AI Agents in Recruitment: What Works

Recruitment is one of the few places where "should we let AI decide?" comes with named legal obligations attached, not just a product opinion. That makes it easier to scope than most AI projects, provided you start from the constraint rather than the capability.

A working split, stated as risk rather than as law:

• Low risk and genuinely useful: sourcing, scheduling, candidate communication, interview prep, note-taking, and structuring what you already collected
• Higher risk, and where most teams should keep a person deciding: initial screening against explicit, job-related criteria
• Highest risk, and where we'd advise against it: autonomous rejection, ranking candidates by inferred traits, and anything scoring people on characteristics you couldn't lawfully ask about

Be precise about why that last line is a recommendation and not a universal legal conclusion, because the obligations differ sharply by jurisdiction and by how you use the tool:

• EU. Employment screening sits in the high-risk category under the EU AI Act, with those obligations applying from 2 December 2027 after the AI Omnibus amendment extended the original date. High-risk status means duties — risk management, data governance, logging, transparency, human oversight — rather than a ban on automated decisions. The nearer prohibition in Europe is data protection law: GDPR Article 22 gives a person the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, subject to its own exceptions.
• New York City. Local Law 144 requires an annual independent bias audit and candidate notice for automated employment decision tools. It regulates how you use the tool and what you disclose. It does not require that a human make the final call.
• Illinois. The Artificial Intelligence Video Interview Act requires notice before the interview, an explanation of how the AI works and what characteristics it evaluates, written consent, limits on sharing the video, and deletion within 30 days of a request. Employers relying solely on AI analysis to decide who advances must report applicant race and ethnicity data annually — which tells you the statute contemplates automated screening and regulates it rather than forbidding it.
• US generally. The EEOC has been explicit that existing discrimination law applies to algorithmic hiring tools, whoever presses the button.

Two things follow. First, check the rules for where your candidates are, not only where you are. Second — and this is the one teams get wrong — a human in the loop is a risk control, not a safe harbour. A rubber-stamped human approval of a biased ranking is still a discriminatory outcome under the EEOC's reading, still an automated employment decision tool under LL144, and still owes an audit. Human review lowers risk; it does not discharge the obligations, and no jurisdiction here treats it as a defence on its own.

Quick Comparison

Where Agents Genuinely Help

Sourcing

The strongest case. Searching profiles against a brief, enriching them, and assembling a shortlist is bounded research work — exactly what agents do reliably. A recruiter reviews the list and decides who's worth approaching.

The failure to guard against: an agent optimising for what's easy to match. Keyword-heavy profiles rise, unusual backgrounds sink, and you narrow your pipeline while feeling more efficient. Check what the shortlist excludes, not just what it contains.

Scheduling

Wholly mechanical, high volume, and irritating for everyone. Coordinating panels across calendars and time zones has no judgment in it and should be automated without hesitation. This is the least controversial win in the whole function.

Candidate communication

Status updates, next-step explanations, answering routine questions about the process. Candidate experience is mostly a responsiveness problem, and responsiveness is what automation is for.

Keep drafting and sending separate. A tone-deaf automated message to someone you're about to hire is expensive in a way the time saving doesn't cover.

Interview support

Structuring notes, transcribing, mapping answers back to the questions asked. This improves fairness rather than threatening it, because the alternative is an interviewer's selective recollection written up two days later.

For the tooling side, our guide to AI meeting notes covers the recording and consent questions — which apply with extra force to interviews, where the power imbalance makes genuinely informed consent harder.

Interview question preparation

Generating job-related, consistent questions from the actual job description. Structured interviews outperform unstructured ones, and the reason most teams don't run them is that writing the structure is work. Removing that friction is a real quality gain.

Where It Goes Wrong

Screening is the whole risk

CV screening is where the value looks highest and the exposure is greatest. Three specific traps:

Learning from historical hires. If you train or tune on who you hired before, you reproduce whatever your past hiring did — including patterns you'd rather not repeat. The system will look accurate because it agrees with history.

Proxies you didn't intend. Postcode, university, career gaps, and name all correlate with protected characteristics. An agent never told to consider them can still weight them heavily through correlated signals.

Unexplainable rejections. If you can't say why someone was screened out in terms of stated job requirements, you have a problem that is simultaneously legal, ethical, and practical.

The workable pattern: the agent extracts and matches against explicitly stated, job-related criteria, and shows its working. A human decides. Extraction — does this CV evidence the three required qualifications — is a defensible task. Judgment about fit is not one to delegate.

