What a hiring copilot is, and where the law says autonomy stops

A hiring copilot is AI that assists a recruiter in real time, suggesting questions, scoring answers against a rubric, and summarizing evidence, while a human still decides who advances and who gets rejected. An autonomous hiring agent makes those calls itself. A hiring copilot is not a junior agent or a waypoint on the road to one, it is a deliberate ceiling on how much autonomy you hand over.

That ceiling is not a philosophy. In New York City, a version of it is already written into rule text.

You will see the same product sold as an AI hiring copilot or a recruiting copilot. The copilot vs AI agent recruiting argument that follows is one question wearing a costume: who owns the decision.

Key Takeaways

A copilot is the correct ceiling for the assessment decision, not a lesser product you graduate out of on the way to agents.
The difference between a copilot and an agent is not capability. It is who owns the rejection.
NYC's Local Law 144 rules define when a tool substantially assists or replaces discretionary decision making. The law never says copilot or agent, but that three-part test is the sharpest line anyone has drawn.
Calling a tool a copilot does not move it out of the EU AI Act's high-risk bucket. Employment AI is listed there by use case.
People take algorithmic advice least when the decision is serious, which argues for hints a human weighs rather than scores a human resents.
A real copilot lets a person see the reasoning, override it, and interrupt it mid-run. If it cannot do all three, it is an agent with softer marketing.

What a hiring copilot actually is, and why every vendor suddenly has one

A hiring copilot sits next to a recruiter and makes that recruiter better at a decision the recruiter still makes.

Microsoft, which popularized the word, defines a copilot as an AI-powered assistant that provides support for tasks, offers insights, and boosts productivity, and agents as specialized AI tools built to handle specific processes. Agents, it says, can respond to user inquiries in real time, or operate independently, taking specific actions based on data and predefined goals. In Microsoft's framing, the copilot is the interface and the agents are the apps behind it.

That is a useful, boring definition, and the recruiting market has spent two years ignoring it. Browse the category and copilot mostly means "our AI", applied to whatever the vendor already sold you. Most of what gets branded a hiring copilot today is inbox work: job descriptions, candidate summaries, outreach drafts, rejection emails, pipeline updates. Useful, but none of it is the moment where a person is judged.

The interesting version sits at the assessment itself. An AI interviewer running a live technical conversation is forming judgments a recruiter then has to own, which is where the copilot question stops being a branding exercise and becomes a governance one.

Copilot or agent: the only difference that matters is who owns the decision

Capability does not separate a copilot from an agent. Ownership of the decision does. The same model, at the same accuracy, is a copilot when a person weighs its output against other evidence, and an agent when its output is the thing that rejects the candidate.

Most vendors sell this as a maturity ladder, with copilots as the training wheels you outgrow. That framing is convenient if you are selling autonomy, and it quietly skips the question a buyer should be asking. Not how autonomous can this get, but where should autonomy stop.

Five questions settle it, and you can ask all five in a demo:

Can the tool reject a candidate with no person in the loop?
Whose name is on the rejection when the candidate asks why?
Can the human see the reasoning before the decision takes effect, or only after?
Is the score one input among several, or the input that decides?
Can a person interrupt the tool while it is running, not just review it afterwards?

Answer yes to the first one and you own an agent, whatever the website says. Question five is the one almost nobody can answer, and we will come back to it. The trade-offs at the interview itself get more room in our comparison of AI interviews and human interviews.

The line is already written into law, and it is sharper than the marketing

New York City has already codified the copilot and agent distinction, though it never uses either word. Local Law 144 of 2021 and the rules that implement it define when an automated tool "substantially assists or replaces discretionary decision making", and that phrase is the line the whole category keeps talking around.

The New York test, in its own words

Under the adopted DCWP rules, a tool substantially assists or replaces discretionary decision making when an employer does any one of three things:

Relies solely on a simplified output such as a score, tag, or ranking, with no other factors considered.
Uses that simplified output as one of a set of criteria, where it is weighted more than any other criterion.
Uses the simplified output to overrule conclusions derived from other factors, including human decision-making.

