You cache AI outputs to save money. So when my organisation changes, am I looking at a decision brief computed on last month’s reality?

The full executive answer

No — because the cache is keyed on your reality, not on a clock. The cache key is a cryptographic fingerprint of the actual inputs: your organisation context, the module, the prompt version, the resolved model, and the state of your document library. Change any of them — complete a new assessment, upload or edit a document, or we improve a prompt or swap a model — and the fingerprint changes, which means the old answer simply can no longer be found. Stale-by-content is structurally impossible; a cache hit means, by construction, that nothing material changed.

Two more layers on top. A staleness registry tracks what each derived artifact depends on — when your maturity scores republish, dependent insights are flagged stale in the interface, so even the edge cases are visible rather than silent. And every screen has force-regenerate, which bypasses the cache entirely and replaces the stored insight in place.

I would also defend the trade-off positively: identical inputs producing an identical, already-validated answer is not just cheaper — it is more auditable and more consistent than re-rolling a generative model and getting a slightly different verdict each time on the same facts. Cache hits are logged in the audit trail and cost nothing against your AI allowance. It is a correctness feature that happens to save money.