the former co-founder of Criteo, Pigment counts tech darlings including Anthropic and Notion as customers, alongside corporations like Coca-Cola and Unilever.

The startup’s AI agents help finance teams at those businesses to make important financial decisions faster by reducing the amount of time they spend collecting, managing and pulling data from spreadsheets. (A problem that Crespo encountered first-hand while working at Google.)

More interestingly for someone not working in financial planning, Pigment is also starting to help its customers rein in runaway AI budgets, better managing their token and model spend. On the show, Crespo and I talk about Uber, which made headlines when one of its executives revealed in May that the company had blown through its annual AI budget in just four months. Uber is a Pigment customer, but not for AI spend tracking yet.

“I would love to build something with them now,” Crespo tells Upstarts. “It’s going to be a big topic for everyone, and it’s a topic that goes beyond just understanding your costs.”

• Listen on Apple Podcasts

• Listen on Spotify

In our conversation, Pigment stands out for other reasons: it maintains an unusual co-CEO structure, with Crespo’s co-founder Romain Niccoli building an engineering team in Paris that Crespo calls “probably one of the most loyal in tech history,” with zero regrettable attrition to date.

Its Paris roots also help it compete for talent with what Crespo compares to a “Chinese buffet,” where AI heavyweights like Anthropic and OpenAI pass through and grab all the steak and lobster first.

Get 20% off for 1 year

But most interesting to me is Crespo’s policy of “unreasonable hospitality” – a customer-obsessed mindset that is reflected, she says, in everything from late-night WhatsApp messages to personal touches like baked goods: she recently visited Meta with cookies featuring dual Meta and Pigment logos.

More on those big takeaways – including a guest-starring role by Tom Cruise – below. Plus, per reader feedback, we’ll try something new: signing off with my own personal thoughts.

• How to make a co-CEO set-up work

• Competing at the AI talent “buffet”

• Using “unreasonable hospitality” to win

• Alex’s (armchair expert) take

A MESSAGE FROM OUR SPONSOR

Your AI is only as good as your data

Rippling AI is built on your live workforce data and permissions. So it gives you full visibility across your startup (on anything from headcount to burn rate) — and can take real action right inside the platform. Want to see how it can work for you?

Get Exclusive Access: rippling.com/ai

How to make a co-CEO set-up work

Before Pigment got started, Crespo ran a painstaking process to identify her ideal co-founder. She knew she wanted an experienced enterprise leader who could build an engineering team, and Niccoli stood out as a former co-founder of Criteo, which went public in 2013.

Crespo spoke to a number of mutual contacts about Niccoli before first approaching him: “I did my own CIA due diligence, and he had no idea.”

To win him over, she brought a 100-page memo book of potential projects to show him. After three afternoons together around Christmas, he told her, “Let’s go.”

How do they make co-CEO status work in practice? The key, Crespo tells me, is organization. “He leads tech, product, and HR. I lead everything else, and we are extremely clear about that.”

The duo have a policy of making tactical decisions on their own, and syncing up on big decisions. For key strategic product updates or company milestones, they’ll spend two hours together in a workshop to reach agreement.

“We have this concept internally of direct ownership, or DRI [directly responsible individuals, an Apple-coined term], to make sure that there is always a single individual that is an owner of a specific topic.”

Most recently, Crespo and Niccoli shared several calls to discuss the right criteria, and compensation package, for a potential C-suite hire. But when it comes to fundraising, Niccoli defers to Crespo, who previously worked as an investor at Index Ventures.

“That’s the type of thing he doesn’t care about,” she laughs.

Competing at the AI talent “buffet”

Building an engineering team in Paris has proven an “incredible edge” for Pigment, Crespo argues, because France currently produces high-quality, AI-trained engineers, who are loyal and mission-driven.

“It’s something that would have been impossible for me to replicate in the Bay Area,” she says, noting that she was recently talking to one of Intercom’s co-founders, and they compared competing with Anthropic and OpenAI for talent to a “Chinese buffet.”

“You have lobster, and you have, whatever, let’s say a good steak, and then you have the noodles, and then you have the rice. And so you come, and Anthropic and OpenAI, they take the lobster and the good steak. And there are only some noodles left, and you’re like, ‘Oh my God.’

That’s what happens today, and it’s very, very, very hard not only to compete, but to retain the talent. So for us, building in Europe has been an incredible strength.”