Read that second and third clause slowly, because they are stricter than most "human in the loop" marketing. A person being present is not enough. If the AI score outweighs every other criterion, or gets used to overturn what your interviewers concluded, you are inside the definition even though a human clicked the button.

Cross that line and the tool is an automated employment decision tool. New York then requires a bias audit within one year of use, public posting of the audit results, and notice to candidates. DCWP began enforcing the law and rule on July 5, 2023.

To be plain about it, the law does not say copilot and it does not say agent. It regulates a function, not a product category. But it is the only place I know of where the assist-versus-decide distinction carries an actual legal test, which makes it a better definition than anything printed on a pricing page.

Does calling it a copilot dodge the EU's high-risk bucket?

No. The EU AI Act classifies by use case, not by how much autonomy a vendor claims. The European Commission's own summary lists "AI tools for employment, management of workers and access to self-employment (e.g. CV-sorting software for recruitment)" among high-risk uses, and says high-risk systems face strict obligations starting on December 2, 2027, including "appropriate human oversight measures".

Which is the quiet joke inside the copilot pitch. Human oversight is not what exempts you from the rules, it is one of the things the rules require.

Colorado went down the same road and then turned around. SB24-205 drew a similar line, turning on whether a system was a substantial factor in making a consequential decision, but it was repealed before it ever took effect and replaced by SB 26-189, which regulates automated decision-making technology in consequential decisions from January 1, 2027. Confirm the current compliance dates with counsel rather than with a blog post, including this one.

We keep a plain-English summary of these regimes in our guide to AI hiring laws. None of it is legal advice.

Where a copilot belongs in the hiring loop, and what stays human

A copilot belongs anywhere its output is evidence, and it stops where that output becomes a verdict. That is a clean operational rule, and it maps onto specific steps in a real loop.

Safe to automate outright: scheduling, reminders, interview notes, transcript search, and the first draft of a rubric.
Copilot territory: question selection during a live interview, scoring an answer against a rubric, flagging a contradiction, and summarizing evidence for a debrief.
Human, permanently: the reject decision, the offer decision, and any judgment about a candidate that your published rubric does not cover.

Structure is what makes any of this work, by the way, not the AI. An unstructured interview with a copilot bolted on is still an unstructured interview, and the hints are guesses. Structured interview software gives the copilot something defensible to score against.

What a copilot looks like inside a live interview, not just in your inbox

Go read the copilot feature pages on ATS vendor sites. What they describe is work that happens around the interview, drafting, summarizing, and updating the pipeline, rather than the interview itself. Expert Hire's copilot sits inside it, and the mechanism is specific enough that you can check it rather than believe it.

On the Growth plan and above, a recruiter can join a live Expert Hire AI interview as a silent observer. When the conversation needs a person, that recruiter can take over mid-interview: the AI pauses, the recruiter speaks directly to the candidate, and control returns to the AI on demand.

That is the interrupt capability from question five. Ask your shortlist that question and see how many vendors can answer it.

For final rounds the arrangement flips. A human interviewer takes the seat, the AI steps back, and copilot-style hints surface in real time while the person runs the conversation. That is the setup we make the case for in the human-led interview.

Either way the artifact is the same. Every interview produces a scorecard carrying the rubric criteria, a score per criterion, transcript excerpts, the code the candidate wrote, and the AI's written reasoning for each score. A recruiter can read that reasoning and disagree with it, which is the entire point of building it this way. How AI interviews are scored walks through one.

The scorecard then syncs back into the candidate record in Greenhouse, Lever, or Ashby, so the evidence sits where the decision gets made rather than in a separate tool nobody opens. The AI interview platform page describes how an interview runs and what the scored report at the end of it contains.

The uncomfortable part: people trust algorithms least when the stakes are highest

Here is the finding that complicates every "a human always decides" reassurance, ours included. People reject algorithmic advice more often as the decision gets more serious, which is precisely backwards from where you want them careful.

In a 2023 PLoS One experiment by Filiz, Judek, Lorenz, and Spiwoks, 143 student participants chose between a 70% accurate algorithm and a 60% accurate human expert across scenarios of varying gravity. In the serious scenarios, only 50.7% picked the algorithm. In the trivial ones, 70.83% did. Each one-unit rise in perceived gravity cut the probability of choosing the algorithm by 3.9%.

