What AI in recruitment means, and what it does not

AI in recruitment is software that reads, writes, ranks or suggests during the hiring process, using models trained on large amounts of text instead of rules someone typed in. That covers four quite different things, and most confusion in buying conversations comes from treating them as one. A hiring manager who says "we want AI" and a security reviewer who says "we ban AI" are often talking about different features.

The glossary entry on AI recruitment gives the short definition. This guide is about the decisions behind it: which parts earn their place, which parts need a person standing next to them, and what the rules now say. Where we describe how something works, we describe how it works in Teamtailor and say so. We make an applicant tracking system, so we have a view, and we have tried to keep it separate from what holds for every vendor.

Four kinds of assistance

KindWhat it doesTypical exampleWhere the risk sits
Writing and summarisingProduces or condenses textA first draft of a job ad, a summary of a CV or a video meetingWrong or generic wording, a summary that drops the one detail that mattered
Ranking and matchingOrders people against a jobCandidates sorted by fit, candidate matching against a talent poolBias in the data, and a ranking that quietly becomes the decision
ConversationAnswers candidates in real timeA recruiting chatbot on the career siteWrong answers given with confidence, a candidate who cannot reach a person
Rules and automationDoes what someone configured, when something happensA message sent when a candidate reaches a stageConfiguration mistakes, not model mistakes

The last row is not AI in the strict sense, and it is the one most teams get the most value from first. A trigger that sends the same, correct message every time a candidate moves is predictable and easy to audit. We come back to that in the chapter on rollout.

How common it is

SHRM's 2025 survey of 2,040 HR professionals, fielded in February 2025, found 43% reporting AI use in HR tasks, up from 26% in 2024.[S12] The respondents work at organisations across the US and the data is unweighted, so read it as direction, not as a headcount of companies. It covers HR tasks in general, not recruiting alone.

A Tuesday detail that explains the adoption curve: a recruiter pastes a candidate's CV into a free chatbot to get a summary before a call, because the call is in ten minutes. Nobody approved it and nobody was careless. The question for a team is not whether this happens, but whether the approved route is easier than the unapproved one.

Where AI helps in the hiring process

The tasks where assistance holds up share a pattern: the output is text or a shortlist that a person reads before anything happens to a candidate. If a person reads it and can change it, a mistake costs a minute. If nobody reads it, a mistake costs a candidate.

Drafting: job ads, questions, messages

A job ad starts from a hiring manager's three lines and the last ad for a similar role. A draft saves the blank page, and the recruiter's edit is where the company's own voice goes back in. In Teamtailor, the support centre lists drafts of job descriptions, question suggestions for interview kits and applications, translations of questions and drafts of rejection emails among the assistance available.[S1]

The rejection email is a good test of your own standards. A draft that is warm but vague is worse than none. The useful version names the real reason where there is one and goes out under a name a person recognises. If your team already has a default rejection that candidates have replied to with thanks, keep it and let the assistance only adapt the details.

Summarising: CVs and meetings

Summaries are the feature most recruiters use first, because the checking is cheap: you have the original in the next tab. In Teamtailor, the support centre lists CV summaries, meeting insights and interview kit answers among the assistance.[S1] The detail to watch is what a summary leaves out. A summary of a ten year career is a choice about which two years matter, and that choice should be one the recruiter could have made.

One interviewer typing notes into a phone in the car park after the final round is the situation this is built for. The notes were going to be thin either way. A summary of the meeting gives the next interviewer something to read, and the interviewer's own scorecard still carries the judgment. See interview kits for how Teamtailor structures that.

Searching what you already have

A talent pool grows every time someone applies and is not hired. Searching it by hand rarely happens, because the people in it applied for different jobs under different titles. In Teamtailor, the assistance includes suggesting candidates you already have.[S2] The value is in recall: someone who applied eight months ago for a neighbouring role and would now be a strong shortlist candidate. The decision to contact them is still the recruiter's, and the data protection question of how long you may keep and reuse their data still applies. See GDPR in recruitment.