Maintaining offices in France and the U.S. has also helped Pigment “think globally” from day one, Crespo says, with employees from a number of countries – far more than just those two – and a mindset she argues is worldlier than some Silicon Valley peers.

That said, Crespo spends a lot of time in San Francisco personally, and plays nice with both Anthropic and OpenAI as partners.

“We spend a ton of time with their teams, and we will keep doing so for the foreseeable future, because it’s true that’s where a lot of things are happening today.”

Using “unreasonable hospitality” to win

To Crespo, “unreasonable hospitality” means being memorable.

“When we come to meetings, the way we show up, we deliver the experience of when they buy Pigment. We make it extremely memorable, and we spend a lot of time trying to really deliver delight there.”

For customers, that means thoughtful, tailored swag, or the cookies like Crespo recently brought to Meta: “We try to create some bonds that go way beyond the prospect, that can create a friendship over time, I would say.”

A “fantastic” customer experience also means Crespo personally responding to WhatsApp messages on the weekend, or late at night.

“I will always answer within the next five minutes, and it’s the same for the rest of my team. We are extremely reactive. We try to fix issues, and we never push back. We really try to be there all the time, and we always do the extra mile.”

That means hopping on planes whenever a customer has a more complex problem, too.

Working with Pigment should feel like buying a Ferrari, Crespo argues, not a mass-produced car. Like Ferrari, that means it remains up-market; such an approach wouldn’t work with SMBs or many thousands of customers, Crespo admits.

“It’s not a mass play here. It’s a play where we serve the largest companies in the world with a solution that is tailored for them.”

That white-glove experience also extends to closing key hiring candidates. One of Crespo’s biggest challenges has been hiring a chief revenue officer, she says; she met with dozens of candidates. When she found the right one, she was eager to find an “unreasonable” touch to close them.

As it turns out, one of her investors, ICONIQ, works with a number of celebrities through its family office practice, and Crespo was invited by founding partner Divesh Makan to a private 90 minute meeting with Tom Cruise, who was going to talk to portfolio founders about filmmaking and what it took to be a standout actor.

“I begged Divesh, ‘Can I bring my candidate? I think it would make a big difference,’” she remembers. Makan agreed.

“It was quite a lifetime moment for [the candidate] and for myself,” Crespo says. “That’s how we managed to get him.”

Alex’s (armchair expert) take

Pigment’s technology isn’t as high-risk as a dog longevity drug or a new system of nuclear reactors. Given its customers, the fact that it partners with the big AI labs, and the straightforward nature of its software, it seems a safe bet to continue growing and build a lasting business.

What is less clear to me is that Pigment will be a $100 billion company versus an attractive acquisition for one of the last-gen giants it’s taking on. I wish I had asked Crespo straight-up: how big do you think this can get?

But at the same time, founders are wired to respond to that the same way: boundlessly big. It’s hard to reach a $5 billion or $10 billion acquisition outcome if that’s what you set out to build at the start. And if that’s Pigment’s eventual fate, it would be an excellent one worthy of celebration. (One dark horse outcome would be that it joins Anthropic or OpenAI directly, gaining a permanent seat at the “buffet”.)

I’m also satisfied with how Crespo presents her co-CEO setup, although I usually am skeptical of such an arrangement (‘too many cooks’). Crespo can bring the energy, public face, customer-facing sales and service, and fundraising; meanwhile, the more veteran, proven co-founder can focus on going deep on product and tech.

It seems like a good match, in large part because of how carefully Crespo approaches it. Generally speaking, I’d be cautious about startups following suit – it seems to carry a lot of risk if you don’t match the right levels of ego.

One place where I’m not sure I fully buy Crespo’s optimism is when I ask her in the episode about Pigment’s AI’s impact on jobs. Crespo concedes: “Probably there are companies that are going to reduce headcount,” as well as rethink processes.

But overall, she takes the techno-optimist angle that most customers have more financial planning to manage than they can handle, and that analysts at customers welcome an AI agent freeing up some of the 80% of time they spend on busywork.

I’m optimistic in many ways about the technology itself, but less about corporate interests and public company executives. If the incentives set by shareholders encourage companies to instead cut headcount and depend more on AI agents for this kind of work, it’s easy for me to picture a CFO who manages an increasingly agentic team of analysts instead.

