Memo: Builder.ai Was Not a Fraud. That's the Part That Should Scare You.
A year after Microsoft's $1.5B no-code unicorn collapsed amid claims its AI was 700 Indian engineers in a trench coat, the more useful question is not "how did investors miss it?" It's "what happens to the dozens of 2026 agent startups that are running the exact same play – with slightly better models?"
In May 2025, Builder.ai filed for insolvency. A creditor named Viola Credit had seized $37 million from its accounts, leaving the company with around $5 million in cash against more than $115 million owed to Amazon and Microsoft alone for unpaid cloud bills. Within weeks, bankruptcy proceedings opened in the UK, US, India, UAE, and Singapore. About 1,000 employees were let go. A handful of small businesses found themselves with half-finished apps and no support contract. Someone spray-painted "AI lied. We died." on the wall outside the Shoreditch office.
The press read this the way the press always reads these stories. The headlines said Microsoft-backed AI unicorn collapses. The think pieces said the AI grift catches up. The LinkedIn posts said we should have known. The Wall Street Journal had, after all, reported in 2019 that Builder.ai's flagship AI – a product called Natasha – was largely 700 engineers in India writing code by hand. Six years later, when the company finally died, the obvious narrative wrote itself: a company faked AI, got caught, ran out of money, the end.
That narrative is wrong. Or rather – it is correct in the small details and wrong in the part that matters.
The part that matters is this: between the WSJ exposé in 2019 and the bankruptcy in May 2025, Builder.ai raised roughly $400 million in additional funding. The Qatar Investment Authority led a $250M Series D in May 2023. Microsoft made an equity investment and announced a strategic partnership the same year. SoftBank's DeepCore was already in. The IFC was already in. None of these are unsophisticated buyers. All of them had access to the WSJ article. All of them invested anyway.
Either every late-stage investor in Builder.ai was negligent – which is possible but unlikely at that price point – or they were buying something other than the AI claim. And the thing they were buying is what the next eighteen months of agent-startup blowups are going to be about.
What Builder.ai was actually selling
Strip the marketing away and Builder.ai was selling a contract that read, in effect: give us money and a spec, and you will receive a working app. The "AI" in Builder.ai – Natasha, the orchestration layer, the reusable component library – was the production system that was supposed to fulfill that contract at software margins instead of services margins. The pitch to customers was outcome-priced software development. The pitch to investors was that the production system would, over time, replace more and more of the human labor inside the contract.
Investors in 2023 were not betting that Natasha worked in 2023. Anyone reading the WSJ piece knew it didn't. They were betting that the gap between the marketing and the reality would close before the unit economics ate the balance sheet.That is a real bet. It is a bet that GPT-3.5 was on the leading edge of a curve, that GPT-4 would close half the gap, that Claude and the next two model generations would close the rest, and that Builder.ai's outcome-priced contracts and customer base would, by the time the models arrived, be a moat instead of a liability.
That bet was not crazy. It was, in fact, the same bet that every serious agent company is making in 2026.
The bet lost at Builder.ai for a specific reason, and the specific reason is the part most of the post-mortems have skipped.
The capability gap and the working-capital gap
When a company sells outcome-priced software – "your app will be delivered, working, for $X" – and the production system can only deliver, say, 60% of the work autonomously, something has to fill the other 40%. At Builder.ai, that something was 700 engineers in India writing code by hand, and the cost of those engineers had to be funded out of the difference between what customers paid and what the system actually cost to run.
In a clean version of this business, the capability gap closes faster than the working-capital gap opens. Models get better. The 60% becomes 75%, then 85%. Margins improve. The human layer shrinks. The flywheel starts to look like SaaS.
In the messy version, three things happen at once.
The capability gap closes more slowly than projected. This is what happened at Builder.ai through 2023 and 2024. Reusable component libraries plus 2023-era LLMs got Natasha to a place where she could glue templates together. She could not, reliably, take a non-technical founder's spec and produce a finished app without an engineer in the loop. The 60% never became 85%. By some accounts internally, it never even became 70%.
The contracts don't reprice. Customers signed for outcomes at fixed prices. When Natasha couldn't deliver, the engineers had to. The cost-per-app stayed roughly flat while the revenue-per-app stayed roughly flat, which means margin stayed roughly flat – at roughly the margin of an outsourced dev shop, not a software company.
The growth narrative has already been priced in. Builder.ai had told the QIA and Microsoft that revenue was on a SaaS-shaped curve. When the actual curve turned out to be a services-shaped curve – Bloomberg later reported revenue was restated from a projected $220M to roughly $55M, and former employees alleged sales figures had been inflated by 20% on multiple occasions – the company had already taken on $50M in debt against the inflated forecast. That was the Viola Credit line. When Viola figured out the forecast was wrong, it pulled the cash. There was no cash left.
