Guide · 7 min read
Using AI in hiring without breaking it
AI is now in most hiring stacks, usually in the two places it does the least good: rejecting candidates and generating outreach at volume. It is genuinely useful for ranking, summarising and drafting. The distinction that matters is not which model you use — it is which decisions you let it make.
Published
Ranking is a good job for a model. Rejecting is not.
Scoring how well a candidate matches a brief is a ranking problem, and models are good at it. They surface people a keyword search misses — someone whose title never says 'platform' but whose last three years were platform work.
Rejection is a different act. It is a decision with a consequence for a person, made on incomplete information, and it compounds: a candidate filtered out silently is filtered out of every future search too.
The workable rule is that a model may reorder a list but may never shorten it. Every rejection stays a human decision, with a human accountable for it.
Match on what candidates told you, not what a model inferred
Inferring seniority, salary expectation or willingness to relocate from a CV is guesswork dressed as data, and the errors are not randomly distributed — they track whoever writes CVs in the style the model saw most.
Asking directly is more accurate and less fraught. Stated salary floor, location, work mode and dealbreakers are facts. Matching on facts avoids most of the fairness problems that inference creates.
It also makes the matching explainable, which matters the first time a candidate asks why they were or were not shown a role.
Volume outreach destroys the channel it uses
Generated outreach is cheap to send, which is exactly the problem. When every candidate receives forty templated messages a week, the response rate for all of them collapses — including yours.
The fix is a cap, enforced in the system rather than in a policy document. On this platform each candidate receives at most three company approaches per month, and that limit is enforced by a database trigger, not by asking recruiters nicely.
Scarcity is what makes the channel work. A message that arrives in an uncluttered inbox with a real salary attached gets read.
Rules worth adopting
No model rejects a candidate. Ranking yes, filtering no.
Match on declared preferences rather than inferred attributes.
Cap outreach per candidate, and enforce the cap in the data layer.
Disclose salary in the first message — it is the single highest-leverage change to reply rates, and no model is required.
Keep a human name attached to every decision a candidate could reasonably want to appeal.
Questions
- Should AI ever reject a candidate automatically?
- No. Ranking and surfacing are appropriate uses; rejection is a decision with consequences for a person and should stay with a human. A silent automated rejection also removes that candidate from future searches, compounding a single error.
- Does AI screening introduce bias?
- It can, particularly when inferring attributes like seniority or salary expectation from a CV, because those inferences track writing style and background rather than ability. Matching on declared preferences avoids most of that exposure.
- How do you stop AI-generated recruiter spam?
- With an enforced cap rather than a guideline. Limiting how many companies can approach a candidate in a given month — in the database, not in policy — keeps the channel worth using for everyone.
- Is candidate data used to train models?
- It should not be, and here it is not. Role briefs and candidate data are used to run searches, not to train third-party foundation models.