The Deep Dive

Adapter CEO Adam Ghetti didn’t expect to be building another startup after selling his previous venture – the product of a decade’s work – to Twilio five years ago.

Then the transformer emerged, and many of the smartest people he knew in tech started racing towards its obvious applications: text generation and large language models.

“It was obvious to us that 10,000 other people were going to go do the content thing,” he says.

For Ghetti, those “flashiest” uses missed the most important questions: how could we use technology to truly understand our own digital footprint, our own precious data? More importantly: if we got there, how would we control those treasure troves, so that we weren’t putting the big AI labs like OpenAI in charge of our own “personal worlds”?

“Ten years from now, could I sleep well knowing I could have helped and didn’t? I couldn’t,” Ghetti says.

Then, as he puts it, a prominent computer scientist focused on data science, professor David Bader (formerly of Georgia Tech, now at the New Jersey Institute of Technology) and a VC firm that had previously backed him, GV, called his bluff. In November 2022, just days before the public launch of OpenAI’s ChatGPT, they got started on a new startup: Adapter.

Now, after three and a half years of operating in stealth, Ghetti is unveiling Adapter to the world. Based in Austin, Texas, the startup employs 17 people and has raised $17.8 million in capital to date from GV, Bond Partners (an early-stage fund within BOND), and a number of other investors including Eric Schmidt’s family office Hillspire, Byers Capital, the funds of former Lookout founder Kevin Patrick Mahaffey and Shana Fischer, and personal investors including Paul Judge, Zach Sims and Ted Schlein.

What they’re betting on isn’t a new AI model or app, but what Ghetti calls ‘Cognition as a Service’ – a new-look infrastructure layer that is intended to better leverage and control your personal and work data for use by AI agents and applications.

Adapter is making available a developer tool and API, called Adapter Mind, intended for developers and tinkerers to start using immediately to connect to existing applications, and to spin up new ones. Users can drop Adapter’s model context protocol, or MCP, into whatever tools they prefer, like OpenAI’s Codex, Anthropic’s Claude Cowork, or Cursor.

Adapter says it will continue to set up Slack channels and ship best-practices and starter kits in a public library to help early adopters test the limits of what it can do in upcoming weeks.

It’s also shared with test users a consumer-facing harness, called Adapter Life, as an example of a tool that can build upon Adapter to operate like a super-charged chatbot or agentic assistant, communicated with via iMessage or WhatsApp.

“Adapter is designed to adapt to your life, your workflow, your needs,” Ghetti says.

But Adapter is explicitly not a new model company or agent builder itself.

Instead, it’s looking to establish new ‘primitives’ that Ghetti compares in potential impact to Amazon Web Services’ release of S3 online data storage, which helped to kick off a new era of businesses from Dropbox to Zoom.

“We don’t want to produce many things, but the things we produce, we hope will enable a whole new world of new things to do,” he says.

If that sounds wide-reaching, even amorphous, it is: Adapter is an ambitious, big-swing venture that looks headed for binary outcomes: kicking off a whole new wave of AI capabilities and startups if it succeeds; failing to land a killer use case and collapsing in on itself if it does not.

But Adapter is already provocative across three hot-button topics in AI today:

• the token economy and efficiency of AI spend

• capabilities and true ROI from AI tools

• transparency and data sovereignty

On the cost front, Adapter claims its knowledge graph, in large part because it lives online using CPUs, not costlier inference-friendly GPUs, is much cheaper to set up and maintain than spending on model tokens.

From a capabilities perspective, Adapter and its advocates say its graph unlocks context currently lacking even from the leading AI interfaces, because it’s not context limited.

And on the privacy and control standpoint, Adapter says it’s a better “custodian” of your data because it doesn’t access or train models off of it, and because Adapter can then parcel out that info to the model labs and wider web on a need-to-know basis, keeping their hands off, too.

Of note: Adapter’s CEO says it currently partners with a number of the big labs behind the scenes, but explicitly not OpenAI. (“Our entire mission is built around trust, transparency and sovereignty. OpenAI does not appear well aligned to these,” Ghetti says.)

Here at Upstarts, we spoke with several backers and users of Adapter’s tools – and went through the onboarding process ourselves to see it first-hand. Our impressions, and Adapter’s efforts in each of those three areas, below.