The moral of the Builder.ai story is not that the company faked AI. The company did fake AI, but plenty of companies have faked AI and survived – the WSJ exposé in 2019 should have killed Builder.ai and didn't. The moral is that the company sold outcome-priced contracts against a capability curve that was forecast to close faster than it actually did, and when the curve underperformed the forecast, the only available pressure-release was accounting fiction. Which works until it doesn't.
This is the part the next wave needs to read carefully.
Why this is going to happen again, and approximately when
In 2026, a substantial fraction of the agent startups raising at premium multiples are running a structurally identical play. The pattern looks like this:
- The company sells an outcome to an enterprise customer. Resolved tickets. Closed loans. Filed claims. Reconciled invoices. Booked appointments. Drafted contracts.
- The price is set as if a frontier-model agent will eventually do most of the work.
- The actual delivery, today, is the agent doing 50–70% of the work and a human – often offshore, often contract – doing the rest.
- The pitch to investors is that the next model release, or the next fine-tune, or the next tool-use breakthrough, will close the gap. Margins will inflect. The flywheel will start.
Some of these companies are going to be right. The capability curve is real. Frontier models in late 2026 are demonstrably more capable at multi-step tool use than they were in late 2024. For the right workflow, the curve does close.
But the curve does not close uniformly, and it does not close on the schedule the investor deck promises. Some workflows close fast – low-ambiguity, low-stakes, well-bounded transactional work. Others close slowly or not at all – anything involving accountability, anything involving ambiguous specification, anything involving an angry counterparty. We saw this with Klarna two memos ago. The three-layer problem applies to every outcome-priced agent contract, not just customer support.
Which means, for a meaningful subset of the 2026 agent class, the Builder.ai trajectory is the base case rather than the tail risk. The model gets better but not fast enough. The human layer doesn't shrink on schedule. The contracts don't reprice. The growth forecast was presented to investors before the curve was understood, and a debt facility was taken against the forecast. When the forecast misses, the lender pulls. The cash is gone. The press writes another piece about AI hype. The pattern repeats.
Builder.ai was the early warning, not the outlier.
Three things to do if you are buying, building, or funding an agent in 2026
1. If you are an enterprise buyer: separate the capability claim from the contract.
Do not sign an outcome-priced agent contract without an audit clause that lets you see the human-in-the-loop ratio, on a per-workflow basis, refreshed quarterly. Builder.ai customers were told they were buying AI-built apps. They were buying contract development at SaaS prices. The same misalignment is being signed today, in support, in legal, in finance ops. The audit clause is the only thing that makes the misalignment legible before the vendor goes insolvent.
2. If you are building an agent company: publish your capability curve internally.
Track, by workflow, what percentage of completed work is being done by the agent versus the human layer, and what direction the line is moving. Builder.ai's leadership either did not have this dashboard or did not believe what it was telling them. If the line is flat or declining inside any workflow you are selling against an outcome price, you do not have an agent business in that workflow – you have an outsourced services business with a model on top, and your contract pricing is wrong. The earlier you face this, the cheaper the correction.
3. If you are funding an agent company: stress-test the capability assumption against the working-capital assumption.
Ask the founder, on the record, what happens to gross margin if the capability curve closes 18 months later than projected. If the answer involves a debt facility, that is the Builder.ai script verbatim. The Viola Credit line was the proximate cause of the bankruptcy. The distal cause was a forecast that no one stress-tested against the actual model-progress slope. In a market where model progress is non-linear and uneven across workflows, every late-stage round into an agent company should price the slope, not the snapshot.
The bottom line
Builder.ai was not a fraud in the way the headlines told it. It was a fraud in a more interesting and more replicable way. The company sold outcome-priced contracts that depended on a capability curve closing faster than it did, papered over the gap with offshore engineers and inflated revenue, took debt against the inflated revenue, and ran out of money when the lender realized the revenue wasn't real. The "fake AI" framing makes this sound like a moral failure of one company. It is more usefully read as a structural failure mode of an entire category – a category that, in 2026, is being capitalized at hundreds of billions of dollars of paper value.
The companies in that category that survive will be the ones whose internal numbers track the capability curve honestly, whose contracts reprice as the curve moves, and whose financing isn't betting on a slope that hasn't been measured. The ones that don't will produce a pattern of mid-2027 collapses that – when they happen, in clusters, across the same six months – the press will once again call a bubble.
It will not be a bubble. It will be the same memo, written 47 times.
Sources: Bloomberg coverage of Builder.ai's insolvency (May 2025), TechCrunch's reporting on Builder.ai's funding history, the Wall Street Journal's 2019 investigation of Engineer.ai's reliance on human engineers, Rest of World's February 2026 explainer on the company's collapse, and public statements from Manpreet Ratia and Sachin Dev Duggal. All figures cited – the $445M raised, the $250M Series D, the $37M creditor seizure, the restated revenue – are from publicly reported statements by Builder.ai or its investors.