Karma drafts.A person decides.
There are good reasons to be sceptical about AI in HR. These are the rules we build Karma against — written down so you can hold us to them, and so our own product decisions have something to lose against.
The rules
Eight rules, written down.
- Nothing is written until a person approves itKarma prepares changes and shows the exact fields it will write. Until somebody presses approve, nothing has happened to your data. This is not a setting with a permissive mode hidden behind it — it is how the assistant is built.
- No automated decisions about peopleHiring, pay, warnings and dismissals are decided by people. Karma may rank a candidate against a job profile and summarise an evaluation, but it does not reject, grade or flag a person — and you cannot switch that on.
- Your data does not train anyone else’s modelKarma works on your dataset. The content of your employee records, applications and review answers is not used to train models that other customers benefit from.
- Karma sees exactly what you see, and no moreThe assistant runs inside the same role model as the rest of the platform. Ask it something your role has no access to and it will not answer — an AI shortcut around access control would be the most expensive kind of convenience.
- A score you can take apartWhere Karma produces a number, the number is explainable. A candidate match is the weighted job profile scored axis by axis, so you see which requirement someone is short on rather than being told they are a 78.
- Everything it did is in the audit logEach suggestion is recorded with what Karma proposed, what was changed before approval, who approved it and when. The log cannot be edited — including by us.
- It can be turned offKarma can be disabled per module. Some organisations want it in recruitment and nowhere near payroll; that is a legitimate position, and the product has to allow it.
- We say when it is wrongAn assistant that is right most of the time still needs checking. Karma marks what it inferred and keeps it editable, and we would rather it say it does not know than produce a confident sentence about somebody’s sick leave.
Where we stop
Things we have decided not to build.
Productivity scoring of individuals. Sentiment analysis of what people write in reviews or pulse surveys. Automatic rejection of applicants. Flight-risk scores attached to a named person. Every one of these is technically straightforward and commercially tempting, and each of them changes what it feels like to work somewhere.
Karma looks for patterns in aggregate — a team drifting over its absence budget, a run of short absences that looks like something else — and hands them to a human being who knows the context. The pattern is a reason to have a conversation, not a verdict.
If a customer asks for one of the things on this list, the answer is no. We would rather lose the deal than find out later what it was used for.
Bring your hardest AI question.
We are happy to have the compliance conversation with your DPO before we show the product at all.