Interested in supporting more exclusive, in-depth startup Deep Dives like this? Take 20% off an annual subscription for a limited time through our Summer Sale.

If your business would like to sponsor a future Deep Dive, you can drop us a line here.

The prime directive

To sign up for Adapter, you give it access to a personal email initially – the idea being that this hub of your data is long-lasting, perhaps more so than any particular job – and then connect a bunch of tools to it, like Google Drive, Notion and Slack. The savvier can set up their own custom connectors to access photos or old file repositories; the more you feed in, the smarter it gets.

Ingesting it all takes a while – for me, 15 minutes or so – and then you can start asking questions. In a demo, Ghetti demonstrates Adapter correctly identifying the earliest touchpoint between us, more than ten years ago when I was at Forbes, despite there being no obvious article link or paper trail of that moment. He’s also used Adapter to write investor updates that know the depth of his relationship with each investor; on the personal side, Adapter Life (his is called Ghetti Life) knows everything about his kids’ pediatrician visits.

Mahaffey, the former Lookout founder and an early backer and user, asks Adapter to serve as a research assistant to update him on any developments in geothermal energy, a current area of focus, as well as for reviews of mountain e-bikes using a new motor he finds promising.

Erik Nordlander, a general partner at GV who previously backed Ghetti’s last startup, Ionic Security, has used it to triage his inbox. His Adapter flags qualified experts and executives in his network to help answer inbound requests from founders, beyond what he says he can do in LinkedIn.

Adapter’s able to do all of it in large part because its system takes the time to understand the “prime intention” behind any task or query, Ghetti explains. (The company has a patent on its prime directive work.)

When using Adapter directly, or when another tool calls the Adapter API, it can then run as few or as many loops of searches as needed to get the answer it believes you need, at the depth it predicts you want. The more you use it, the better your Adapter Mind gets at balancing that.

Unlike an LLM, Adapter doesn’t forget context over time, or inject unwanted context from a past conversation. And unlike many agents that load up a prompt behind the scenes to re-search for all the relevant information, once Adapter knows something, it stays that way.

“You can do this with all the questions you wouldn’t typically think to ask an AI, because it has nuance,” Ghetti says.

Pricing power of CPUs

How is Adapter doing it? The big breakthrough is the knowledge graph, explains Bader.

Bader’s government-level work and Ghetti’s experience in data security allowed them to build with less than $20 million a system they argue might take a corporation hundreds of millions to set up.

Adapter’s graph only has to read your data one time; then it maintains it on a secure, always-updated private website hosted through Cloudflare. Adapter continuously updates that information as new data points (an email, a Slack message, a sale) come in, but once Adapter knows something, it doesn’t necessarily need to call an LLM or search the wider web, only doing that for “truly novel” questions.

That allows Adapter to ingest many gigabytes of data and sift through much of it for just one query, far exceeding the context windows of most AI models, Ghetti argues. And because the data is in memory, performant on CPUs, they can do much of that searching locally, on device, and using smaller parameter models.

The hope for Adapter is that it can then pass on much lower costs to customers, making it a cheaper and more efficient building block for other applications, and making it more practical for prosumer users. (In my personal use case of managing email inboxes, I’ve hit token limits asking AI agent applications to ingest and label the high volume of PR pitches I get.)

It all adds up to “bell curve” cost to unlock “log curve” value, Ghetti argues, which means Adapter is cheaper to operate the longer it works with a customer. “The delta from ‘never’ to ‘now’ is a lot, but the delta incrementally between ‘now’ and ‘what’s next’ is very small.”

Guardians of your galaxy brain

In an era in which AI labs and advertisers are both hungry for your data, Adapter’s supporters describe it as a third option in a brewing battle between secure walled gardens of data that must be entrusted with one company and not shared widely, and open ecosystems where your privacy and data aren’t secure.

Adapter users who access models from Anthropic, OpenAI or elsewhere don’t share their specific data with the LLM, Ghetti notes, because Adapter sends the models an aggregate stream of traffic.

Mahaffey compares it to a data ‘Switzerland,’ sharing your data on a need-to-know basis that could serve as a compromise moving forward for advertising and agent-to-agent communication.

Imagine a model or site communicating with your Adapter graph to share a few versions of an ad, he posits. Instead of the company scraping all of your data and web behavior to target the ad to you directly, Adapter could respond with what ad might resonate best, without sharing any additional info.

