18 July 2026
What does an AI recruiter actually do?
A plain English guide to what AI recruiting software really automates in 2026, what it must never decide, and the questions worth asking any vendor.
An AI recruiter is software that does the search and admin half of recruiting. It reads the role, searches a candidate market, ranks people against your must-haves with written reasons, drafts the first outreach message, and remembers everyone it has seen. It does not interview, negotiate or decide who gets hired. A person does.
That is the whole category. The rest is detail: what is genuinely automated, what is marketing, and how to tell them apart before you pay.
Four words the market keeps mixing up
Most of the confusion is vocabulary. Vendors use four words interchangeably. They mean very different things.
| Word | What it actually means | What you get |
|---|---|---|
| Automation | Rules you configure once. Same trigger, same action. | Scheduling, status emails, ATS updates. Reliable and dull. |
| Assistant | You ask, it answers or drafts. You drive every step. | A faster blank page. Nothing happens unattended. |
| Matching model | Ranks candidates against criteria and produces a number. | A shortlist, useful only if the number carries reasoning. |
| Agent | Takes a goal, plans the steps, uses tools, comes back with work done. | Rarer than the marketing suggests. |
That last row deserves scepticism. In June 2025 Gartner predicted more than 40 percent of agentic AI projects would be cancelled by the end of 2027, and estimated only around 130 of the thousands of vendors claiming agentic capability were genuine. Gartner calls the rest agent washing: an old chatbot or rules engine with a new label. So the label tells you nothing. Look at the work instead.
What the software actually does, one role end to end
One requisition from open to shortlist, with the honest split at each stage.
| Stage | What the software produces | What a human still decides |
|---|---|---|
| Market research | Pool size for the role and region, salary range, who else is hiring, realistic time to fill | Whether to open the role at this salary and location |
| Brief into criteria | Structured must-haves and nice-to-haves, pulled from a plain description | Which requirements are real and which are habit |
| Sourcing | A ranked list from a searchable market, not only people who applied | Whether the list fits the role, and what it missed |
| Ranking and explanation | A per-candidate verdict against each must-have, with the case for and against | Who is worth an hour of your time |
| Outreach drafting | A first message built from that person’s profile, in their language | Whether it is good enough to send, and then sending it |
| Screening | Structured answers to structured questions, parsed into comparable data | Who reaches an interview |
| Memory | Everyone you touched stays searchable and gets re-matched against later roles | Who to call back when a similar role opens |
Read the right-hand column again. None of it is automated by anyone credible.
That split matches the industry. SHRM’s State of AI in HR 2026 report found recruiting to be the most common practice area for AI in HR, at 27 percent, and described real-world applications as “concentrated in transactional, process-driven tasks” (SHRM). Transactional is most of a sourcing week.
Five things an AI recruiter cannot do
1. It cannot make the hire. In Europe that is not only good practice. AI used for recruitment and selection is high risk under Annex III, point 4 of the EU AI Act, which carries human oversight obligations. A verdict you cannot inspect or overrule is a compliance problem as well as a bad tool.
2. It cannot send on LinkedIn for you. LinkedIn’s help centre is explicit: third-party software that scrapes, alters or automates activity on the site is not permitted, and members using it risk having their accounts restricted (LinkedIn). Software can draft the message. You send it from your own account. Anyone promising to send for you is selling you a suspension.
3. It cannot hold the relationship. The candidate nervous about relocating, the one weighing a counter-offer, the one whose partner has an opinion. Those conversations decide whether a role closes.
4. It cannot know what nobody wrote down. The hiring manager’s real bar, the politics behind the headcount, why the last person left. Systems work from what is written. The gap between the brief and the truth is a human job.
5. It cannot prove it made you a better hire. The efficiency evidence is strong. Independent evidence that AI improves who gets hired is not there yet, and most organisations are not checking: SHRM’s State of AI in HR 2026 found 56 percent do not formally measure the success of their AI investments at all (SHRM). Treat any quality-of-hire claim as a hypothesis to test on your own roles.
A search box, or a system that remembers?
This distinction decides what you are actually buying, and it never appears on a feature list.
A search box gives you results, and every search starts from zero. Close the tab and the work is gone. Most sourcing tools are this, and a good one is still worth money.
A system that remembers keeps the output of every search: the people, the reasoning, the replies, the rejections. Role two starts with role one’s shortlist in it. The candidate who reached your final round last spring resurfaces when a fitting role opens, instead of being re-sourced as a stranger. That is the silver medalist problem, and it is the quietest waste in recruiting.
Four tests that separate them:
- Does your second role visibly benefit from your first? Ask for a demo on two roles, not one.
- Does it re-match past candidates to new roles without being asked?
- Can it show why it passed on someone, not only why it picked someone? Explainable matching is the difference between evidence and a black box.
- If you leave, does the intelligence leave with you, in an export you can read?
The questions to ask a vendor
Take these to the demo. The answers are more diagnostic than the product tour.
- Which of the four is your product: automation, assistant, matching model or agent? Make them pick one word.
- Show me a candidate you rejected, and why. A tool that shows only its hits is a highlight reel.
- Who presses send on LinkedIn, me or you? There is one correct answer.
- Where does candidate data live, who are the subprocessors, and what happens if I cancel? In writing.
- Under the EU AI Act, are you the provider and am I the deployer? What documentation do I get?
- Can I run this on twenty of my own past candidates, where I know the outcome?
- What is in the price and what is not? ATS connection is usually a higher tier, ours included.
- What does month six look like that month one does not?
If a vendor cannot answer the last one, you are buying a search box.
Where Avery sits, honestly
Avery is one system in this category, built for white-collar hiring in Europe. It searches 125M+ vetted European professionals, ranks each person against your must-haves with the reasoning, drafts outreach you send from your own inbox, and keeps everyone in a talent hub that re-matches them to later roles. It drafts LinkedIn messages, it does not send them. The product pages, the AI recruiting explainer and the comparison with LinkedIn Recruiter go further, including where it is the wrong choice.
It is not the only credible option, and the right answer depends on how you work. The freelance recruiter and in-house team pages set out the two versions of the job, which need different things from the same software.
Two questions people always ask next
Is an AI recruiter the same as an ATS? No. An ATS records candidates who already reached you. An AI recruiter works upstream, finding people who never applied, ranking them and drafting the first contact. Many teams run both, which is why ATS integration is worth asking about early.
What should it cost? It depends what is included. Avery is free to search at EUR 0, Solo is EUR 299 per month, and Team and Max are scoped on a call. Judge any tool against the cost of the role staying open, not your current software bill; the ROI calculator is one way to do that.
If you want to test the difference between a search box and a system that remembers, book a call and bring the role you are most stuck on.