Here is a test any enterprise can run this afternoon. Take a question whose answer feeds real decisions, what was our exposure to this supplier last quarter, how many open high-severity findings do we carry, and ask your newest AI tool twice. If the two answers differ, you have learnt something important about where that tool may be allowed to sit.
Generative models are, by design, samplers. Fluency is the product and variation is the mechanism, and for language tasks that is exactly right. But a number that can change on refresh has a specific set of things it cannot do: it cannot be reconciled by two people who received it on different days, it cannot be reproduced for an auditor, and it cannot be traced when the decision it fed goes wrong. AI errors are already reaching boards and external audiences often enough to be benchmarked (the verification gap), and errors are survivable when they are reproducible. An error nobody can regenerate cannot even be diagnosed.
What determinism buys, concretely
Same inputs, same number, every time. Hold an intelligence system to that sentence and four governance properties follow directly:
- Reconcilability. Two viewers, or the same viewer twice, can agree on what the system said, which is the precondition for disagreeing productively about what to do.
- Auditability. Any historical figure can be regenerated from its inputs and checked. The audit trail is not a log of claims but a recomputable fact.
- Attributability. When the number moves, it moved because an input moved, and the driver can be named. Explanation stops being a story and becomes a calculation.
- Accountability. A decision made on the number can be reviewed against exactly the number that was shown, not a best guess at what the model probably said that day.
An answer that cannot be regenerated cannot be governed. It can only be believed or disbelieved, which is not a control environment.
The placement line
None of this is an argument against language models. It is an argument about placement, and the line is clean once stated.
Language belongs at the edges. Turning a plain-language question into a precise query. Summarising a resolved analysis for a human reader. Drafting the narrative around the figures. These are language tasks, ambiguity is native to them, and generative models do them better than anything before. The words of a summary may vary between runs; that is native to language too. What must not vary is the substance beneath them: the facts, figures and citations are deterministic, the narration adds nothing to them, and the version that was shown is kept.
The middle must be deterministic. The retrieval that selects the data, the computation that produces the figure, the aggregation that respects the viewer's permissions, the model that issues a calibrated probability: this pipeline runs the same way every time, and the generative layer is never allowed to be the number, only to talk about it.
The failure mode of the current copilot wave is collapsing that line: letting the sampler both find the data and state the figure, then wondering why two executives are holding different totals. The fix is not better prompting. It is architecture that routes each kind of work to the kind of machine that can be held accountable for it.
A useful procurement question falls out of this: ask a vendor which parts of their answer pipeline are deterministic, and how they prove it. A system built on the right line answers with an architecture diagram. A system built on vibes answers with a roadmap.
Prophesee is built on that line. Deterministic computation and calibrated models produce the figures, and generative language serves at the edges. Same inputs, same number, every time. To see where the line runs on your own stack, start here.