“We need a world with less noise, not one with more,” he argues.

At GV, Nordlander argues that even the tech giants can find value in such infrastructure: “I don’t think in OpenAI’s wildest dreams they think you’re going to be doing all of your productivity stuff, and your social stuff, in the model.”

Ghetti, for his part, sees himself as a ‘custodian,’ or protector, of your data. When I note the irony that a startup founder is effectively asking people to trust it – and him – with that data in order to secure it, Ghetti responds that users can verify their own data security cryptographically with Adapter’s documentation.

He’s also exploring legal ways that Adapter can establish itself as a long-term good actor more structurally, and says he has no plans to ever cede board or equity control. “You should have sovereignty over your coal mine,” he argues.

Adapter’s CEO says he’s confident his startup has found a way to build a successful business model ensuring all three.

For it to work, it needs one thing above all: people to try it out.

“When people feel like the information and understanding asymmetry in their life begins to balance in their favor — not just because they come to understand more, but because they come to achieve more simply — when that is true, we are winning,” Ghetti says.

Materials science hasn’t changed much in 150 years. Enter Radical AI CEO Joseph Krause.

In a big new lab in Brooklyn’s Navy Yard, the startup founder is building a “next-gen” approach that he says makes other autonomous labs look like hand-free driving mode on a car, compared to a fully self-driving Waymo.

It’s an approach that plenty of corporate scientists tell Krause must be too good to be true. “How could you really discover more alloys in a week than we have discovered in our company history?” he says they ask.

But since its founding in 2024, Radical AI has already produced valuable, and secret, alloys for customers, Krause insists. And the startup has done it after raising just $55 million in its last funding round, not the $1 billion of certain well-heeled competitors.

We’ve written about other material science startups at Upstarts already, including our scoop of Apple’s former AI chief joining CuspAI. Another, Periodic Labs, was reported by Forbes to be raising $500 million at a valuation of $7.5 billion in May. Radical AI stands out not just because it’s based in New York, but because it’s arguably the only startup that isn’t outsourcing the actual lab work to partners.

Krause is also a straight shooting, no-nonsense founder who makes for a lively interview. Will he ‘win’ this category when DeepMind alumni, Jeff Bezos and others can also raise huge sums? It’s hard to definitively say. But it’s easy to see Radical AI competing, and thriving, in this fast-evolving category.

• Watch on YouTube

• Listen on Apple

• Listen on Spotify

On The Upstarts Podcast, Krause explains how a SpaceX breakthrough helped kick off a category; why cloud labs and deep tech companies often fail; and why concurrent engineering is as important as next-gen lab equipment itself.

Plus, he shares his Upstart Moment: a 100-year-vision to become one of the world’s most important companies.

Our three biggest takeaways for busy builders can be found below.

• Look for ‘something more’

• Embrace resource pressure

• Stick to your guns

A MESSAGE FROM OUR SPONSOR

Your AI is only as good as your data

Rippling AI is built on your live workforce data and permissions. So it gives you full visibility across your startup (on anything from headcount to burn rate) — and can take real action right inside the platform. Want to see how it can work for you?

Get Exclusive Access: rippling.com/ai

Look for ‘something more’

Krause was in a PhD program and working as a fellow at the U.S. Army Research Lab when he cold-emailed 100 VCs about in New York about working with them to invest in materials science.

Kevin Ryan, the serial entrepreneur and AlleyCorp founder, offered him an internship.

“This AI wave was taking off, everyone was investing, and there was pressure to deploy dollars in really exciting companies that are going to be the big companies of the future. “

Reviewing AI pitches with one of his Radical AI co-founders, Jorge Colindres, that Krause got the idea for Radical — out of frustration.

“We were just like, ‘How many more recruiting pitches can we get to ‘redefine the recruiting industry,’ right? I'm sure it's a good problem. Nothing against those businesses, but we were looking for something more.”

They thought about ideas in materials, robotics, manufacturing, and physical industry, but materials science seemed the best chance to have a “fundamental impact” on humanity.

“Regardless of what industry you care about, automotive and aerospace, manufacturing and defense, climate, energy, semiconductors, electronics, the most important industries in the world are all a direct result from materials R&D.