Reading the numbers

The support centre also lists a chart builder for reports.[S1] This is the quietest use and sometimes the most useful: asking for time to hire by source and by team in plain words instead of building a filter. It only produces something trustworthy if the process runs through one system, as the reporting page explains.

Where AI fails, and what the evidence says

Three failures account for most of the harm, and none of them is hypothetical.

Bias carried in from the past

A model learns from what it was shown. In 2018 Reuters reported, in the version HR Reporter republished, that Amazon had built an experimental recruiting tool, starting in 2014, trained on resumes submitted over ten years, most of them from men. The tool downgraded resumes containing the word "women's" and graduates of two all women's colleges. Amazon edited the programs to neutralise those terms, but Reuters noted that this gave no guarantee the system would not find other discriminatory patterns, and the team was disbanded.[S11] The tool, as reported, was never relied on alone by recruiters.

A newer test points the same way. University of Washington researchers varied 120 first names across more than 550 real resumes and compared them against more than 500 job listings in nine occupations, over three million comparisons in all. The models they tested preferred white associated names 85% of the time against 9% for Black associated names, and male associated names 52% of the time against 11% for female associated names.[S10] The authors approximated real systems, since many are proprietary, and tested only names, three models and nine occupations, so the figures do not describe any specific vendor's product. What they show is that the problem has not gone away, and the lead author noted that, aside from a New York City law, there is no regulatory, independent audit of these systems.[S10]

The practical reading: any feature that ranks people needs a way for you to see on what basis, and a person who can disagree with it. Our glossary pages on hiring bias and unconscious bias cover the human side of the same problem. People are biased too, and the case for assistance is only strong when its bias is easier to find and correct than a person's.

Confident wrong answers

Generative models produce fluent text whether or not it is true. Teamtailor's support centre says plainly that accuracy of the output cannot be guaranteed and strongly encourages customers to verify it.[S2] That is the right posture for every vendor: assume the draft is a draft. A chatbot that tells a candidate the role is remote when it is hybrid does real damage, and it does it at three in the morning when nobody is watching.

The ranking that becomes the decision

The subtlest failure is behavioural. A list sorted by fit is read from the top. Candidates below the fold are not rejected by anyone, they are simply never opened. Over a few months the sort order becomes the process. The fix is a rule about use, not a setting: a ranking narrows what the recruiter reads first, and the recruiter opens anything that could plausibly fit regardless of position.

A test for any ranking feature: ask the vendor to show a case where the system's top candidate is wrong, and what in the interface lets the recruiter see that. A vendor who cannot show one has not looked.

What "AI ATS" should mean when you are buying

People who search "AI ATS" are usually asking one of two things: does my applicant tracking system have assistance in it, or should I add a separate tool next to it. Both are reasonable. The trade-off is where the data goes and who can see what.

A separate tool reads CVs and sends back scores. It needs a copy of candidate data, an integration to keep it current and a second set of access rules. Assistance inside the system works on records that are already there, under the permissions already set, and leaves a trail in the same place as everything else. The cost is that you take the vendor's choices instead of picking the best tool for each task. Fewer places where candidate data lives is easier to govern. Weigh that against what a specialist model adds, most of all when you hire at high volume. For high volume with unusual requirements, the answer can differ. Our complete guide to applicant tracking systems covers the rest of the buying checklist.

Questions to put to any vendor

  1. What does it read, and what does it write? Name the fields. A summary that reads only the CV is a different product from one that reads messages and notes.
  2. Who switches it on? In Teamtailor the assistance is an add-on that only Company Admins can activate.[S2]
  3. Where is the data processed, and for how long is it kept? The answers differ by feature, as the next chapter shows.
  4. Is customer data used to train the vendor's or a provider's models?
  5. Can a recruiter see why a candidate was suggested?
  6. Can you see what ran and when? Logs are a buying criterion now, for reasons covered in the chapter on the law.
  7. What happens when the feature is wrong? Ask for the correction path, not the accuracy claim.

