Deterministic scores, generative drafts: drawing the line inside an AI feature

We shipped an AI sales layer on top of our CRM this week: lead and deal scoring, next-best-action, email drafting, funnel analytics, and an autonomous outreach agent. In building it, the design decision that mattered most was not which model to use. It was which outputs a model is allowed to produce at all.

Our answer: the model writes prose. It does not produce numbers.

Why scores are different

A drafted email is inspected before it matters. A human reads it, and if it is wrong, that is obvious on the page. Errors surface at the moment of use, and the cost of one is a rewrite.

A score is different in kind. A lead scored 84 does not look wrong. It looks like 84. Nobody re-derives it, nobody can, and it silently determines what gets worked first. If that number came from a language model, it is unreproducible — the same lead can score differently on Tuesday — and unexplainable, which means the first time it ranks an obviously-hot lead low, the entire feature loses credibility permanently.

Generated prose fails loudly. A generated number fails silently, and gets acted on.

So the numbers are computed

Lead scores, deal scores, forecast bands, and funnel drop-off are computed deterministically from CRM signals. Same inputs, same output, every time, and a reason you can point at. The AI sits around that: interpreting the ranking, suggesting the next action, drafting the message, answering read-only questions over the aggregates. It advises about numbers it did not invent.

This is not a statement about model quality, and it does not get revisited when models improve. It is a statement about where an unverifiable output is safe to put. A model good enough to produce a plausible score is precisely a model good enough to produce a plausible wrong score.

The same test elsewhere

It generalizes past sales. Anywhere a system emits a number a human will act on without re-deriving — a risk score, a priority, a confidence level, a compliance percentage — the question is whether that number can be explained and reproduced. If it cannot, an AI-generated version of it is not a feature, it is an unfalsifiable claim rendered in a font that makes it look measured.