So to us, it felt like the perfect blend where you have this technology that is completely changing the paradigm in discovery, and you have this field that's 100-something years old and hasn't had it happen yet. That's a place that we should attack.”

Embrace resource pressure

Radical AI’s most recent funding round, a $55 million seed announced a year ago from RTX Ventures, Nvidia, AlleyCorp and others, seems quaint compared to some of the billion-dollar raises for pre-launch companies in AI.

When I mention this to Krause, he first stresses that Radical AI’s competitive edge doesn’t come from automating tools, or navigating a robotic arm, but from the interdisciplinary team of scientists he’s put together to run such experiments, capture the results, and build AI systems for analyzing them on top.

 “The connection and the collaboration between all of those units is imperative. It is not impossible — we did it — but it takes serious time to really build it. You cannot get interdisciplinary work unless you build interdisciplinary products. So the timeline to even just catch up to us is two, three years, and we are compounding that more and more, the more technology we build on top of it.”

Still, I ask Krause straight-up: if he could have raised $1 billion to start, would he really have turned it down?

“ I think it's dishonest to say that more money is not helpful. It is,” he replied.

Money means hiring faster, bigger facilities, and more data collection.

“In the material science space, we do not have experimental data sets, and so the more experiments you can run, the larger the data set gets, so the better the AI models perform. The number of experiments you could run is correlated to the amount of capital you have, because of the tools that you can buy.

So more money's helpful, but you do not need a billion dollars to start this. As a matter of fact, we didn't have a billion dollars, as you know, and yet we've still proven the self-driving lab works.”

Krause says he doesn’t want a billion dollars today, because he doesn’t need it, and it keeps his team focused.

“I'm a big, big believer as an entrepreneur in ‘back against the wall.’ When you get comfortable is where great ideas go to die, because you have no incentive and no chip on your shoulder to execute.”

In the future, Radical AI will likely need $1 billion, and it will raise it. But for now, it would be a distraction, says the CEO.

”Having pressure to deliver is a very healthy thing.”

Stick to your guns

At this point in Radical AI’s journey, Krause says he and his team are confident that they’ve found the “correct way” to build in materials science. It will take time for the market to agree, however, in what he calls Radical’s “capital-efficient, but very directed, approach.”

 “We don't want to boil the ocean in materials science, because you'll never drive any value, and then you'll never prove the system drives value.

So for us, now it's about sticking to your guns, not being persuaded by VCs, by the market, or even by what talent is telling you. Staying true to the mission, true to the execution plan, and driving against that.”

According to Krause, this is a particular challenge for deep tech companies, specifically.

“They go to market, they hit a wall, and then they quickly try to start moving around that wall with a lot of input from the outside world.

Not at all do I think you shouldn't pay attention to what customers or what people are saying. I think you can't be influenced by what people are saying, either, particularly if you are confident you're right. It is making the bet that you are right that typically pays off in the long run.”

GetYourGuide CEO Johannes Reck has a love/hate relationship with Google.

The search platform is still a key source of new customers for GetYourGuide, the still-private online marketplace that Reck co-founded with three others in 2009; even in the age of AI chat tools, Reck doesn’t see that changing any time soon.

But then there’s all that Google doesn’t do: such as seem to care much about whether a travel ticket site is official, or a tourist trap paying the tech giant to appear as such.

Take the Sagrada Familia church, which I visited earlier today with my spouse’s family here in Barcelona. Tickets to see the recently-completed final tower, which opened in February, sell out weeks in advance.

When I meet with Reck in New York earlier this spring, I’m already on the hunt, running into posts warning that many of the top ticket pages on Google aren’t official, but scams.

“We have a lot of beef with Google with it,” Reck scowls. “We believe that they should be filtering this, and that it’s not in the consumer interest.”

Small comfort: Checking back today after snapping photos atop the tower, I notice that the seemingly most prominent offender has been taken down and de-listed from Google; perhaps Reck’s visit last month during Pope Leo XIV’s official visit, during which the church recognized GetYourGuide as one of 15 official collaborators, has paid off.

But it’s a problem that persists across the world’s most popular tourist sites. A few days ago, in Rome, I have to wade through sites offering up Galleria Borghese tickets; ironically, the official website is broken that day, leaving only scalpers.

I buy three entry tickets from one; one ticket from another. The last two of our group — including my brother-in-law Zach D. Roberts, whose excellent photography you’ll find below and on his website — I book onto a tour they immediately ditch so we can sync up.

