A model can draft. It cannot decide, and it should never sound like it did.
Most product tools bolted a chat box on and called it AI. The harder question is which parts of product judgement you are willing to hand over — and what happens when the answer is confidently wrong.
Suggestions are labelled, and they are cheap to reject
Anything a model produced arrives marked as a suggestion, in your own record, with the input it was given. You accept it, edit it, or throw it away. Nothing is written into your decision log on a model’s say-so, because a log you cannot trust is worse than no log at all.
The Intelligence tools are about your AI work, not ours
If you are shipping AI features yourself, the same rigor applies one level up:
- Intelligence Strategy — decide what you will let a model decide, and write down what stays human.
- Intelligence Design — the prompt, the guardrail and the fallback designed together, because the fallback is the product when the model is wrong.
- Intelligence Validation — score output against ground truth before it ships, not after a customer finds it.
- Intelligence Operations — watch for drift in production, where the distribution moves and the evals do not.
Which engine answers is decided by your plan
Paid plans use hosted inference; the free plan runs against your own local model. That is a real difference in where your text goes, so it is stated on the pricing page rather than buried in a settings menu.
Cost is metered in credits, and a credit maps to underlying inference cost rather than to a made-up unit — so a heavy week costs more and you can see why.