AI Recruiter: What the Job Looks Like When You Actually Use One

AI Recruiter: What the Job Looks Like When You Actually Use One

Cristian García is Nova's Chief Product Officer, across the Nova Talent network and Nova Recruiter. His read on AI recruiting: the tools are ready, and almost nobody is set up to actually use them.

Updated 9 min read

An AI recruiter is a recruiter who delegates the boring half of the week to an agent and keeps the judgment. Not a bot that screens people for you, and not another layer of automation bolted onto your ATS. The setup is closer to onboarding a junior: you explain how you hire once, and from then on the repetitive work gets done while you’re doing something else.

The reason this matters now isn’t that the models got better. It’s that they can finally reach your actual tools, which is the difference between an assistant that writes nicely and one that can pull a shortlist, check it against your ATS, and leave four drafts in your inbox before Monday starts.

Most recruiters who’ve tried this gave up at the search bar. Here’s what the job looks like past it.

What actually changes in your week

Think about the Monday you have now. Weekend replies sitting unread. Three roles opened Friday with nothing in them. The same outreach intro, rewritten for the thirtieth time with a different name at the top.

Now the version with an agent running underneath:

  • The weekend replies are already triaged, each with a draft response under it, in the candidate’s language.
  • The three roles have a fresh shortlist, with a paragraph on each person explaining why they fit.
  • The week’s outreach is written one by one rather than one template thirty times.
  • Friday’s report is already in drafts.

None of that is a decision. That’s the point. Every item on that list is work you’d do yourself if nobody else could, and none of it is the reason anyone hired you.

Do this: write down the three things you did last week that you’ve done more than fifty times. That list is your first delegation, and it’s almost certainly more specific than “use AI.”

The three layers, and why most setups fail at one of them

An agent that feels useless is almost always missing one of three things. Naming them makes the gap obvious.

Memory is who you are. What you recruit for, how you screen, what disqualifies someone, what tone you write in. Without it, every session starts from zero and you spend the first ten minutes re-explaining your own job.

Skills are how you do things. A workflow written down once and run the same way forever: building an ICP from a brief, writing three outreach variants, producing a two-sentence fit rationale per profile. The test is simple. Any prompt you’d type twice belongs in a skill.

Access is reach. Your inbox, your ATS, your sourcing platform. Without it the agent is articulate and blind, and you’re back to pasting threads into a chat window.

Most people set up the first one, skip the other two, and conclude the technology is overhyped. It’s the same as giving a new hire a handbook, no process, and no logins.

Do this: don’t write your memory file. Have the agent interview you, one question at a time, and write it for you. Ten minutes of talking produces something more accurate than an hour of typing, because when you type you summarize. When you talk you mention the hiring manager’s complaint from last time and the deadline nobody wrote down.

Parallel for coverage, sequential for depth

This is the one structural idea worth taking away, and it’s the difference between an agent that’s mildly helpful and one that compresses a day into an hour.

When several pieces of work could each be done independently against the same goal, run them at the same time. Sourcing is the obvious case: three angles on the same role, opened in parallel, merged at the end. Mapping five target companies at once rather than five rounds of “okay, now the next one.”

When each step needs the previous one’s output, chain them. Vetting is the case: check technical fit, then seniority, then red flags, with each pass writing its notes onto the candidate’s record before the next one runs. Each layer reads what the last one found, so the final read is deeper than any single pass.

Sequential is for depth. Parallel is for coverage. Run the wrong one and you either get four shallow opinions or one very slow pipeline.

Done properly, a vetting chain gets through about 30 candidates in five minutes, with a score and a note on each from all three passes. The two hours of manual review that used to produce the same thing are the part that disappears, and the judgment is still yours, applied to a shorter list.

Do this: before delegating anything multi-step, ask one question. Does step B need what step A found? Yes means chain it. No means run them together.

The deletions are the valuable part

Here’s the habit that separates people who get better results every week from people who get the same mediocre results forever.

When the shortlist comes back, open it and remove everyone who doesn’t genuinely convince you. Then, instead of moving on, ask what the removals had in common. Too junior. Wrong company stage. Impressive on paper, nothing built. The pattern in your deletions is your actual criteria, which is almost never what you wrote in the job description.

Feed that pattern back into the file that defines the role, and the next search starts from it. Six rounds in, the first pass returns roughly what you’d have hand-picked.

This is also the honest answer to anyone worried about AI flattening hiring standards. The system doesn’t have taste. It has yours, to the extent you’ve bothered to teach it, and the teaching happens through what you reject.

Do this: after every shortlist, spend two minutes on one question: what did I cut, and why? Write the answer where the agent can read it.

Where it gets dangerous, said plainly

Three places, and they’re worth more attention than any productivity gain.

Scheduled work should never send. Routines that fire while you’re asleep are genuinely useful for triage, interview prep and weekly summaries. Give one of them permission to send email and sooner or later it sends something embarrassing, under your name, to a candidate who screenshots it. Drafts and summaries only.

Your ATS key is the whole database. Connecting an ATS usually means handing over a key with full access to candidate data: names, emails, interview notes, offers, salary expectations. Left sitting in a plain text file on your laptop, every backup and every mis-shared folder is a path out. A leak from your machine is a GDPR-relevant breach with your name on it. Putting that key in a password manager takes five minutes and it’s the only part of this setup that’s genuinely urgent.

