Introduction
Updated: August 2026 · By Álvaro Arrescurrenaga, CEO of Voicit
You've been reading for months that artificial intelligence is going to transform human resources. And meanwhile, you're still spending half an hour writing the report for each interview, struggling with incomplete notes and the feeling that you missed something important the candidate said.
This guide focuses on the latter. Not on what AI promises, but on what it does today in a real selection team: what tasks it performs, where it makes mistakes, what European regulations require starting this month, and where to begin if you're not a technician.
In short: Artificial intelligence in human resources is primarily used for eliminate administrative workTranscribing interviews, writing reports, preparing minutes, and organizing information. Where no Its usefulness lies in deciding who to hire: there it amplifies biases and the AI Act classifies it as a high-risk system.
Content
What artificial intelligence in human resources is and what it is not
In short: It's a layer that automates information tasks—transcribing, summarizing, writing, classifying—within the people cycle. It's not a system that makes hiring decisions, and treating it as such is the most expensive mistake you can make.
It is important to separate two things that are constantly mixed up in conversations about AI and human resources:
| AI that processes information | AI that decides about people | |
| What are you doing | Transcribe an interview, write the report, summarize a meeting, organize data | Score candidates, discard resumes, and compile a list of finalists. |
| Who decides | The person with the best information ahead | The system, and the person ratifies |
| Risk of bias | Low: does not value, transcribes | Stop: inherits biases from the data it was trained on |
| AI Act | General use, transparency obligations | High risk (Annex III, Employment) |
| Real savings | Direct and measurable in hours | Debatable: reviewing what he proposes takes almost the same amount of time |
Almost all the practical value a national team receives today is in the first column. And almost all the headlines focus on the second.
Where it really works
In short: Time-consuming tasks that don't contribute to judgment: documenting interviews, writing reports, taking minutes, and searching through past conversations. These are writing tasks, not judgment tasks.
Document interviews
This is the case with the clearest return on investment. A consultant spends between 30 and 45 minutes writing the report for each interview, and they do so afterward, with partial notes and only partially fresh memories. With the conversation transcribed and structured, that time drops to about five minutes of review. In a process with six candidates, that's between two and four hours recovered.
And there's a side effect that often outweighs the savings: You stop taking notes during the interviewWhile you're writing, you're not listening.
Standardize the team's output
In a consulting firm with several consultants, each one writes in their own style. The client receives reports in different formats depending on who handled the process. Defining a template and ensuring all reports use it solves a perceived quality problem that can't be fixed with more internal training.
Recover what was said
Being able to search within conversations—what a candidate said about their availability, what the client requested in the kickoff meeting—makes the historical record accessible. Without transcripts, that knowledge lives only in the mind of the person who was there, and it goes with them when they change companies or go on vacation.
Where it fails, and it's good to know beforehand
In short: AI fails in three areas: poor audio quality, nuances not explicitly stated, and any assessment that depends on the customer's context. No tool can fix this; what can be done is to design the process around these limitations.
| Failure in | What's happening | To do |
| Bad audio | An echoing room or a distant microphone degrades the transcription more than any accent. | Record close to the speaker; in person, use a dedicated recorder on the table |
| The unspoken | The hesitation in answering, the body language, the feeling of fitting in: none of that is in the audio | Add your observations as a note before generating the report |
| Customer context | The model doesn't know that this client rejects very junior profiles even if they are technically qualified. | Let it be in the template or the instructions, not in your head. |
| Invention | When asked to assess competencies, a model may fill in the gaps with what sounds reasonable. | Require that each assessment quote a verbatim phrase from the interview |
What does the AI Act require, and since when?
In short: from the August 2, 2026 The obligations are in force transparency of article 50. The specific obligations of the systems of high risk Annex III —which includes employment and selection— were postponed to December 2, 2027 by the Digital Omnibus approved in June 2026.
This is the point where there is the most misinformation right now, because the calendar changed recently and many articles are still citing the old dates.
| That | Since when | Who is affected? |
| Prohibited practices | In force | Recognition of emotions in the workplace, among others |
| Transparency (art. 50) | August 2, 2026 | Notify users that they are interacting with AI and label generated content |
| High risk, Annex III | December 2, 2027 | Employment and selection systems: human supervision, risk management, data quality |
| High risk, Annex I | August 2, 2028 | AI integrated into already regulated products |
Sources: analysis of the AI Act timetable following the Digital Omnibus in LIS Data Solutions and SapientiaDEV, consulted in August 2026. This guide is not legal advice; for a specific case, consult a professional.
Recording interviews without violating the GDPR
In short: You can record a selection interview, but you have to inform the candidate that it is being recorded and for what purpose, to have a legal basis for the processing and not to keep the recording longer than necessary.
- Inform before you beginNot in the middle. A clear sentence at the beginning is enough.
- Explain why: prepare the interview report, do not automatically evaluate the candidate.
- Define how much is kept and delete it when the process is finished.
- Check where it's processedIf the provider processes the data outside the EU, you need additional guarantees.
- Put it in writing in your candidate privacy policy.
You have it developed in the GDPR guide for recording interviews and meetings.
Where to start: a 30-day plan
In short: start with a single task And the least risky option: documenting interviews. If you try to transform the entire process at once, the team will revert to its old way of working within two weeks.
Before making any changes, note how much time each consultant spends documenting each process. Without that data, you won't be able to tell if anything has improved, and you'll just be arguing based on feelings.
Take a report you already like and turn it into a template: what sections it has, what each section covers, and what tone to use. This step determines whether the result is suitable or needs to be rewritten.
