An AI Agent in HR: CV Screening and Onboarding

Three hundred CVs arrive for one advert. A recruiter works through them in two days, rejecting nineteen out of twenty within five seconds. Candidates wait meanwhile, and some accept another offer before anyone replies.

It is understandable that automation pushes in exactly here. It is also the area where we recommend the most caution — not for technical reasons, but because decisions about people are being made.

The difference between those two words is the whole difference between a useful deployment and a problematic one.

Sorting means the agent establishes verifiable facts: does the candidate hold a driving licence, do they state the required language, how many years of experience do they claim in the field, is the CV complete. It extracts these from the document into a structured form.

Scoring means the agent assigns candidates a number and ranks them. This is what we advise against. A score looks objective, but it comes out of a model nobody sees in full — and a recruiter naturally reads top-down and never opens the bottom of the list.

A useful way to phrase the line: the agent may say what is in the CV. It may not say who is better.

If a model learns from who the company has hired so far, it learns the patterns that were there — including ones nobody would introduce deliberately. The classic example is a preference for unbroken employment history, which systematically disadvantages people returning from parental leave or illness.

Practical measures that work:

• Do not give the agent data it has no business judging — name, age, gender, photograph, address.
• Check the output regularly — compare the make-up of those advancing with the make-up of those who applied.
• Keep the full list available, not only the part the agent flagged.

Recruitment is personal-data processing, and under EU rules it counts among the higher-risk uses of AI. In practice that means settling: the legal basis for processing, informing candidates that an automated tool is used in the process, retention periods for CVs, and the option of human review.

We cover this in more depth in the piece on AI agents and GDPR. The principle is simple: anything that affects a person has to be explainable and reviewable.

A few uses where nobody is being assessed and the saving is immediate:

• Adding structure — turning CVs in many formats into one consistent overview so they can be compared.
• Replying to candidates — acknowledging applications, status updates, answers to repeated questions about the role.
• Scheduling interviews — finding a slot across calendars, which is a surprisingly large share of the work.
• Preparing the interview brief — a summary of the CV and suggested questions about what is missing or unclear in it.

That last one has a side effect companies value: interviews become more consistent, because every candidate is asked about the same areas.

This gets overlooked. A new starter's first weeks are a process with dozens of steps that repeat identically — accounts and access, equipment, training, documents, introductions, check-ins after a week and after a month.

Nobody is being assessed in it, so none of the constraints above apply. It is also the process most often done by hand and inconsistently — and a new joiner forms their first impression of the company precisely here.

Here an agent can track what is done and what is late, remind the people responsible, answer a newcomer's routine questions from internal documents, and flag when a step falls through.

Not by the number of CVs processed. Track time from application to first reply (the strongest influence on whether a candidate stays in play), time to fill, the share of candidates who dropped out mid-process, and for onboarding the share of steps completed on time.

We take the same approach to measurement in the piece on the ROI of deploying an AI agent.

Start with onboarding, not recruitment. It carries less risk, the benefit is visible within a month, and the team gets used to working with an agent on a process where it cannot do harm. Move to recruitment afterwards — and even there, start with sorting rather than scoring.

If you would like to go through what can and cannot sensibly be automated in your process, get in touch. More on our approach to AI and automation, and examples in our case studies.