How the assistance is arranged in Teamtailor

In Teamtailor, the assistance sits in the same candidate record and job editor the team already uses. The support centre groups the available assistance into six areas: understanding candidates, preparing interviews, meetings and feedback, creating content, analytics, and additional tools such as CV parsing. A Company Admin activates the add-on.[S1] The product overview is on the AI recruiting software page.

What the rules say: EU AI Act, GDPR and New York City

This chapter is a map, not legal advice. Take it to your counsel or data protection officer with the primary texts.

The EU AI Act

The AI Act entered into force on 1 August 2024. Prohibitions on certain practices and the AI literacy requirement have applied since 2 February 2025.[S4] The literacy requirement reaches every team using these tools, whether or not they are high risk.

Annex III lists systems used for the recruitment or selection of people as high risk, in particular placing targeted job ads, analysing and filtering applications and evaluating candidates.[S5] According to the implementation timeline, the obligations for those Annex III systems apply from 2 December 2027.[S4] Which of your tools fall inside that list depends on what they do, not on what the vendor calls them. A tool that drafts a job ad is a different matter from one that screens applications.

For an employer using such a system, the Act sets out what a deployer does:

  • Assign human oversight to people who have the competence, training and authority for it.[S7]
  • Keep the automatically generated logs you control for at least six months, unless other law says otherwise.[S7]
  • Before first use, tell workers' representatives and the affected workers that they will be subject to the system. For a hiring tool the affected people include candidates, and the exact duty towards them is a question for your counsel.[S7]

The oversight article asks that the people overseeing the system can understand its capacities and limits, stay alert to over-reliance on its outputs, and decide not to use it or to override its output.[S6] That describes the ranking problem from the previous chapter in legal language: the law expects someone who can say no.

Teamtailor's trust centre lists Responsible AI Principles, an AI Act page and an ISO/IEC 42001:2023 badge, and its support centre states that the assistance complies with the AI Act requirements already in force.[S2][S3] Read the documents and the scope of the certificate yourself, as you would for any vendor.

GDPR and automated decisions

Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing, including profiling, where it has legal or similarly significant effects. There are exceptions for contract necessity, authorisation by law and explicit consent, and where they apply the controller must provide at least the right to obtain human intervention, to express a view and to contest the decision.[S8] A rejection decided only by a score is the textbook case. A recruiter who reads the profile and decides is outside the problem, as long as that reading is real.

New York City Local Law 144

If you hire in New York City, employers may not use an automated employment decision tool unless it has had a bias audit within one year of use and information about the audit is public. Notices must be given to candidates, and the city's department clarified that the notice must be provided 10 business days before the tool is used. Enforcement began on 5 July 2023.[S9]

Other US states and other countries are moving too, and rules change faster than guides. Check the date on any summary, ours included: the sources behind this page were read on 29 October 2026.

Candidate data: what to ask before you switch anything on

Most features in this area work by sending text to a model run by someone else. That makes the questions concrete: which provider, which region, how long is the text kept, and who may use it for anything else.

What Teamtailor states

In Teamtailor, the assistance runs on OpenAI's Enterprise API, not the public ChatGPT product. Customers choose between EU (Ireland) and US West (Oregon).[S2] Retention differs by feature:

CaseWhat the support centre says
Most featuresOpenAI does not retain API input and output, and zero data retention is enabled[S2]
Suggesting existing candidates (batch processing)OpenAI may retain input and output for up to 24 hours[S2]
CV parsing (background processing)OpenAI may retain input and output for up to about 10 minutes[S2]
Abuse monitoring logsMay be retained by OpenAI for up to 30 days[S2]
Interview transcripts (interview kit answers and meeting insights)Generated by AssemblyAI, a separate sub-processor[S2]
Meeting recordingsVideo is deleted automatically after 30 days unless saving is enabled. Transcripts and summaries stay and can be deleted by hand from the candidate card[S13]

On training: OpenAI is a sub-processor committed to using the data only to provide the service, not for its own purposes such as training its models.[S2] The page also says OpenAI applies human intervention in the design and development of its models.[S2] That sentence is the sort a data protection officer will want to read twice, which is why we put it here instead of leaving it for them to find.

