Every recruitment founder I talk to wants an AI that reads CVs. Very few have written down what that AI should be looking for. That's why most screening bots end up switched off after a month.
A screening agent is not magic. It does one job: it reads what a candidate sends you, compares it with what the role needs, and tells a recruiter what it found and why. If your recruiters can't explain how they screen today, the agent can't either. So the build starts with the process, not with the tool.
Here is how I build one, in the order I build it.
Step 1. Write the scorecard before you touch any tool
For each role type you fill often, write three short lists.
- Knock-outs. Things that end the conversation: no right to work, no driving licence when the job needs one, not available for months.
- Must-haves. Things the client will reject a CV for: a certificate, years in a specific trade, a language.
- Nice-to-haves. Things that move a candidate up the list but don't decide it.
Sit with your two best recruiters for an hour and ask them how they decide. You'll be surprised how different their answers are. That difference is the real problem the agent exposes, and it's worth fixing even if you never build the agent.
Step 2. Decide what goes in
Candidates don't send tidy CVs. They send PDFs, Word files, a photo of a certificate, a WhatsApp voice note, a LinkedIn link. Decide which inputs the agent accepts and in what form. One simple rule works well: everything becomes text first, then the agent reads the text.
The first agent I built for my own work does exactly this. It takes a messy CV, a WhatsApp intake or a call transcript and turns it into a clean, anonymised candidate profile. Before anything goes to a client, it checks that no name, phone number or email has slipped through.
Step 3. Give the agent one narrow job
The agent should do four things, and nothing else:
- Pull out the facts: experience, certificates, languages, availability, location.
- Check them against the scorecard for that role.
- Give a result (strong match, possible match or not a match) with one line of reasoning for each point.
- Flag anything it isn't sure about, so a person checks it.
The fourth point matters most. An agent that says "I couldn't confirm the safety certificate" is useful. An agent that quietly guesses is dangerous.
Step 4. Keep the decision with a person
The agent recommends. The recruiter decides. That isn't only good practice, it's where the law is heading. Under GDPR, candidates have the right not to be subject to a decision based solely on automated processing when it significantly affects them. Under the EU AI Act, AI used to screen or filter job applications counts as high-risk. Since this summer's Digital Omnibus, those high-risk obligations apply from 2 December 2027.
That sounds far away. It isn't, if you're building now. So build to that standard from day one:
- No protected characteristics in the scorecard. No age, gender, nationality or judgement based on a photo.
- Every result comes with a written reason a recruiter can read.
- Every decision is logged: what the agent suggested, and what the recruiter did.
- Candidates are told that AI helps screen their application.
Step 5. Put the result where recruiters already work
If recruiters have to open a separate tool to see the agent's output, they'll stop using it within a week. The result should land in your ATS or CRM as a note on the candidate, with the status updated. Most ATS platforms have an API, and tools like n8n or Make can connect the pieces without a developer on staff.
Step 6. Test it against your recruiters
Before you trust the agent, run it in the background for two weeks. Recruiters screen as usual, and you compare.
- Agreement rate. How often does the agent reach the same result as the recruiter?
- Missed candidates. Check a sample of the ones it rejected. Were any of them good?
- Time per CV. Before and after.
When agreement is high and missed candidates are rare, switch it on. Keep checking a sample every month, because roles, clients and candidates change.
What it's worth
In Bullhorn's 2026 GRID report, a survey of nearly 2,300 recruitment professionals, 46% said AI had cut their screening time by half or more. That's the prize. But look at what comes before it: a scorecard, a clear process, and a person who owns the decision.
The agent doesn't fix your screening. It makes a good screening process faster.
Process before automation. Write the scorecard this week, even if the agent comes later. Your recruiters will screen more consistently from day one.