What an AI Recruiter Can and Cannot Do

Two products are being sold under the same two words, and they are not at the same stage of maturity.

One of them books the interview: it reads a panel's calendars, finds the intersection, sends the confirmation, handles the reschedule at eleven at night and tells the recruiter it happened. That product works now, unattended, at volume. The other one reads a resume and tells you the candidate is a 78 out of 100. That product does not work, and the way it fails is well documented.

Both get called an AI recruiter. Which one a team bought explains almost everything about whether their AI rollout is delivering or sitting in a pilot.

What is an AI recruiter?

The word doing the work in that definition is executes. A tool that drafts an outreach email for a recruiter to send is an assistant. A tool that sends it, reads the reply, books the screen and updates the ATS is an agent. The distinction is not marketing, it is the whole question of whether a human stays in the loop on every action or only on the exceptions. We drew that line in more detail in AI agent vs chatbot.

The label covers at least five different things in the current market, and a buyer should insist on knowing which they are looking at.

* Sourcing agents. Search a candidate database, write the outreach, run the follow up sequence, book the interested replies.

* Screening agents. Read applications and rank or score them against the requisition.

* Interviewing agents. Conduct a recorded or live interview and produce a rating.

* Coordination agents. Collect availability, resolve panel conflicts, book rooms and links, send confirmations, absorb reschedules.

* Companion agents. Sit inside the interview, take structured notes and write the scorecard draft the interviewer edits.

Nothing about that list is controversial. What is worth arguing about is that these five have wildly different reliability, and the market prices them as though they were the same product.

What an AI recruiter does reliably today

This is the half that is finished, and it should be said plainly rather than hedged. Scheduling a panel is a constraint satisfaction problem with a hard answer. Five calendars, a room, a time zone, a training rule that says the new interviewer shadows before they lead. A machine is better at that than a person is, not marginally but by an order of magnitude, and it does not get worse at six on a Thursday.

The size of the prize is set by how much of a recruiter's year the stage consumes. Greenhouse's Hire Standard benchmark report, published March 2026 across more than 6,000 companies and 640 million applications, puts a recruiter at 650.5 interviews scheduled a year, up 128 percent since 2022, while recruiters per organization fell from 10.43 to 4.62. The volume went up and the people went away. On our own platform a manually coordinated event averages 243 minutes of coordination time against 27 minutes when the candidate self schedules against real interviewer availability, and roughly 14 percent of events get renegotiated for structural reasons that have nothing to do with the candidate.

That is the work an agent takes. Not a suggestion, not a draft for review: the availability request goes out, the intersection gets found, the invite lands, the reschedule gets absorbed and the recruiter finds out afterwards. Coordinators running this way handle volumes that were not previously possible, which is the argument in 153 interviews a week.

Outreach sequencing, application acknowledgement, interview reminders, ATS field updates and post interview status notes belong in the same category for the same reason. Each has a definition of done that a system can confirm without waiting for a human to form an opinion.

What an AI recruiter still gets wrong

Kyra Wilson and Aylin Caliskan of the University of Washington Information School ran more than three million resume to job comparisons across 550 real resumes, 120 first names and 500 job listings, and presented the results at the AAAI/ACM conference on AI, Ethics and Society. The models favoured white associated names 85 percent of the time against 9 percent for Black associated names, and male associated names 52 percent against 11 percent for female. In the male comparison, the systems never once preferred a Black associated name over a white associated one.

The instinct is to read that as a training data problem that a better vendor has solved. It is more stubborn than that. A scoring model is asked to predict something nobody has measured. Even at a company with excellent records, the outcome data covers only the people who were hired, so the model learns to reproduce past selection rather than future performance. There is no counterfactual for the candidate who was rejected and would have been excellent.

Interview scoring runs into a second ceiling that has nothing to do with AI. In the most recent revision of the selection literature, Sackett and colleagues put the operational validity of a structured interview at .42 against .31 for a cognitive ability test. The predictive power in that number belongs to the structure, the same questions asked in the same order and scored against the same anchors, not to who or what does the scoring. An agent that enforces structure is doing something useful. An agent claiming a better prediction than the structure supports is claiming something no method in the field has demonstrated.

This is also where AI implementations stall. Our own read of the market has 88 percent of TA leaders adopting AI in some form and only 11 to 14 percent with anything production ready, with 62 percent stuck in pilots. The teams in that gap did not pick bad vendors. Most of them started with the scoring half, which cannot pass a pilot, instead of the execution half, which passes in a week.

The dividing line is whether the answer is checkable

Every reliable use of AI in recruiting sits on the checkable side of that line, and every unreliable one sits on the other. It is a more useful test than asking whether a task is hard, because coordination is genuinely harder than ranking in the computational sense and it still works better.

A checkable task gives the system a correction signal on the timescale of minutes. The invite bounced, so retry. The interviewer declined, so rebuild. The candidate picked a slot, so confirm and stop. An agent operating there improves against ground truth and its errors surface immediately, which is what makes unattended operation defensible in the first place.

An unmeasurable task gives the system nothing. It produces a number that looks like the checkable kind, carries no error bar, and gets treated as evidence in a decision about a person's livelihood. That is the entire failure mode, and it explains why the same underlying model can be excellent at one recruiting job and unfit for another.

The practical version for a TA leader: automate the stage where the decision has already been made and the work is executing it. Screening decides who continues, so a human owns it. Coordination executes a decision someone already made, so an agent owns it. That boundary also happens to sit exactly where the hours are, which is the arithmetic in full cycle recruiting.

What candidates do when the AI makes the decision

The Greenhouse candidate survey ran across 2,950 job seekers, 1,200 of them in the United States, and the numbers are worth reading as a design brief rather than a verdict on AI.

