The Contract Review That Paid for Itself

A few months ago, I was sitting across from a payments-industry CEO I'd been coaching on AI workflows. That morning, he had signed an MOU for a US franchise investment — roughly $150K, a nine-figure sum in won, going to someone he trusted personally. The wire transfer was supposed to happen within days. The deadline had already slipped once, and the pressure to just send the money was real.

The deal was built on a relationship, which is exactly the kind of deal that doesn't get scrutinized. He had told me earlier, in so many words, that he thought of the money as a relationship cost — an investment in a person as much as a business. The contract itself was about eight pages, in English, drafted by the other side. He hadn't read it closely. In fairness, most people in his position wouldn't have. He runs several businesses, his calendar that day alone included three other meetings, and the counterparty was someone he'd known for years.

Earlier in the session, before we'd even opened the document, I made a simple pitch: "Before you wire this, let's run it through Claude once. Photograph the pages, upload them, and it'll take fifteen, maybe thirty minutes. It's a lot of money. Thirty minutes before sending it costs you nothing."

He agreed. Not because he suspected anything, but because the math was hard to argue with.

What the session actually looked like

There was nothing exotic about the workflow. We photographed the contract page by page, uploaded the images, had the text extracted, and then asked for a structured review: summarize each section, flag anything missing, flag anything one-sided, and explain what each clause means for the investor if things go wrong.

The first anomaly surfaced fast, and it was almost embarrassingly simple: Section 3 didn't exist. The document's numbering referenced it, but the section itself wasn't in the body. In a casual read, your eye slides right past something like that. A machine that's been asked to account for every section doesn't slide past it.

From there we went clause by clause, and the picture got worse:

• The entire return depended on an IPO. The contract promised shares in connection with a public listing. If the listing was delayed or never happened, there was no clause for getting the money back. No repayment trigger, no remedy.
• There was no buyback clause. Nothing obligated the company — or the founder personally — to repurchase the stake under any conditions. The only exits were an IPO, an acquisition, or finding another buyer for shares in a small private company. In practice: no exit.
• The issuing entity was incorporated in the US. Any dispute would be litigated in an American court, under American law, by a Korean investor. For a sum this size, the cost of enforcement would eat most of what could be recovered.
• The company carried existing debt, disclosed only in passing. New money would sit behind it.
• The valuation didn't hold up. The business had a handful of locations, and the implied company value was off by roughly an order of magnitude from what those locations could plausibly support.

At one point I summarized it for him in one line, because the clause-by-clause detail was starting to blur: as written, this contract says the company doesn't have to go public, doesn't have to return the money, and if it all goes wrong, your only recourse is a lawsuit in a foreign country. Two hundred million won, in exchange for that.

I want to be precise about what the AI did and didn't do here. It didn't "detect fraud." It read page eight the way it read page one. That sounds trivial until you consider that no human reads a friend's contract that way — not for lack of ability, but because the attention doesn't get sustained. The relationship reads alongside you, softening every clause; by the middle pages, fatigue and a quiet surely it's fine have finished the reading for you. The machine carried none of that. It noticed that a numbered section was absent, and answered the question "what happens to the investor if the IPO doesn't happen?" with "the contract is silent." Mechanical observations — but nobody in the deal, not the investor, and I'd guess not even the drafter's own reviewers, had made them, because nobody had given the document that kind of even attention.

What he did with it

What impressed me most wasn't the analysis. It was the speed of the decision that followed.

That morning, the contract had been a stack of pages signed on trust, effectively unread. That afternoon, the same pages sat between us annotated clause by clause, and he restructured the deal on the spot: no equity. Instead, the same amount as a loan, secured by collateral — a share pledge, a personal guarantee, an asset pledge, or a personally guaranteed buyback right, counterparty's choice. We drafted the counter-proposal in the same session, along with three affirmative rights to request: information rights on how the funds would be used, a warranty on the company's financial statements, and a right to force a sale of the stake under defined conditions. If the other side refused everything, he would walk away entirely.

Notably, he didn't kill the relationship. He changed the form of the capital while keeping the personal commitment intact — "I'll still lend you the money, but as a loan, with security." That distinction, between trusting a person and signing an unprotected contract, is the whole game.

Somewhere in the middle of drafting the counter-proposal, he said — half to me, half to the screen — "Honestly, every time I look at this thing, I'm just grateful." Then, a few minutes later, the other half of the emotion: frustration that he had to do any of this at all, that a deal built on trust required this kind of defense. Both reactions were right.

As I told him at the end: if we hadn't met that day, the money might simply have been gone.

What I actually take from this

I'm wary of turning this into a bigger claim than it deserves, so let me state the limits plainly.

AI did not replace a lawyer in this story. The output went to his retained attorneys for a second, legally authoritative review — as it should. An AI reading of a contract is a first pass, not an opinion you can rely on in court.

But the first pass is where this deal was going to fail, because the first pass wasn't going to happen at all. No lawyer had been asked. The document was in a second language, the counterparty was a friend, the deadline was near, and the amount was large enough to hurt but not large enough to trigger the institutional reflexes that a ten-million-dollar deal would. That's the exact profile of money that gets lost.

Most of what gets sold as workplace AI is offense — make more, faster. This story is the other lens, and I think the underrated one: defense. AI as the thing that keeps money you already have. The useful frame isn't "AI catches fraud." It's narrower and more durable: the AI reads page eight like page one, and before you sign anything you haven't fully read, that even attention now costs thirty minutes and almost nothing. It won't negotiate for you. It won't know your relationship with the counterparty. It will, however, reliably notice that Section 3 is missing — and sometimes that's the entire ballgame.

We've since made it a standing rule in his operation: any new investment, loan, or contract gets an AI review before money moves, then goes to counsel. Not because the AI is brilliant. Because the alternative, most of the time, is that nobody reads the contract at all.