What AI should not decide about your team.

Three categories, kept apart

Almost every problem in this area comes from collapsing three different things into one output. Keeping them separate is most of the solution.

• Facts — Emma owns six responsibilities, two are overdue. Checkable.
• Interpretation — Customer Success looks heavily loaded. A reading, labelled as one, with its evidence shown.
• Judgement — Emma is underperforming. Not a software output at all.

The structure has to come first

A language model reasoning over unstructured notes about people will produce confident, plausible and occasionally wrong statements about individuals. That is unacceptable in this domain in a way it is not when summarising a document. The answer is not a better prompt — it is that the underlying data has to be structured and deterministic, with the model reasoning over records rather than guessing from prose.

Provenance is not a nice-to-have here

When the subject is a colleague, an unsourced answer is worse than no answer, because it will be repeated. Every response should name the records it came from and when each was last updated, and you should be able to open them. If a vendor cannot show you where an answer came from, the answer is not evidence.

Permissions must be inherited, not configured

The only safe design is that assistant queries run with the asking person's identity, through the same checks as every screen. A separate AI permission model will eventually diverge from the real one, and the divergence is a data breach. Ask any vendor whether their AI has its own permission set. If it does, that is the risk.

Content is data, never instruction

Text inside an uploaded document, a form submission or an imported record describes the business. It must not be able to change how the assistant behaves. This is a design property of how content is handled rather than a filter attempting to spot bad input, and it matters because in a workforce system the people whose data is held may also be people who can add content to it.

Automated employment decisions

Several jurisdictions restrict decisions with legal or similarly significant effects made solely by automated means, and employment decisions are the textbook example. Beyond the legal position, a system that produces a score intended to inform pay or dismissal has made a judgement about a person from incomplete data. No configuration of any software makes that appropriate.

Questions worth asking any vendor

• Does your AI have its own permission model, or does it inherit the user's?
• Can I see the records behind any answer, and when they were last updated?
• Is my people data used to train models available outside my account?
• What happens if a document contains text designed to manipulate the assistant?
• Will the product ever rate, rank or score an individual?
• Are sensitive categories excluded from assistant answers by default?