The Handshake
Daniel signed the letter of intent on a Friday afternoon, in a borrowed meeting room, with his co-founder reading over his shoulder and a bottle of something decent waiting in the fridge.
Forty million dollars. Cash. For a company he had started in a spare bedroom four years earlier.
The business made route-optimisation software for regional logistics fleets: a cloud dispatch platform, a driver app, and a small on-premise connector that sat inside customers’ warehouses and talked to their legacy systems. Thirty-eight people. Healthy retention. Revenue growing at a rate that made investors lean forward.
And an engineering team that shipped like a company three times its size.
That line was in the pitch deck, and Daniel believed it, because it was true. Eighteen months earlier the team had adopted AI coding assistants across the board. Release cadence doubled. The roadmap stopped slipping. When the buyer’s corporate development team asked how twenty engineers had built so much so quickly, Daniel told them.
“We use AI for everything. It’s like having a second team.”
Nobody in the room wrote that down as a risk.
They wrote it down as a strength.
Day 6 — The Clean Report
The buyer was a listed supply-chain software company with a warehouse management suite installed at more than six hundred sites. The plan was simple: fold Daniel’s routing engine into that suite and ship it to every customer in the next major release.
Diligence ran on a 45-day exclusivity clock. Deal codename: Project Harbour.
The data room filled the way data rooms do. Cap table. Customer contracts. Employment agreements with IP assignment clauses—all signed, all clean. A list of open-source dependencies, generated from the build manifests, licences attached.
On day six, the buyer’s standard software audit ran against the repositories. It checked every declared dependency against known licences and vulnerabilities.
It came back clean. No copyleft packages. No unexpected licences. Nothing to flag.
On paper, it was the tidiest target the deal team had seen all year.
Day 11 — The Question Nobody Had Asked
Grace led the technical due diligence. Her firm had run more than a hundred software audits. She had a checklist for everything: architecture, security, scalability, key-person risk.
The checklist said nothing about AI.
It came up almost by accident, in an interview with Daniel’s co-founder, who ran engineering. Grace asked how the team had rebuilt the driver app in under three months.
“Honestly? The assistant wrote a lot of it.”
“How much is a lot?”
He didn’t know. Nobody did. There was no field for it in the pull request template. No tag in the commit history. No policy that said to record it. Why would there be? It was simply how people wrote code now.
Grace asked for the assistant’s usage export. The admin console could tell her how many lines of suggested code had been accepted, per developer, per month. It could not tell her which lines, in which files, or whether they had survived later edits. And it counted nothing pasted in from the chat window—or from the general-purpose chatbot half the team kept open in a browser tab.
Her team spent four days triangulating: usage telemetry, commit timestamps, interviews, and the tell-tale shape of machine-written code.
Their estimate: 23% of the first-party codebase had been generated or heavily assisted by AI over the previous eighteen months.
Grace wrote “floor, not ceiling” beside the number and underlined it twice.
Day 16 — One Function, One Afternoon
A number is a finding. It is not yet a problem.
The problem arrived on a Wednesday afternoon, when one of Grace’s engineers got curious about an unusually elegant function in the on-premise connector: a parser for a legacy warehouse file format that almost nobody still used.
It didn’t look like something a startup would write under deadline. It looked like something a specialist had refined over years.
He pasted a few distinctive lines into a public code search.
The first result was a near-exact match. Same structure. Same odd edge-case handling. Two variables renamed. The source was a long-established open-source project, GPL-licensed.
No licence header had come with it. No dependency. No import. The assistant had simply produced it, and a developer had accepted it, the way the team accepted hundreds of suggestions a week.
The team sampled two hundred more AI-heavy functions by hand. They found two further matches: one lifted from a permissively licensed library without the required attribution, and one whose origin they couldn’t pin down at all.
And those were only the matches that had survived close to verbatim.
Rename a handful of variables, and a text search finds nothing.
POV: The Founder
Daniel took the call in his car. His own lawyer was on the line, and so was Helen, the buyer’s deputy general counsel.
At first, he didn’t understand why a few dozen lines of parser code should matter to a forty-million-dollar deal. The connector was a minor component. They could rewrite it in a fortnight.
Then Helen asked a question he couldn’t answer.
“Which other files should we be worried about?”
