When AI was the wrong answer — and we shipped plain software anyway

James: Every investor update you sent that quarter had the word "AI" in it. And the first thing you shipped had none. Walk me through that.

Priya: The board wanted an AI triage assistant. The customer wanted one number on a screen that they could trust at 3am. Those aren't the same product. When we actually watched nurses use the prototype, the model's "explanation" was the thing they distrusted most.

James: So what did you ship instead?

Priya: A rules engine. Maybe two hundred lines. A clinician could read every branch and sign off on it. It went live in three weeks, it cost us almost nothing to run, and it passed the procurement review because we could show exactly why it did what it did.

James: This is the thing I keep saying on this show — deep on AI when it matters, plain software when it doesn't. How do you know which one you're in?

Priya: I ask what breaks if the answer is wrong. If a wrong answer is cheap and reversible, a model is a great bet. If a wrong answer ends up in someone's medical record, you want something you can audit line by line. We earned the right to add a model later, on the parts where being approximately right was genuinely good enough.

James: And the customer never asked whether it was "real AI."

Priya: Not once. They asked whether it was right, whether it was fast, and whether they'd get fired for trusting it. We optimised for those three. The sophisticated version came in version four — after we had revenue, and after we knew exactly which decisions deserved it.