And it’s one that seems like it could easily get more challenging in the AI era. When it’s easier than ever to marshal armies of bots to make listings, simulate visits and skew site traffic, what will appear bona fide, and what AI slop?

At GetYourGuide, Reck says he’s optimistic about the future of AI-assisted travel.

“There are still a lot of question marks in the market,” he concedes. “But right now, I think it’s a huge accelerator.”

A few factors underpin that confidence.

First, there’s what AI can do to help make an overwhelming number of listings – GetYourGuide currently shows about 200,000 experiences across 50,000 supply partners to an audience of about 33 million travelers each year – feel more tailored and personalized.

Then, there’s the competitive dynamic: GetYourGuide and similar sites like Viator and Headout have databases of proprietary info that an LLM can’t easily scrape from the web.

Zooming out, there’s how the travel booking experience evolves alongside AI tools. Reck envisions a world where travelers talk through experiences with AI in a more haptic, visual experience.

Below, we break down how GetYourGuide is leveraging AI, threatened by it, and the “geeky” future its CEO envisions.

• AI horses, human guides

• Resisting our AI overlords

• Travel’s TikTok-ification

We get into the risks, too. But Reck is optimistic: “The good news is that agents can’t really go on the trip for us.”

Those insights are for Upstarts paid subscribers only. Take 20% off an annual subscription for a limited time, as part of our Summer Sale.

Take 20% off

Read more

As I sit down to meet Brian Elliott, co-founder and CEO of Boston’s new startup unicorn Blitzy, I’m excited to ask him about his experience as a U.S. Army Ranger.

Not what it was like to go through Ranger School, the famously grueling program that has been the subject of TV shows. I want to know about something more hardcore: Elliott’s LinkedIn profile.

There, more prominent than any word about his combat deployment in Afghanistan, or his Bronze Star, Elliott shares that he “managed $30M in assets” as chief of staff of a 140-person team with, he adds, 10 direct reports.

There’s something so perfectly LinkedIn about it: the framing of elite, brave service to America as valuable within the corporate ladder.

When I get to ask Elliott about it, he tells me he’s been fascinated by systems – how things work – since he was a kid. Even a ‘direct action raid’ involving ground and air movements is, to hear him tell it, a “very elegant system.”

“Synchronizing all of that is IRL orchestration, and I loved it,” Elliott says. “Because precision, and operating under stress with precision, mattered more than anything”

It’s an answer that any ‘LinkedIn Lunatic,’ would approve of, but it’s helpful in understanding how Elliott, who co-founded Blitzy with friend Sid Pardeshi in 2023, has been able to build it into a promising business valued at $1.4 billion so fast.

Blitzy’s agents can understand 100 million-plus lines of code, helping corporations to automate and update systems that would otherwise take millions of dollars, paid out to consulting firms over months, to barely crack.

“For eons, we have been limited by how much a human context can hold in their brain,” Elliott argues. Now Blitzy can offload that context to its AI systems: “We can do changes at a size and scale that were previously impossible.”

• Watch on YouTube

• Listen on Apple

• Listen on Spotify

On this July 4 edition of The Upstarts Podcast, Elliott shares how he built Boston’s newest tech unicorn by becoming a CFO’s friend; why Cursor and Claude Code only see enterprise code through a straw; and what West Point and the Army Rangers taught him about high-stakes performance.

Plus, he shares his Upstart Moment: catching the 6am train to New York to try to close a customer, with his co-founder’s visa at stake.

Our three biggest takeaways for busy builders can be found below.

• Follow the ‘dissonance’

• A CFO’s ‘best friend’

• A Ranger’s lesson

A MESSAGE FROM OUR SPONSOR

Your AI is only as good as your data

Rippling AI is built on your live workforce data and permissions. So it gives you full visibility across your startup (on anything from headcount to burn rate) — and can take real action right inside the platform. Want to see how it can work for you?

Get Exclusive Access: rippling.com/ai

Follow the ‘dissonance’

The best time to build a startup, Elliott argues, is when “you have a view on the world that everybody tells you is incorrect, and despite that, you know it’s true.”

For Blitzy, that was the conviction that large language models, for all their capabilities, would not grow into efficient tools for updating large-scale code bases like the one he’d seen at what’s now called the Department of War.