He had no list. He had never needed one. He had treated the AI assistant the way he treated a faster laptop: a tool that helped the same people produce more of the same work. It had never occurred to him that it was also a source—that code arriving through it might carry obligations, or authors, of its own.
For eighteen months, speed had been the story. Now speed was the exposure.
POV: The Acquirer’s Counsel
Helen had negotiated software deals for twenty years. She knew how to price a known open-source problem. You found it, scoped it, fixed it, and wrote a specific indemnity. Routine.
This was not routine. There were three problems, and nobody could size any of them.
How much copyleft code was in the product, and where?
How much else had been copied without the attribution that even permissive licences require?
And how much of the codebase did the seller actually own?
The last one kept her awake. Under current US Copyright Office guidance, material generated by AI without sufficient human authorship isn’t protected by copyright. The seller’s standard representations would say the company owned its IP outright. For an unknown share of the code, Helen wasn’t sure there was anything to own.
Distribution made it worse. The GPL’s obligations bite when software is distributed, and for a cloud service that question had been largely academic. But the connector already ran on forty customer servers, and the buyer planned to ship the whole platform, on-premise, to six hundred more.
The assistant’s vendor offered an IP indemnity on its business tier. It depended on a setting that blocked suggestions matching public code. For most of the eighteen months, much of the team had been on individual plans. The setting had been off.
And because the issue was now known to the deal team, the representations and warranties insurer excluded it from cover. Known issues always are.
Helen couldn’t price a risk she couldn’t bound. So, she priced the uncertainty instead.
POV: The Due Diligence Lead
Grace’s final report contained a section her firm had never written before. There was no template for it, because there was no methodology. She had built one in under two weeks, out of telemetry exports, git history, interviews and a sample small enough to embarrass a statistician.
She was candid about what it could and couldn’t show. The sample proved the problem was real. It could not say how big it was. Extrapolate three matches in two hundred functions across the AI-assisted code and you get a range, not a number: a handful more, or several hundred.
The confidence interval was wider than the deal.
Day 38 — The Restructure
The forty-million-dollar deal died in a conference room on a Thursday.
A different deal took its place.
The headline price fell 35%, to $26 million. Four million of that was held back in escrow for twenty-four months, as a special indemnity against IP claims arising from AI-generated code. Closing was made conditional on a clean-room rewrite of the on-premise connector.
Daniel’s investors did the arithmetic on their liquidation preferences. His employees did the arithmetic on their options. It was still a good outcome. Everyone said so.
It was also fourteen million dollars less than the one they had celebrated.
Nobody had lied. Nobody had cut corners. Every developer had done exactly what the company asked of them.
The company simply couldn’t prove what it was selling.

The Gap
It is tempting to read this as a cautionary tale about AI. That misses the point.
The assistant did what it was built to do. The developers used it as intended. The buyer’s audit worked as designed: it examined declared dependencies, and there were none to declare.
The failure was structural. Traditional IP diligence assumes open-source code arrives as a package, with a name, a version and a licence. AI-generated code arrives as a suggestion: anonymous, unlabelled and, once committed, indistinguishable from the code around it.
Every signal the deal team knew how to read said clean.
The absence of signals was the finding.
The Questions Every Deal Now Needs
AI-assisted development isn’t going away, and nor should it. But IP due diligence was designed for a world where every line of code had a traceable author. That world is gone.
The deals that close at full value will be the ones where both sides can answer a new set of questions.
For buyers:
- What share of the codebase was generated or assisted by AI—and how does the seller know?
- Can the seller show where that code lives, file by file and commit by commit?
- Has the code been checked for open-source matches at snippet level, or only its declared dependencies?
- Will your plans for the product—distribution, on-premise deployment, embedding—trigger licence obligations the seller never faced?
- What do the seller’s AI tool terms promise, and were their conditions actually met?
For sellers:
- Could you produce an AI provenance record tomorrow, if a buyer asked?
- Is your code checked for open-source origin every time it changes, or only when someone finally asks?
- Would your IP representations survive contact with your own commit history?
- Which of your AI tools carry indemnities, and have you met their conditions?
Sellers who can answer these questions will command full value. The rest will discover their price in someone else’s conference room.
Speed didn’t cost Daniel fourteen million dollars.
Not being able to account for it did.