What to check in a vendor that is not us

The same shape of answer should come back from anyone: provider, region, retention per feature, training use, sub-processor list and the date the page was last updated. A vendor who answers "secure and compliant" without a retention figure has not answered. If the data protection officer's first reaction is to ask for a table, give them a table. The GDPR recruitment entry and the trust centre are a start.

Rolling it out without losing the recruiters

Teams that do well with assistance rarely begin with the most impressive feature. They begin with the one whose mistakes are cheapest.

Start where a person always reads the output

A sensible order for most teams is drafting, then summaries, then interview preparation, then search across existing candidates, and ranking last. Each step adds a decision that used to belong to a person. By the time you reach ranking, the team has a feel for where the assistance is right and where it needs correcting, and a habit of reading before acting.

Write down what a person decides

One page is enough. Who decides to reject? Who decides to invite? What may the assistance do without anyone looking? A usual answer is that it may draft and suggest, and that every message to a candidate and every stage change is sent or made by a person or by a rule someone configured. That page is also the start of the oversight record the AI Act expects of deployers of high risk systems.[S7]

Use rules for what rules do well

Not every time saving needs a model. In Teamtailor, triggers send messages and surveys and share candidate profiles when a defined event happens, and the behaviour is the same every time. A recruiter who has configured a trigger can say exactly what it does. For many of the repetitive tasks in a funnel, that predictability beats a model, and see the glossary on recruitment automation for the range.

Measure something you already track

Choose two numbers from your existing reporting, such as time to hire for one role family and the share of candidates who receive an answer, and look at them before and after for the same roles. Use your own history. Benchmarks from other companies are about their processes. The candidate screening entry describes where a funnel usually loses time.

Tell people

Tell your recruiters what the assistance is for, tell hiring managers what it does not do, and tell candidates where it touches their application. A short line on the career site and in the privacy notice costs little. Say it in plain words before a candidate has to ask.

Candidates are using it too

The applicant on the other side has the same tools. Cover letters have become smoother and more alike, and a CV that was tailored to a job by a model looks tailored to every reader. A recruiter who notices three near identical openings in one morning has learned one thing: the cover letter no longer carries the information it used to.

Two responses work. Ask for something a model cannot easily produce from the job ad, such as a short answer about a specific part of the work. And move the weight to the stages where a person is in the conversation. A structured interview with the same questions and a shared scorecard shows differences between candidates that a polished application hides. For further reading, see our article on hiring with AI and ChatGPT and the one on keeping the human touch.

Penalising candidates for using AI to write clearly rarely finds the applicant you are worried about. Judge the substance, and test it in the conversation.

Next steps

If you are deciding whether to use assistance in your hiring, a plain order of work looks like this.

  1. List the tasks your recruiters repeat weekly and mark which ones end in a person reading the output.
  2. Write the one page on what a person decides.
  3. Ask your current vendor, or any vendor you are considering, the seven questions in this guide and ask for the answers in writing.
  4. Start with drafting and summaries, review after a month, and add search across existing candidates next.
  5. Give your data protection officer the retention table and the sub-processor list early, not at the end.

If you want to see how this is set up in Teamtailor, with the admin settings, the candidate record and the reports in one place, book a demo and bring your list. Pricing is quoted per organisation, and the pricing page shows what is included. For the wider picture of what an applicant tracking system should carry, the complete guide is the place to continue.