At trade fairs we are regularly asked whether our AI can also rate candidates. The question comes in a friendly, expectant tone, because everyone is offering it at the moment.
The answer is no. And not “not yet”.
What Lynn does
Our assistant, we call her Lynn, works in four places:
She suggests staffing that takes qualification, availability and workload into account. For detected disruptions she supplies a way out, with a qualified stand-in and prepared messages. She drafts communication for cancellations, rescheduling and follow-up questions. And she points out anomalies in the operational data so that they do not disappear into a chart.
All four have one thing in common: they are about organisation, not about suitability. Lynn knows who has time on which day. She does not know who the better candidate is, and is not meant to.
The professional reason
Aptitude assessment is a craft with a methodology of its own. It asks about requirement profiles, about the validity of procedures, about inter-rater agreement, about whether an instrument is admissible at all for this purpose with this group.
Psychologists practise that craft. They have studied it, they carry professional responsibility for it, and they can explain why they arrive at a judgement.
A language model can formulate a plausible judgement. That is not the same thing. The difference does not show as long as nobody asks, and it shows catastrophically the moment somebody does.
We build a system for organisations where selection decisions have consequences: cockpit staffing, merit-based selection in the public sector, leadership positions. “The model rated it that way” is not a justification that holds there.
The second reason: who plays which part
If a platform organises the procedures and rates the candidates, something fundamental shifts. The supplier becomes a co-decider, and the question of why someone was rejected leads to a company that does not disclose its models.
We do not want to stand in that place. flynne is neutral on test instruments too: we supply none of our own, we connect to your providers. Anyone who orchestrates and judges at the same time has a conflict of interest that cannot be organised away.
What that costs
It would be dishonest to present this as purely a matter of principle. The decision has a price.
We lose deals to suppliers who promise more. “AI-supported pre-selection” sounds stronger in a tender than “AI-supported scheduling”. In innovation comparisons the restrained position looks like falling behind.
We accept that, because the alternative is more expensive: a product that becomes unusable the moment a customer has to justify it.
What the AI Act has to do with it
Little, at least as the reason. The decision was taken before the AI Regulation took effect.
That it is now also convenient in regulatory terms pleases us. Systems that take part in hiring decisions count as high risk under Annex III of the AI Act, with corresponding obligations for providers and deployers. Anyone not using such systems does not have that discussion.
But we would have arrived at the same decision had the regulation never existed. A position you only take because the legislator demands it does not survive the first opportunity to drop it.
What we keep open
No dogma without a boundary. There are areas in which we will expand our use of AI, everywhere it is about organisation: capacity forecasts, better detection of disruption patterns, suggestions on how appointments sit across several weeks, help with drafting.
And there is one area that stays clear: the assessment of people.
That is not a technical boundary. It is a boundary of responsibility. It runs where accountability cannot be handed to software without something being lost.
More on what Lynn can do and where we deliberately stop: flynne AI: Lynn.