“We would go everywhere and be like, ‘The model's not gonna do it all. Inference-time compute is the most important scaling law for code quality.’

This is before reasoning models, way before reasoning models were a thing, right? And if that was true, the limit case was gonna be understanding the underlying code base.”

This was not, Elliott says now, a popular position to have in 2023. “Everybody told us it was the stupidest thing they’d ever heard.”

 “They just thought that the model would solve the problem, right? It was like, ‘Oh, if GPT-3.5 can't do it, then GPT-6 will do it.’

It's not true. It's a physics problem, right? But these models are so amazing, it creates this illusion that it will be able to do everything, because you can experience GPT-3, you can experience GPT-3.5, right? And then you extrapolate what that's going to mean for size and scale in a way that's very logical, but it's not rooted in the reality of what a transformer is, right?”

This view informed how Blitzy set up its software: to run continuously, without a human in the loop, for weeks. The big unlock was what Elliott calls a knowledge graph, or orchestration system, that could handle millions of lines of code — not the thousands of an LLM’s context window.

 “What we're doing is we are removing the pressure or the load of having something... As a human, it would be like, ‘What do you kind of remember top of mind?’

So we're removing that load off of the agent, and we're only putting in what they need, just in time, every time, dynamically, done hundreds of thousands of times. And so at any single moment in time, an agent can only do one thing. We parallelize work, and then we sequence work based on how it can be sequenced.”

A CFO’s ‘best friend’

If you listened to our episode with May Habib last week, this one might sound familiar.

But Blitzy also considers itself a CFO’s ‘best friend,’ and Elliott is also bearish about ‘token maxing,’ which Blitzy’s CEO compares to evaluating a sales rep ‘SDR’ by the number of phone calls they make.

“It could be related. Sometimes it's related. But really you should be counting an SDR by bookings, right? That is their incremental value to the company. So, if an SDR created a script that dialed a billion people, that's not necessarily good.

That's my general view on token maxing: It's not sufficient, although it is not an unreasonable way to think about who is using AI. It's really not a sufficient way to think about adoption.”

Blitzy keeps its work usage-based, but aligned with a company’s existing roadmap, Elliott says.

“You're like, ‘Literally, I am buying these projects.’ What are the jobs to be done? These projects, way faster, from Q4 brought into Q2. That's what I'm buying, and there's incremental value both in labor savings and top-line revenue if those things are generating product value.

That is against the fixed amount. It's gonna be roughly X lines of code, right? So there's no variance swing from the CFO, like they're experiencing when they just buy random licenses for other AI products. What you're doing with the other AI products is you're throwing Opus 4.7 at everything, you're on 'max think' mode. And you're just putting that on loops, right?”

Blitzy’s value-add rests in part with the fact that it selects the best model for each task in real-time, Elliott says.

“Might be Gemini, it might be Flash, it probably won’t be Flash, but it might be Opus, it might be Sonnet, right? It might be an earlier version of Sonnet, right?

It’s baked into how we build our product in a way that really resonates with the CTO, the CIO, and the CFO.”

A Ranger’s lesson

When I ask Elliott what his top lesson across West Point, his active service and Harvard would be, he invokes a slogan from the Rangers: sua sponte.

"It means ‘of your own accord,’ and for the tech world, this is high agency,” Elliott says.

 “You will not succeed in the regiment if you're not an extreme sua sponte leader, because you're often on the ground, and you don't have comms, you don't have information, et cetera. You just have to make decisions live, based on understanding the larger intent. And you're not asking for permission. You are operating against a designed intent and mission.

That's exactly what it's like to be an entrepreneur, right? You are operating, and you are owning the outcomes of those decisions, and you're doing so at an incredibly high pace.

It goes all the way back to the conversation of: ‘Action begets information.’ Information doesn’t beget action, right? And so the ability to move with incredibly high agency translates from a prior life to this current life.”

When Elliott says this, it reminds me of startup founders who’ve said they hire especially from the ranks of elite collegiate athletes, who can combine initiative with discipline. The CEO agrees.

 “We hire a disproportionate amount of former athletes. Beth behind the camera [Beth York, Blitzy’s head of media and creative, and our volunteer camera person for the taping in Boston] is an Iron Woman. A really high signal way to vet through people is looking for athletic backgrounds, both on the technical side and go-to-market.”