Recruiters use AI on the paperwork either side of the conversation — the ad, the intake note, the search string, the submission summary, the follow-up — and not on the conversation itself. That is not modesty; the submission is a representation you make to a client.
Key takeaways
Ask a working recruiter where the week went and the answer is rarely “talking to people”. It is writing the ad, chasing the hiring manager for a proper brief, rewriting the ad, searching, writing 30 messages, writing submission summaries, coordinating five diaries, writing the debrief, updating the record, chasing the feedback, writing the rejection notes. The conversations are the job. Everything around them is the reason there is not enough time for the job.
That is the honest scope for AI on a desk. It compresses the writing and the structuring. It does not compress the judgement, and it should not touch the representation you make when you put a candidate forward.
If you write ads for Ontario clients, this is now the first thing to know. Under the Employment Standards Act, 2000, as of 1 January 2026, an employer with 25 or more employees on the day a publicly advertised job posting goes up must include in that posting: information about the expected compensation or compensation range; a statement disclosing the employer’s use, if any, of artificial intelligence to screen, assess or select applicants; and a statement disclosing whether a vacancy exists. Employers are also prohibited from including any Canadian experience requirement in the posting or in any associated application form.
Two details matter enormously for an agency. First, the guidance is explicit that where an employer uses a third party to advertise the posting, or advertises on a third-party site, the employer is responsible for ensuring the requirements are followed — which in practice means your client will look to you to have got it right. Second, the disclosure does not have to be detailed: it is enough to state that artificial intelligence is used to screen, assess or select applicants.
The Act’s definition of artificial intelligence is deliberately broad — a machine-based system that, for explicit or implicit objectives, infers from the input it receives in order to generate outputs such as predictions, content, recommendations or decisions. If your stack ranks or filters applicants, assume you are inside it and ask the client the question early.
So the useful AI job here is not “write me a job ad”. It is a template that holds the mandatory elements, into which a model drafts the role-specific copy, plus a checklist that fails if the compensation range or the AI disclosure is missing. That is a five-minute build that removes a recurring client-facing error.
Most bad searches are bad briefs. A recorded, transcribed and structured intake call — must-haves, nice-to-haves, the reason the last person left, the interview process, the decision-maker, the salary reality — is the difference between two submissions and eleven. A model is genuinely good at turning a rambling 25-minute call into that structure, and at flagging what the manager did not answer.
Send the structured brief back to the hiring manager for confirmation before you source. Half the value is the transcription; the other half is that the manager now has to look at their own requirements written down.
Two useful modes. The first is translation: describe the person in plain English and get a search string, then iterate on the string rather than on the description. The second is expansion: from a shortlist of people who obviously fit, generate the adjacent titles, the equivalent employers and the alternate credential names you would not have thought of at 4pm on a Thursday.
What it cannot do is tell you whether someone will move. That is a phone call.
This is the part of the job most exposed to a fluent writing tool and least suited to it. Under the ACSESS Code of Ethics, member agencies commit to take all reasonable steps to provide clients with accurate information on each candidate’s employment qualifications and experience, and to present only candidates who have authorised the agency to represent their application. A generated summary that smooths a nine-month gap, upgrades a title, or infers a certification from a course name breaches both halves of that commitment while reading beautifully.
Use it, but invert the workflow: you write the facts, it writes the prose. The tool should never see anything the candidate did not tell you, and every claim in the summary should be traceable to the CV or your notes. The ethical AI principles ACSESS endorses put the same point structurally: human in command, with clear processes keeping systems under human direction at all times, and transparency about where the technology was used.
Volume messaging is where a desk quietly acquires legal exposure. Canada’s anti-spam legislation sets three requirements for a commercial electronic message: obtain consent, provide identification information, and provide an unsubscribe mechanism. Consent may be express or implied, express consent does not expire but can be withdrawn, and the onus of proving consent sits with the sender. The CRTC enforces it.
A generator that produces 200 personalised messages does not change any of that; it multiplies it. The sane pattern is AI-drafted, human-approved, sent from a system that records consent and honours an unsubscribe automatically.
A four-desk agency in Mississauga fills accounting and finance roles for mid-market clients. The senior recruiter runs a week like this.
Monday: two intake calls, both recorded and transcribed, structured briefs back to the managers by 3pm. One manager corrects the salary band in writing — a correction that would previously have surfaced at offer stage.
Tuesday: ads drafted from the client-specific template. The template refuses to export without a compensation range and, for the client that runs a screening tool, the AI disclosure. Search strings drafted from the briefs, then hand-tuned.
Wednesday and Thursday: calls. No tool involved. Notes dictated after each and structured automatically into the record.
Friday: six submissions. The recruiter writes the factual bullet points; the model turns them into the house format. She reads every one against the CV before it goes, and cuts a sentence that implied a designation the candidate is still working towards. Rejection notes for the 14 people who will not go forward are drafted in a batch and sent individually — the part of the job that gets skipped when there is no time, and the part candidates remember.
Pick the intake call. It is the single change with the largest downstream effect, it requires no client permission beyond the ordinary courtesy of asking before you record, and it produces an artefact your client can correct. Add the ad template second, because it removes a compliance risk. Leave submission summaries until you have a rule about what the tool is allowed to see.
Should I tell candidates I used AI to write the message?
You are not required to label a drafted email, but you are required to be accurate and identifiable under the anti-spam rules. The line that matters is different: if a tool screens, assesses or selects applicants for a client’s Ontario posting, that must be disclosed in the posting itself.
Can AI write the candidate summary I send to a client?
It can write the prose from your facts. It should not generate the facts. The industry code requires accurate information about qualifications and experience, and that you only present candidates who authorised you to.
Is it worth it on a desk that is not high volume?
The intake and note-taking gains are volume-independent — they come from better briefs, not faster typing. The sourcing and messaging gains scale with volume. A search-led desk should expect the first, not the second.
A 30-minute call is enough to tell you whether AI pays for itself here.
