~/aakash
Writing ← Portfolio
Agentic AI June 22, 2026

What a Transcript Loses, and Why It Matters in a Contract

My system wrote a perfect SOW. It was quietly wrong, and the reason says something about where AI delivery actually ends and the human still begins.

01

The SOW That Looked Finished

The SOW looked finished. Every section filled in, the scope written out in plain language, deliverables listed in a sensible order. My PM system had drafted it from the call transcript, and reading it back, I couldn't find anything to fix. Honestly, I almost just sent it.

Then I hit one line and stopped. Not because it looked wrong, but because I remembered the moment it came from, and the moment didn't match what was on the page.

Here's what had actually happened on the call. The client was talking through how he wanted to work with the prompts in the system. At some point he started thinking out loud about having a place in the dashboard where he could see all the prompts the app used, tweak them himself, save his changes, and generally not have to come back to us every time he wanted to adjust something. It was a nice idea. But that's all it was at that point: an idea, said in the way people say things when they're musing rather than asking. Something like, it'd be good if, somewhere down the line, we could do something like that. He offered it and more or less set it back down in the same breath. Nobody wrote anything down. Nobody treated it as a decision. It was a maybe.

By the time it reached the SOW, it was a deliverable.

And the frustrating thing is that nothing had gone wrong, exactly. The system did precisely what I built it to do. The transcript had the client describing a prompt-management module, that description pointed to a genuine feature, so the system put a genuine feature into the document. Every individual step was correct. The result was still wrong.

02

What the Transcript Couldn't Hold

That's the part I keep turning over. A transcript captures what gets said. It has no way of capturing how it was meant. The hesitation, that it'd be good if, down the line, lived entirely in the room. It was in his tone, in the way he floated the idea and let it drift, in the fact that all of us moved straight past it without it landing anywhere. None of that makes it onto the page. What makes it onto the page is the feature. And a feature sitting in a list of deliverables doesn't look tentative. It looks promised.

So the module didn't end up in scope because anyone chose to put it there. It ended up in scope because a loose, half-formed thought gets quietly firmed up the moment it's written down. The transcript drops the qualifier, the system works with what's left, and a maybe turns into a will with nobody actually deciding it should.

Pull Quote · 01

A transcript records what was said. It has no way of recording how it was meant.

03

The Cost in a Contract

In a contract, that's not a small thing. A prompt-management module isn't a line you tack on. It's a real build. You're looking at an interface, the logic to edit and save, validation so a broken prompt can't take the app down with it, and all the testing that comes with that. If that SOW had gone out and been signed, we'd have been on the hook to build a whole module the client had never actually asked us for.

And then comes the conversation you really don't want: either we swallow the cost and build it to honour a commitment we invented ourselves, or we go back to the client and explain that something sitting in his signed scope was never really in scope. Neither of those is a fun email to write over a feature nobody had even decided on.

I only caught it because I'd been on the call. The system couldn't have caught it, because the thing you needed in order to catch it was never in the transcript in the first place. It was sitting in my memory of how he'd said it. I don't hold that against the system. It can't be in the room. I can. That's the whole difference.

04

The Trap of Going Fast

And here's the part that actually worries me. I caught it because I'd been there and I was listening. But the whole appeal of a system like this, the reason it's so easy to lean on, is that it lets you not do the slow, effortful work. You don't have to sit through the call fully present. You don't have to read the transcript line by line. You feed it in, a clean SOW comes out the other side in seconds, and it looks great. That speed feels like a gift, especially to someone new to working this way, or to anyone who was half-listening on the call because they were juggling three other things.

That's the trap. The faster and smoother the output, the less you feel any need to question it. A document that arrives polished and complete doesn't invite scrutiny. It invites a signature. If you weren't really present for the conversation it was built from, you have no way of knowing that one of those tidy deliverables was never actually agreed to. You'd send it. You'd feel efficient doing it. And you'd find out weeks later, when the client is surprised by an invoice or you're surprised by a scope they think they're owed.

The uncomfortable bit is that the tool is most dangerous exactly when it's serving you best. The smoother it runs, the more it tempts you to stop paying the kind of attention that catches things like this. Going fast and going unwatched feel identical right up until the moment they don't.

05

Where the Human Belongs

This is the bit people skip when they talk about keeping a human in the loop like it's a checkbox you tick to feel safe. The human isn't there to recheck the machine's sums. The human is there to carry the context the machine has no way of holding: the qualifier that never made the transcript, the gap between what a client said and what he actually meant, the quiet knowledge that one particular line was someone thinking aloud, not someone placing an order.

The system can write the SOW. It does it faster than me and more consistently than me, and most of the time it gets it right. What it can't do is know that one line was a maybe. Knowing that means you had to have been there to hear it, and hearing is exactly the thing a transcript leaves behind.

I've stopped expecting that part to automate away. Not because the tools will catch up eventually, but because what's being judged here, the intent, the tone, how much weight a person actually put behind a sentence, was never written down to begin with. You can't pull out of a document something that was never in it.

So that judgment stays a human job, and I've made my peace with that. If anything, it's a useful reminder of where the tool quietly ends and where being in the room starts.

Takeaway

The tool is most dangerous exactly when it's serving you best.

THE LINE Going fast and going unwatched feel identical, right up until the moment they don't.