CV screening with AI: how to do it carefully

On a vacancy with lots of applicants, the first hour goes on reading. Opening CVs, looking for experience, checking availability, ruling out half. That work can be well supported by AI. But it's about people, which brings rules and responsibilities that don't apply to an invoice.

What does CV screening cost you today?

Work it out for one vacancy. How many applications do you get, how many minutes does a first review take, and how many candidates do you end up speaking to? Add the lead time: how many days are there between an application and first contact?

That lead time often matters more than the hours. In a tight market, the candidate you call after four days is already talking to someone else. Responding faster gets you more than reading faster. For staffing and recruitment agencies, this is worked out further on the page about recruitment and staffing agencies.

What can and can't AI do in a hiring process?

The distinction you need to make is between supporting and deciding. AI can summarise, structure, sort and flag. Rejecting is a decision about a person, and that belongs with a person.

Two frameworks surround this. The GDPR sets requirements for decisions based solely on automated processing that significantly affect someone. The EU AI Act classifies AI systems used for recruitment and selection as high-risk, with requirements including human oversight, logging and informing the people involved.

That doesn't mean you can't use it. It means you need to be able to show how you use it. Have your set-up reviewed by your data protection officer or a lawyer; this article isn't legal advice. What else is involved is covered in our article on the AI Act and the GDPR.

Step 1: write down what you're selecting on first

The biggest risk isn't the model, it's criteria that have never been written down. As long as selection is based on gut feeling, you can't explain why someone is ruled out, and you can't check whether it's fair either.

For each vacancy, put on paper what's really needed: which tasks someone must be able to do, which certification is required, what availability, what travel distance. Separate hard requirements from nice-to-haves. Anything without a demonstrable link to the work has no place in the selection.

Step 2: let AI summarise and sort, not reject

Use the model to produce a summary in the same format for each candidate: relevant experience, education, availability, and the points where information is missing. Have it indicate for each hard requirement whether there's evidence for it, and where in the CV it appears.

Don't give an overall score that takes over the decision. A score of 72 invites people to stop reading. An overview per requirement, with the source alongside, forces people to look. The recruiter decides, and sees on one screen what the decision is based on.

Don't let candidates who seem to fall short on a requirement disappear automatically; put them on a separate list that someone still goes through. That's where you find the people with an unusual CV, and they're often not the weakest. How such a role is set up is under CV Assistant.

Step 3: connect it to your ATS

Screening that sits next to your system gives you an extra screen and extra retyping. The gain only comes when the summary is in the candidate record itself, with the application it belongs to.

That means integrating: fetching the application and the CV, writing the summary back to the candidate, and updating the status when the recruiter makes a choice. If your ATS has an API, that's a manageable job. What we can integrate for each system is under integrations, and indicative prices are under what an integration costs.

Step 4: record what happened

For each candidate, keep which version of the instructions was used, which summary came out and which team member made the decision. You need that when a candidate asks why they were rejected, and when you have to show there was human oversight.

Also tell candidates that you use AI for the groundwork, and what for. That's more transparent and it avoids trouble later. Stick to your retention periods: don't keep CVs of rejected candidates longer than necessary, and ask for consent if you want to keep someone on file for future roles.

How do you avoid reinforcing bias?

A model learns patterns from language and from examples. If you feed it your own hiring history, it learns who you've hired so far. That's not a good measure of who you need.

• Leave name, age, gender, nationality and photo out of what the model gets to see.
• Assess on demonstrable requirements, not on the school or employer someone comes from.
• Don't train on your own rejections; use written criteria instead of past behaviour.
• Use samples to measure whether outcomes are comparable for different groups, and record that check.
• Have a second person review a share of the rejections.

When is AI CV screening not a good idea?

With few applications per vacancy, there's no point; just read them yourself. For roles where selection mainly comes down to personality and motivation, a CV summary adds little. And if your selection criteria aren't written down, start there: otherwise you're automating a judgement you can't explain.

If your bottleneck is actually further along the process, for example in scheduling, timesheets or contracts, look there first. That's often more work than the screening itself, and it comes with fewer sensitivities.

Where do you start?

Start with one type of vacancy that attracts lots of applications, with written requirements and a recruiter who owns it. Run summaries alongside the process for four weeks without anything happening automatically, and compare the outcomes with your recruiters' assessments.

If that shows the groundwork is right, expand. If you first want to know which part of your recruitment process really takes time and what automating it would deliver, the AI Opportunity Audit is the logical first step. The build itself is under AI agent development.