Every business tool now advertises AI. Very few say what theirs is permitted to do, which is the only part that matters once it is sitting on top of your company's data.
Here is ours, written as rules rather than as a promise, because a rule can be checked.
The eight rules
1. The model only ever sees what the person asking could already see.
An AI feature makes the same permission scoped call the human's own screen makes, with the same workspace, the same permissions and the same record scope. It is not given a wider view and told to be careful.
This matters because the alternative is extremely common: give the model broad access and put an instruction in the prompt saying it should only discuss what the user is entitled to. That is a polite request to a model, not a control, and it fails the first time somebody phrases a question cleverly.
2. Small groups are suppressed before the model is called, never after.
Where a figure covers fewer people than the minimum group size, the value is removed from the data on the way in. The model never receives it, so there is nothing for it to leak.
Filtering the answer afterwards is the same idea done in the wrong order, and it only works if the filter anticipates every phrasing the answer could take.
3. Facts are computed in code. Only the prose comes from the model.
Every number on an AI assisted screen is calculated by the engine and displayed beside the text.
Two things follow. The screen still works when the model is unavailable, because the figures were never coming from it. And a wrong sentence cannot make a right number disappear, because the number is sitting next to it.
4. AI proposes. A named human applies.
Nothing on this platform lets an AI decide pay, approve a payment, reject a candidate, close a request or write to a person's record.
The most an AI feature does is fill in a form and hand it to somebody, in the same form with the same validation and the same permission check they would have used themselves. They press submit, and their name is on it.
5. Flag the artefact, never the person.
Patterns, transactions, themes, groups and documents. Not scores, rankings, characterisations or inferred intent about an individual.
6. Refuse rather than degrade.
Where there is no published policy to answer from, the assistant says it does not know instead of producing something plausible. That refusal is the feature, and it is the hardest one to hold, because a confident wrong answer demonstrates better.
7. Record it.
Every suggestion is written down before a human sees it, along with what the human then did with it: accepted it, edited it, or threw it away. That record is what makes it possible to answer the question "why did this happen" a year later.
8. Disclose it.
Every AI surface says that an AI assisted and a person decided.
What we refused to build
This is the part most vendors leave out, and it is more informative than the feature list.
We do not score individual people for flight risk. A per person prediction that somebody is likely to leave would demo extremely well and is asked for by name.
We will not build it, for three reasons. It is acted on quietly. Its subject never sees it and cannot correct it. And it is unfalsifiable in a way no accuracy figure captures: a manager told that somebody may leave treats them differently, that person then leaves, and the score is recorded as correct.
What we build instead is the team level version. Where in the organisation are the conditions that historically precede people leaving, reported as a group figure, with the contributing factors named, and with no individual identified anywhere. It answers the same business question and it cannot be printed and put in somebody's file.
We do not screen or rank job applicants with AI. No résumé scoring, no automated shortlisting, no ranked candidate summaries. AI in hiring here drafts a job description and structures a document somebody uploaded. It does not form an opinion about a person.
We do not price by inferred willingness to pay.
These are recorded as permanent refusals in our own planning documents rather than as deferred work, and the wording matters: deferred work gets picked up by somebody who was not in the room for the reasoning.
Why publish this at all
Because the asymmetry runs one way.
Shipping one of those features and withdrawing it later means telling a paying customer we have removed something, which never happens cleanly. Not shipping it stays reversible in both directions.
And because a rule that is written down in public is much harder for a future version of us to quietly drop.