Recruitment

AI in recruitment. Useful, and genuinely risky.

Recruitment is unusually well suited to AI — it is matching, searching and communicating at volume. It is also one of the few applications where getting it wrong has legal and human consequences beyond a bad quarter, because the decisions affect whether people get work.

Where it helps clearly

Searching your own database. Most agencies have thousands of candidates they cannot effectively search. Semantic search over CVs — finding people by capability rather than keyword — surfaces candidates who were always there.
Reactivating the database. Candidates placed two years ago are moving again. Nobody has time to work through a list that size manually.
Job description drafting from a client conversation, for the client to approve.
Candidate communication at volume. The industry's worst reputational problem is silence after application. Automated, honest status updates fix a real grievance cheaply.
Business development. New client acquisition is the constraint in most agencies, and it is exactly the kind of research-and-write work that automation suits.
Formatting and summarising CVs into a consistent client-facing format.

Where it is risky

Screening and ranking candidates is where recruitment AI attracts regulatory attention, and the concerns are well founded. Models trained on historical hiring data can reproduce historical patterns, including discriminatory ones, and they do so in a way that looks objective — which makes it harder to challenge, not easier.

Automated rejection without human review is the highest-risk use available. Several jurisdictions have introduced or are introducing rules requiring transparency or auditing where AI materially affects hiring decisions.
Proxy discrimination. A model can learn to use a signal that correlates with a protected characteristic — a postcode, a school, a career gap — without ever being given the characteristic itself.
Opacity. If you cannot explain why a candidate was ranked low, you cannot defend it to them, to your client, or to a regulator.

A defensible position: use AI to surface and summarise, never to reject. Widening the pool a human then assesses is a use that improves outcomes for candidates. Narrowing it automatically is the use that generates the harm and the liability.

Why candidate data belongs in-house

A recruitment database is CVs, salary histories, references and personal circumstances — a concentration of personal data with strong protections in every jurisdiction. It is also your entire commercial asset.

Running matching and search on your own infrastructure means candidate data never reaches a third party. For an agency whose database is the business, that is both a privacy position and a competitive one. How private AI deployment works.

Common questions

Can AI screen candidates for me?

It can summarise applications and surface people matching stated criteria, which saves substantial time. Using it to automatically reject is where legal risk concentrates — several jurisdictions now impose transparency or audit requirements where AI materially affects hiring decisions, and the discrimination exposure is real. Surface and summarise; let a person decide.

Is AI in recruitment discriminatory?

It can be, and often in ways that are hard to detect. A model can learn proxies for protected characteristics — postcodes, schools, career gaps — without being given the characteristic. That is why human review of anything affecting an outcome, and the ability to explain a decision, matter more in recruitment than in almost any other application.

What is the safest high-value use?

Searching and reactivating your own database. It has no adverse-decision risk, it works an asset you already paid for, and it surfaces candidates for opportunities rather than filtering them out — the direction of effect is toward more people being considered, not fewer.

Should candidate data stay on our own systems?

There is a strong case for it. A recruitment database is a dense concentration of personal data and simultaneously your core commercial asset. Running matching locally means it never reaches a third party, which addresses both the privacy obligation and the commercial exposure at once.

Keep reading

Running AI locally AI and your business data B2B lead generation