The authors call it the tragedy of algorithm aversion. Hiring is a high-gravity decision, so the prediction is that recruiters will under-use good machine advice exactly where a bad call costs the most.

So you get failure in two directions at once, and most vendors write about only one of them. Automation bias is deference to a score nobody understands. Algorithm aversion is dismissing a score that was right. A copilot is the shape that fights both, because a hint a person weighs is harder to rubber-stamp than a ranking and harder to resent than a verdict.

That is a design argument, not proof. The closest thing to real evidence cuts against us. Hoffman, Kahn and Li (Quarterly Journal of Economics, 2018) studied job testing across 15 firms and found managers who hired against the test's recommendation made worse hires on average.

Human override is not automatically the safe setting. It is a check that can itself be wrong, which is why the reasoning has to be visible enough to argue with.

Frequently asked questions

What is a hiring copilot, and how is it different from an AI recruiting agent?

A hiring copilot assists a recruiter in real time and leaves the decision with the human. It suggests questions, scores answers against a rubric, and summarizes evidence. An AI recruiting agent acts on its own, advancing or rejecting candidates with no person in the loop. The test is not capability, it is who owns the rejection and whether a human can see the reasoning and override it before it takes effect.

Does calling an AI hiring tool a copilot keep it out of high-risk classification under the EU AI Act?

No. The EU AI Act classifies by use case, and the European Commission lists AI tools for employment and recruitment, including CV-sorting software, among its high-risk uses. Human oversight is not an exemption from those obligations, it is one of them. Strict obligations for high-risk systems apply starting December 2, 2027.

How does an AI interview copilot integrate with an existing ATS like Greenhouse, Lever, or Ashby?

Expert Hire has live integrations with Greenhouse, Lever, and Ashby. The candidate is pushed from the ATS, the AI interview runs, and the completed scorecard syncs back into the candidate record with the rubric, transcript excerpts, and the AI's reasoning attached. Slack alerts and calendar scheduling run off the same workflow, so the evidence lands where the hiring decision gets made.

Can a recruiter join or take over an AI interview while it is still running?

In Expert Hire, yes, on the Growth plan and above. A recruiter can join a live AI interview as a silent observer, then take over mid-interview when the conversation needs a person. The AI pauses, the recruiter speaks directly to the candidate, and control returns to the AI on demand. Ask any vendor on your shortlist whether a person can interrupt the tool mid-run rather than only review it afterwards.

Is a hiring copilot subject to NYC Local Law 144 bias audits if a human makes the final call?

Often, yes. Local Law 144 applies when a tool substantially assists or replaces discretionary decision making, and the DCWP rules define that to include weighting a simplified output more heavily than any other criterion, or using it to overrule conclusions drawn from other factors. A human clicking the final button does not by itself put you outside the definition. Ask counsel about your specific configuration.

Which hiring decisions should stay human when you use an AI copilot?

The reject decision, the offer decision, and any judgment your published rubric does not cover. Scheduling, reminders, note-taking, and transcript search can run without you. Scoring, question selection, and evidence summaries are copilot work, where the output is one input a person weighs. Once the AI's output is the deciding input, you have crossed from copilot into agent.

A copilot is the ceiling, not a waypoint

Most of this category is arguing about how autonomous a hiring tool can get. For the assessment decision, that is the wrong question. New York already answered the right one in rule text: once the machine's output outweighs every other criterion, or overturns what the humans concluded, it has stopped assisting.

A copilot that stays on the correct side of that line is not an immature agent. It is a product that knows where it is supposed to stop, and that is a design choice worth paying for rather than a limitation to apologize for.

If you want to check whether we hold that line ourselves, read the Expert Hire methodology. It documents the IO psychology research behind the rubric, the rubric design, and the scoring approach, and it is the page we would want a regulator to read first.

By TK, Growth at Expert Hire. Last updated September 22, 2026. Reviewed by Anand Suresh, CPO at Expert Hire.