Where AI helps in HR and where it has no business being

Echipa HR 365 · reviewed 2026-09-05 · 5 min read

The rule that separates the good uses from the bad ones: AI helps where it produces a draft a person checks anyway, and harms where somebody’s right follows from its output — a hire, a promotion, a review, a decision to let somebody go.

The three-question test

Before using AI for an HR task, run it through three questions. If the answer to any of them is “no”, the task stays with a person.

Applied to any task

1. 1 — Does somebody check the result before it has any effect? If not, it is not used.
2. 2 — Is a mistake easy to spot and cheap to fix? A bad text is visible; a candidate wrongly screened out is not.
3. 3 — Is the data I am sending free of personal information? If not, that gets solved first.

Where it helps, concretely

The common denominator of those rows: in all of them, the output is a draft that passes through a person’s eyes before it has any effect. The gain is the time from a blank page to a first version, which is often half the total effort.

Where it has no business being

• Automatic selection decisions: rejecting a candidate without anybody looking.
• Performance appraisal. A score generated from text measures nothing and is impossible to explain to the person being assessed.
• Predicting departures at individual level, used as the basis for decisions about that person.
• Analysing tone or sentiment in messages or named survey responses. That is surveillance, whatever it is called.
• Any decision affecting somebody’s salary, job or continuity, made without a person who owns it.

The grey zone: sorting and prioritising

Between the two lists sits the most frequent real use: ordering a stack of CVs. It is neither a checked draft nor an automatic decision — it depends entirely on what happens to the candidates at the bottom of the list.

The practical rule: sorting is acceptable if nobody is removed, only reordered, and if the recruiter also looks at the tail rather than only the top ten. The moment “ordering” becomes, in practice, “I read only the first twenty” is the moment you crossed, without deciding, into the second list.

Where to start

Take the tasks where you already use AI and run them through the three questions. Usually one or two fail, and usually they are the ones introduced without discussion.

Then pick a single task from the first table and measure how much time it saves, over a month. You will need that figure in the next conversation, which will be about expanding.

The cost, which is not zero

The discussion about AI in HR almost always starts from what it saves and skips what it costs. It is not just the subscription: every use carries a verification cost, and if verification takes as long as writing would have, the gain is zero.

The conclusion from the example is not the figure but the condition: the saving exists only if verification stays serious. In practice it decays — after twenty good texts, the twenty-first is no longer read carefully. That is the moment the mistake appears.

How you introduce it without producing resistance

Start with a task nobody enjoys and which decides nothing about people: summaries, a first draft of a text, extraction from documents. The result is visible, the risk is small, and the team gains confidence before the hard conversations.

What does not work: introducing it through recruitment, where the stakes are highest and the scepticism most justified. A bad first experience there closes the discussion for a year.