Headhunters use AI on the research half of a retained search — mapping a market, building a longlist, drafting the position specification and the scorecard, and structuring reference notes. The approach itself, and the assessment, stay with the consultant, because that is what the retainer buys.
Key takeaways
A retained assignment is four pieces of work. Define the role and the market. Map the market and build a longlist. Approach, assess and shortlist. Support the client through selection, references and close. The fee is charged for judgement in the third and fourth pieces, and most of the hours are spent in the first two.
That asymmetry is the whole opportunity. Research is where the hours are; assessment is where the value is. AI moves hours, not value — which in a retained model is exactly the right direction.
The reliable uses are structural rather than clever. Turn a role definition into the set of adjacent titles, equivalent employers, alternate credential names and plausible feeder functions — the list a researcher would build over two days from experience. Normalise inconsistent job titles across an industry so a longlist can actually be sorted. Summarise a target organisation’s public structure to work out where the role sits. Draft the search strings, then hand-tune them.
What it cannot do is tell you who is movable, who is about to vest, who has just been passed over, or who would never leave a particular boss. Those facts are the reason the client is paying a retainer, and they arrive by telephone.
A good specification is a negotiation with the client rendered as a document. A model is a useful drafting partner for it — taking the intake discussion, the organisation chart and the compensation frame and producing a structured draft the consultant then argues with. The same applies to the interview scorecard: define the competencies, and generate the behavioural questions that probe each, so the panel assesses the same things in the same order.
The value here is consistency across a panel, not speed. A scorecard that every interviewer actually uses is worth more than a faster longlist.
In contingency recruiting the client is usually named in the market. In retained and confidential search the client’s identity, the incumbent’s situation and sometimes the very existence of the assignment are the secret — and a leak ends the relationship regardless of the outcome.
Every tool that touches the brief holds that secret. The ACSESS Code of Ethics commits members to respect the confidentiality of records in accordance with law and good business practices and to maintain the highest standards of digital security and data privacy, ensuring personal and professional data is protected and stored securely. Applied to software, that is a procurement standard: where does the data go, is it used to train anything, who at the vendor can see it, and how is it deleted.
Contractually, the obligation you already owe does not soften because a tool is involved. It is worth knowing the difference between a confidentiality clause and a standalone NDA before an assignment starts, and what remedies follow a breach of one before you find out. Candidate data is separately governed — PIPEDA applies to the personal information you hold about people you have never placed, and the definition is broad.
Two different restrictions get confused, and a search firm needs both straight.
The first is the firm’s own. The code commits members not to recruit, encourage or entice a candidate they have previously placed to leave the employ of that client, and not to misuse membership privileges to recruit another member’s staff. That is an off-limits obligation the firm carries, and it belongs in the database as a flag, not in a partner’s memory. This is a legitimate and easy automation: mark placed candidates and their client, and have the longlist tool surface the flag rather than rely on recall.
The second belongs to the candidate. A departing employee may be bound by a non-solicitation clause restricting them from recruiting former colleagues, and such a clause can be enforceable in Ontario where it is reasonable in duration and scope and protects a legitimate business interest such as workforce stability. It is not caught by the ban on non-competes, because it does not restrict where the person may work — only whom they may actively try to recruit. The distinction between non-solicitation and non-competition clauses is worth understanding before you build a lift-out around one hire, because the exposure lands on your candidate and, commercially, on your client.
An approach message drafted by a model is fine; an approach campaign is not a search. Where messages are sent at volume, the anti-spam requirements — consent, identification information and an unsubscribe mechanism, with the onus on the sender to prove consent — apply to a search firm exactly as they do to anyone else.
The candidate report is where restraint matters most. Members commit to take all reasonable steps to provide clients with accurate information on each candidate’s qualifications and experience, and to present only candidates who have authorised the firm to represent them. A fluent generated profile that infers seniority from a title or smooths a gap breaches that quietly and convincingly. Write the assessment yourself; let the tool structure and tidy it.
Reference calls are the same shape. Recording and transcribing with consent is a genuine improvement, because it lets the consultant listen. The judgement about what a reference actually meant is not transcribable.
A three-partner search firm runs a confidential search for a chief financial officer at a mid-market manufacturer. The incumbent does not know.
Research: the brief is stripped of the client’s name before any tool sees it, and the market map is built against a described company rather than a named one. The longlist runs to 140 names across four adjacent sectors, deduplicated against the firm’s off-limits flags automatically.
Drafting: the position specification is drafted from the intake notes and rewritten twice by the partner. The scorecard is generated from the agreed competencies and used by all three interviewers.
Approach: every one of the 22 first approaches is made by a partner, by phone. No sequence, no automation, no exceptions — because the first thing a candidate asks is who the client is, and the answer is a judgement call about that person.
Reporting: candidate reports are written by the partner from her own notes, then formatted. One draft is cut because it implied a board-level mandate the candidate had described as advisory. That correction is the job.
Put your off-limits obligations into the database as flags and have every longlist check them automatically. Then move market mapping and specification drafting. Leave approaches and assessment where they are — they are the retainer.
Can we use AI to write the candidate report?
To format and tighten your assessment, yes. To generate the assessment, no. The code requires accurate information about qualifications and experience and authorisation from the candidate to represent them — both of which are yours to give, not the tool’s.
Is it safe to put a confidential brief into a general-purpose tool?
Not by default. Strip the client’s identity, and treat the vendor question as a procurement matter: data location, training use, staff access and deletion. The confidentiality commitment is not satisfied by a setting you have not read.
Does an off-limits agreement stop us approaching anyone at a client?
That depends on what you agreed with the client, which is a contract question. Separately, the industry code commits members not to recruit a candidate they previously placed away from that client. The two overlap but are not the same, and only one of them is negotiable.
A 30-minute call is enough to tell you whether AI pays for itself here.