Not in a demo. In two real processes, with one consultant who is enthusiastic and another who is skeptical: the skeptic will give you useful feedback.
Compare the documentation time to that of the first week. If it hasn't decreased, the problem lies with the template or the tool, not the team.
How to tell if it's working
In short: Four indicators are enough: documentation time per process, coverage time, 12-month stay and consistency among consultantsIf after three months none of them have moved, the tool is not solving a real problem.
| Indicator | What does it tell you? | Alarm signal |
| Documentation time | Direct savings are the easiest to measure. | It's not going down: the template doesn't fit and everything needs to be rewritten |
| Coverage time | If the savings translate into closing earlier | The first one goes down but it doesn't move: the bottleneck is elsewhere |
| 12-month stay | If the quality of the decision is maintained | It's getting worse: hiring is happening faster and worse. |
| Consistency | If all reports are at the same level | It still depends on who writes it |
The details of each metric and how to calculate it are in the Recruitment and Selection Report Guide.
What type of tool do you need?
In short: It depends on where your conversations take place. If they're video calls, a browser extension; if they're in-person or by phone, a recorder; if you already have the files, you just need to be able to upload them.
| Your interviews are… | What you need |
| Video calls (Meet, Teams, Zoom) | An extension that records without putting a bot in the room |
| Telephone companies | A recorder that captures the call; the browser cannot do it |
| In-person | Record from the computer or with a recorder on the table |
| Already recorded | Being able to upload the file and receive a transcript and report |
If it's the latter, you can try it right now with the tool of transcribe audio to text or with the online voice recorder.
Five mistakes that are repeated
- Start with the decision, not the paperwork. It is where there is the most risk and where the least time is saved.
- Not defining the template. Without it, the report comes out generic and has to be rewritten, so the savings disappear.
- Measuring "productivity" in the abstract. If you don't measure the documentation time before you start, you won't know if it worked.
- Record without warning. Besides being illegal, it destroys the candidate's credibility if they find out later.
- Buy the demo. The demos use clean audio and ideal scenarios. Try it with an interview of your own, in your own room, and with your own microphone.
Actionable summary
- Use AI to documentnot to decide: that's where the real savings and low risk are.
- Define your template Before testing anything. That's what separates a useful report from one that needs to be rewritten.
- Measure the documentation time Before and after. Without that number, everything is just opinion.
Frequently Asked Questions
What is artificial intelligence applied to human resources?
It's the use of AI models for tasks within the human resources lifecycle: filtering candidates, transcribing and summarizing interviews, writing reports, analyzing employee data, or answering frequently asked employee questions. It's not a system that makes decisions for you; it's a layer that reduces manual work and leaves the decision in the hands of the professional.
Can AI decide who gets hired?
It shouldn't, and in Europe, it also has legal consequences. A system that automatically rejects candidates falls under Annex III of the AI Act as a high-risk employment system, subject to obligations regarding human oversight, risk management, and data quality. AI proposes; the person decides and responds.
Since when has compliance with the AI Act been mandatory in personnel selection?
The transparency obligations of Article 50 have been in force since 2 August 2026. The specific obligations for high-risk systems in Annex III, which include employment and selection, were postponed to 2 December 2027 by the Digital Omnibus adopted in June 2026. The postponement provides leeway, but does not exempt: it is advisable to document now how AI is used.
Does AI introduce bias into selection?
These biases can be amplified if AI is trained on historically biased decisions. Therefore, the criterion isn't "to use AI or not," but rather at what stage of the process. Using it to transcribe and structure information is low-risk; using it to score or reject people is where problems arise and where the regulations become strict.
Which HR tasks should be automated first?
Administrative and repetitive tasks that do not involve judgment about people: documenting interviews, writing reports, preparing minutes, and summarizing meetings. These are the most time-consuming, add the least value, and have the lowest regulatory risk.
Do you need to know about technology to use AI in human resources?
No. Today's useful tools work like any other application: you upload a recording or activate an extension and receive the result. What is needed, however, is the discernment to review the output, because the responsibility still lies with whoever signs the report.
How much time is actually saved?
It depends on the task. In interview documentation, the typical time saved is going from 30-45 minutes per interview to about 5 minutes of review. In a process with six candidates, that's between two and four hours per process. It's a real and measurable saving, unlike generic promises of "multiplying productivity."
Is it possible to record an interview for AI to process?
Yes, always informing the candidate that the conversation is being recorded and for what purpose. Consent and the legal basis for processing are mandatory under the GDPR. It is also advisable to indicate how long the recording will be kept and allow for its deletion.
What KPIs measure whether AI is contributing anything?
Time spent on documentation per process, coverage time, 12-month retention period, and consistency of reports across consultants. If none of these changes occur after three months, the tool is not solving a real problem.
Will AI replace recruiters?
Not in the short term, and probably not in the medium term either. What it replaces is the administrative part of the job: writing, organizing, and searching. Evaluating a person, negotiating, and building relationships with candidates remain human, and these are precisely the areas where the recruiter adds value.
Transparency: Voicit is an AI-powered transcription and reporting tool. This guide is written to be helpful even if you don't use it; the product is mentioned only where relevant.
CEO and co-founder of Voicit. He works with recruitment consultancies and HR teams that document interviews and meetings with AI.
Looking for AI recruiting tools?
We compared 13 AI recruiting tools by hiring stage: sourcing, screening, scheduling, interview and report, and ATS, with what each one automates and real pricing.