Candidates are not objecting to automation as such. Look at what they said they walked away from: a pre recorded video interview scored by AI with no human present, 33 percent. Undisclosed AI use, 27 percent. Monitoring during the process, 26 percent. Every one of those is the scoring half. Meanwhile 46 percent want a human interview option and 44 percent simply want to be told, which are requests a team can satisfy this quarter.

Two other figures deserve attention. Only 21 percent of candidates believe most employers use AI responsibly, and 51 percent of people who completed an AI interview got no response at all afterwards. That second number is not an AI failure, it is a coordination and communication failure wearing an AI costume, and it is the thing that turns a candidate into a detractor. Silence reads as a decision, which is the same mechanism behind the drop off in our recruiting funnel benchmarks.

An agent that confirms, reminds and updates without being asked is the cheapest fix available for the trust problem the scoring agents created.

What the law now requires of an AI recruiter

The regulatory line falls in the same place as the reliability line, which is not a coincidence. Regulators wrote rules about systems that evaluate people, not systems that book meetings.

* EU AI Act. Recruitment, selection, targeted job advertising and candidate evaluation are listed in Annex III as high risk. The original 2 August 2026 compliance date was pushed to 2 December 2027 by the Digital Omnibus, which entered into force on 27 July 2026. The obligations themselves have not softened: risk assessment, bias testing, logging, meaningful human oversight and disclosure to candidates.

* Illinois HB 3773. In force since 1 January 2026. Employers must notify applicants when AI is used in an employment decision, and using zip code as a proxy for a protected characteristic is prohibited outright.

* New York City Local Law 144. In force since July 2023. An automated employment decision tool needs an independent bias audit every year, the results published, and candidates given at least ten business days of notice.

Read the compliance burden as a cost that attaches to one half of the category. A coordination agent that never scores anyone is largely outside all three regimes. A scoring or interviewing agent brings an annual audit, a notice obligation and a documented human review step, in every jurisdiction where any of your candidates happen to live. That cost belongs in the business case, and it rarely appears in the demo. None of this is legal advice, and the jurisdiction rules turn on where candidates sit rather than where you do, so it is worth putting in front of counsel before you buy.

How to tell which half a vendor is selling

Five questions that separate the two halves faster than a feature list.

* What does it do when it is not confident? The good answer names a person and describes the exception. The bad answer is a confidence score, which is the model rating its own homework.

* What does it decide, and what does it execute? Make the vendor put every claimed capability into one column or the other. Anything in the deciding column needs the audit trail below.

* Show me the bias audit. If the tool scores or ranks people, the audit already exists or the vendor is not selling into New York City. Ask for the date and the vendor's own name on it.

* How many customers run this unattended? Not licensed, not piloting. Running without a human approving each action. This is the question that separates the 88 percent who adopted from the 12 percent in production.

* What breaks if it is wrong? A misbooked interview costs an apology and a reschedule. A miscalibrated score costs a candidate a job and you a claim under the Illinois statute.

Sequencing matters as much as selection. Start with the stage that has a checkable answer, get it running unattended, then decide whether the judgment tools have improved enough to be worth the compliance overhead. That order is the one that gets teams out of pilot purgatory, and it is the argument in building an agentic AI recruiting strategy.

See how autonomous coordination handles the scheduling stage.

Frequently asked questions

What is an AI recruiter?

An AI recruiter is software that executes recruiting tasks autonomously rather than assisting a human with them, covering sourcing outreach, application screening, interview coordination, scheduling and candidate communication. In practice the term covers several different products with very different reliability, so the useful question is which stage a given tool automates.

Can an AI recruiter replace a human recruiter?

No. It replaces a set of tasks, not a role. Agents reliably take over the execution work such as coordination, scheduling, reminders and status updates. The judgment work, deciding who advances and closing a candidate, stays with a person, and in the EU, Illinois and New York City there are legal obligations attached to letting software make that call.

What can an AI recruiter do well?

Anything with a verifiable right answer. Collecting availability, resolving panel conflicts, booking rooms and links, sending confirmations and reminders, absorbing reschedules, updating the ATS and keeping candidates informed. On our platform this moves a coordination event from 243 minutes of manual work to 27.

What can an AI recruiter not do?

Predict who will succeed in a job. Resume ranking and interview scoring have no reliable ground truth to learn from, since outcome data exists only for people who were hired. A University of Washington audit of three commercial models found white associated names preferred 85 percent of the time against 9 percent for Black associated names.

Is it legal to use an AI recruiter?

Yes, with conditions that depend on what the tool does and where your candidates live. Tools that evaluate candidates trigger notice requirements in Illinois, annual independent bias audits in New York City, and high risk obligations under the EU AI Act from 2 December 2027. Tools that only coordinate and communicate carry far lighter obligations. Confirm your own position with counsel.

Do candidates mind being interviewed by AI?

Many do. Greenhouse found 38 percent of candidates have walked away from a hiring process because of an AI interview, with the top reason being a pre recorded interview scored by AI with no human present. Disclosure helps: 70 percent were never told upfront that AI would evaluate them, and 46 percent want a human interview option.

The bottom line

An AI recruiter is not one thing, and the half of the category that works is not the half that gets the attention. Execution is solved. Judgment is not, and the reason is structural rather than temporary, because one kind of task can be checked in minutes and the other cannot be checked for years.

The teams getting real return from AI in 2026 did not bet on better judgment. They handed over the work with a right answer, kept the work without one, and got a coordination function that runs unattended while everyone else is still evaluating scoring engines. Book a demo and we will run your own coordination numbers through it.