The volume problem cuts both ways

Candidates now use AI to write applications, so volume is up and differentiation is down. Responding with an agent that screens harder produces an arms race where both sides optimise against a model and nobody learns anything about anybody.

The teams handling this well go the other way: fewer, more specific requirements, and a short work-sample task early. A sample of the actual work is harder to fake and more predictive than any CV screen — with or without AI.

Making It Defensible

1. Write the criteria before you automate. If you can't state them explicitly, an agent will infer them, and you won't like what it infers.
2. Keep the reject decision human as a default, and make the review substantive. Advancing someone incorrectly costs an interview slot; rejecting someone incorrectly is the harm regulators care about. Treat this as risk reduction rather than as compliance — the obligations above attach to the tool and the process, and they survive a human sign-off.
3. Log every decision with its reasoning, retained. Local Law 144-style audits and candidate notice requirements are answered with records, not with policy documents.
4. Test outcomes across groups before rollout and periodically after. Bias testing at launch only tests launch.
5. Tell candidates an automated tool is used, what it does, and how to ask for human review. Increasingly a requirement, and it's good practice regardless.
6. Re-test after any model change. A vendor updating the model underneath you changes behaviour without your deployment — the continuous-evaluation point from AI governance tooling.

Key Points

• Sourcing, scheduling, communication, and interview support are the safe wins — and scheduling alone justifies most of the effort
• Screening is high-risk in the regulatory sense, not just the reputational one: EU AI Act high-risk obligations apply from 2 December 2027, and NYC Local Law 144 requires bias audits and notice today
• Obligations vary by jurisdiction and by use — high-risk status means duties, not a ban, and GDPR Article 22 is the closer restriction on solely-automated decisions in the EU
• A human in the loop is a risk control, not a safe harbour — the audit, notice, and non-discrimination duties still apply
• Never train on your own hiring history unless you're confident it's a pattern worth reproducing
• Proxies do the damage — postcode, university, and gaps correlate with protected characteristics without ever being named
• Extraction is defensible; judgment isn't. Match against stated criteria and let a human decide
• Work samples beat CV screening in an era where both sides are using AI
• Log reasoning, test across groups, disclose, and re-test after model changes

If your team's problem is more that the useful AI setups never get running, Taku mirrors a working AI setup into a desktop workspace and runs it, so a workflow someone proved out becomes something you can use and keep. Taku is in Beta, and the Mac app is available now.

FAQ

What can AI agents do in recruitment?

Source and shortlist candidates, run outreach sequences, handle all scheduling, answer routine candidate questions, generate structured interview questions, and transcribe and structure interview notes. Hiring decisions themselves should stay with people.

Is it legal to use AI to screen job applicants?

It's legal in most places but regulated, and the regulation is about process and disclosure more than about prohibition. NYC's Local Law 144 requires an annual independent bias audit and candidate notice; Illinois requires notice, an explanation, and written consent for AI-analysed video interviews, plus demographic reporting if AI alone decides who advances; Maryland regulates facial recognition in interviews; and the EU AI Act treats employment screening as high-risk, with obligations applying from 2 December 2027. In the EU, GDPR Article 22 is the provision that actually restricts decisions made solely by automated processing. Check the rules where your candidates are, not just where you are.

Can an AI agent reject candidates automatically?

In most places it isn't flatly prohibited, and we'd still advise against it. Illinois, for example, contemplates AI-only screening and requires demographic reporting from employers who do it rather than banning it; the EU AI Act imposes high-risk duties rather than a ban, though GDPR Article 22 restricts solely-automated decisions with significant effects. Rejection is the decision with legal exposure and no recovery, and it's where bias claims land. Let an agent surface and evidence its matching, keep the decision human, and take advice for the specific jurisdictions your candidates sit in — this is a risk judgment, not a compliance checkbox.

How do I stop AI screening from being biased?

Don't train on historical hiring outcomes, define explicit job-related criteria in advance, test outcomes across demographic groups before and after launch, watch for proxies like postcode and university, and keep decision logs. Assume bias is present until measurement says otherwise.

Does AI in recruitment actually save time?

Scheduling and sourcing, clearly and immediately. Screening saves less than expected once you add proper review, logging, and auditing — and skipping those is what creates the liability. Budget for the review step or the saving is illusory.

Candidates are using AI to apply. How should we respond?

Not by screening harder — that's an arms race between two models. Narrow your stated requirements and introduce a short, realistic work sample early. Work samples are more predictive than CVs and considerably harder to fake.

Should we tell candidates we use AI?

Yes. It's already required in several jurisdictions, it's likely to become more broadly required, and disclosure with a route to human review costs almost nothing while removing most of the reputational risk.