"The Customer Said" Is the Most Dangerous Sentence in Product.

"The customer said they want this."

Five words. They end arguments. They unblock roadmaps. They justify quarters of engineering. And almost nobody in the room asks the obvious follow-up: which customer, when, in what words?

It's the most powerful sentence in a product meeting and the least examined. We grant it the authority of evidence while asking it for none of the proof.

Memory is not a citation

Here's how the sentence gets made. A PM ran some calls last quarter. A few moments stuck. Over weeks, those moments smooth into a confident generalization: "enterprise wants approval workflows." The specific customer fades. The hedge fades. What's left is a clean claim that sounds like a finding and is a memory of a vibe.

That isn't dishonesty. It's how memory works. We keep the gist and lose the source. We remember the calls that confirmed what we suspected and forget the ones that didn't. By the time it reaches the PRD, "one or two customers mentioned something adjacent" has become "customers want," and nobody can reconstruct the difference.

The claim is load-bearing. The feature gets built on top of it. If the claim is a smoothed-over memory, the feature is built on sand, and you find out when it ships and nobody uses it.

In 2026 the sentence got cheaper to forge

It used to take a PM weeks to produce a confident-but-unverified claim. Now it takes a model twelve seconds.

You paste a brief into a chat and out comes a tidy PRD with twenty customer claims arranged into user stories. Some trace to something a customer said. Some are pattern-matched plausibility the model produced because the prompt implied they'd fit. They are written in the same confident voice. They are indistinguishable on the page.

The reviewer can't tell which is which. The designer takes them as input. Engineering builds against them. "The customer said" now carries the same authority whether a customer said it or a language model guessed it would be reasonable. That should worry you.

The fix is not more skepticism

You can't out-discipline this by reading more carefully. Confident-sounding claims are free to produce and impossible to vet by eye. The bottleneck moved from making the claim to checking it, and squinting harder at the doc checks nothing.

What checks it is a corpus: every customer conversation typed into structured, searchable feedback, so "the customer said they want approval workflows" stops being a sentence you trust on the speaker's confidence and becomes a query. Run a claim against the record and three answers come back:

• Supported: fourteen customers said it, here are the clips.
• Contradicted: two said it, six asked for the opposite.
• Unsupported: nobody said this, anywhere.

That last answer is the one no model can give you, because it's the model's blind spot by construction. The thing that wrote the claim cannot tell you the claim is unsupported. Only the record can.

Make the sentence earn it

The healthiest product cultures have one quiet rule: "the customer said" is an invitation for a receipt, not the end of the conversation. Not as a gotcha. As a reflex. You say it, you can show it, and if you can't, you say "I believe" instead, which is a fine thing to say and a completely different thing to build on.

That reflex is impossible without somewhere to look. You can't ask for receipts when finding one means re-watching fifty hours of calls at 11pm. You can ask for them when the moment is one query away.

---

We built Fragment so "the customer said" comes with the clip attached, and the claims that don't survive the check get caught before they reach the roadmap, not after.