The decision stays human, and now it’s also the law. The EU AI Act classes candidate evaluation systems as high risk, and the responsibility sits with the employer deploying them, not the vendor. The gap between what employers think they’re doing and what they’re actually doing is already documented: the UK’s data regulator reviewed automated decision-making in recruitment across more than 30 organizations and found that while most believed they were using these tools as decision support, in practice they were making solely automated decisions without meaningful human involvement. Its test is specific. Meaningful review means a person with the authority, discretion and competence to change the decision before it’s applied. A rubber stamp doesn’t count, and neither does having designed the system yourself.

Do this: automate the yes, never the no. Advancing someone to a conversation is reversible. Rejecting them isn’t, and it’s the decision you’d have to defend.

What this has to do with where the candidates come from

An agent only works on the pool it can reach, which is why the sourcing layer underneath it matters more than it used to.

Nova Recruiter indexes 800M+ public profiles and ranks them with a Talent Score built on the quality of someone’s experience, education, international exposure and languages, rather than on how well they described themselves with the right keywords. That distinction does a lot of work. A senior engineer who spent ten years building distributed systems but never wrote “distributed systems” on her profile is invisible to keyword search and perfectly visible to merit-based ranking.

Underneath that sits something a bigger index can’t replicate: a curated network of 25,000+ professionals who’ve been assessed and who opted into being discovered. Reaching one of them through Nova is a different act from sending a cold InMail to a stranger. It’s the same row in a comparison table and a completely different thing in practice, and it’s part of why multichannel sequences land around 2.5x the replies of a single channel.

For the agent setup specifically, the practical part is that Nova Recruiter connects to Claude, ChatGPT and Cursor, with human approval required before anything gets created. The free tier doesn’t ask for a card, so the setup above can be tested on a real role before anyone signs anything.

How to start, this week

Not with a system. With one task.

  1. Pick the thing you repeated most last week.
  2. Spend ten minutes letting the agent interview you about how you do it, and keep the file it writes.
  3. Connect one tool. Your inbox is usually the one that pays off fastest.
  4. Run it on a live role, then delete everyone who doesn’t convince you and write down why.
  5. Add the next layer when the first one has earned it.

If you’d rather have the whole thing walked through, with the prompts and the troubleshooting, that’s what Become an AI Recruiter covers across sixteen lessons.

The recruiters pulling ahead right now aren’t the ones with the best tools. They’re the ones who’ve written down how they hire, which turns out to be useful whether or not an agent ever reads it.

For in-house recruiters and small agency teams running more than two roles at once, this setup pays for itself in the first week. If you’re hiring once or twice a year, the honest answer is that the setup time costs more than it saves, and a good sourcing tool on its own will serve you better.

Frequently asked questions

What is an AI recruiter?

An AI recruiter is a recruiter working with an AI agent that handles the repetitive layer of hiring: building search strings, pulling and deduping candidates, drafting individual outreach, summarizing replies, and preparing reports. The agent does the volume. The recruiter decides who moves forward, how the role is positioned, and what a weak signal means. It isn't an autonomous system that hires people, and anyone selling it that way is overselling it.

Will AI replace recruiters?

Not the ones doing the judgment work. What AI replaces is the part of recruiting that's effectively data entry: reformatting profiles, chasing duplicates, rewriting the same intro. What it can't replace is deciding whether a non-obvious background is a risk or an edge, or convincing a passive candidate that a conversation is worth their time. The gap that's opening is between recruiters who delegate the boring layer and recruiters who still do it by hand.

Do you need to know how to code to use an AI recruiting agent?

No. The setup is a folder on your computer, a few plain-text files describing how you hire, and connections to the tools you already use. The files are written in normal language, and you can have the agent interview you and write them for you. If you can brief a junior recruiter, you can brief an agent.

How long does it take to set up?

The first useful version takes an afternoon: one folder, one memory file, one connected tool. Where people lose time is trying to build the whole system before running it once. Start with the single task you repeat most this week, get that working, and add the next layer when the first one has earned it.

What should an AI agent never be allowed to do in recruiting?

Send anything outward on its own, and make the final call on a candidate. Scheduled tasks that run while you're away should produce drafts and summaries, never outbound messages. On decisions, the EU AI Act treats candidate evaluation systems as high risk and puts the responsibility on the employer using them. The UK's data regulator has also set a specific bar for what counts as human review: a person with the authority, discretion and competence to change the decision before it takes effect. Approving a list without the standing to alter it doesn't meet it.

How is this different from the AI already built into recruiting tools?

Most tools apply AI inside their own walls: better search in the sourcing tool, better matching in the ATS. An agent sits above them and works across them, so it can pull a shortlist, check it against your ATS for people you've already spoken to, and draft the intro email in one pass. The two aren't alternatives. The agent is only as good as the tools it can reach.

Sources and method

The EU AI Act's classification of candidate evaluation systems as high risk, and the allocation of responsibility to the deploying employer, follow the Act's own text. The findings on automated decision-making in recruitment are from the UK Information Commissioner's Office report and draft guidance published 31 March 2026, based on a review of more than 30 organizations between March 2025 and January 2026, https://ico.org.uk. Nova Recruiter figures (800M+ profiles, 25,000+ vetted members, 2.5x reply rate, 60+ ATS integrations) are Nova's own, as published on the product page.

Cristian García
Cristian García

Chief Product Officer

Cristian García is Chief Product Officer at Nova, leading product across the Nova Talent network and Nova Recruiter. He previously built Banco Santander's global app as a product manager, after starting out in risk and technology at BBVA.

Connect on LinkedIn →
AI in recruiting How to Connect Your ATS to an AI Without Leaking Candidate Data AI in recruiting Recruitment Automation: From 30 Hours to 30 Minutes