Can AI do candidate screening calls

I get asked this by agency owners who have seen the demo. A voice agent rings the candidate, runs through notice period, salary, right to work, a few competency questions, scores the lot and marks them qualified or not, and the recruiter's afternoon of phone screening disappears. It is a good demo. I have built most of the pieces it is made of, so I want to explain what is inside it, what happens when you point it at real candidates in 2026, and what we chose to build instead.

What an AI screening call is made of

Strip it down and an automated screening call is a loop of three models. Speech-to-text turns what the candidate said into words. A language model decides what to say next from a script and the transcript so far, and text-to-speech says it. Around that loop sits a scoring step that reads the finished transcript against the questions and produces a number per answer. None of this is exotic now. The speech models are good, the latency is under a second, and a competent team can have one on the phone in a few weeks.

So the honest answer to the literal question is yes. AI can make a screening call, and make it well, in the sense that it will complete the script, capture the answers and produce a score, with a voice most candidates will find acceptable. The question that matters is what the call finds out, and that is where it falls apart.

Three things happen when you switch one on

The first is that the candidate's AI picks up, sometimes literally and more often in the sense that anyone applying for jobs in 2026 has an assistant open, and a scripted question from a bot gets a scripted answer read off a screen. The call becomes two language models exchanging text through a phone line, and the transcript is as informative as the CV was, which is to say polished and identical to every other one. The thing you built the call to recover, some signal about the person, is exactly what the format removes.

The second is that the scorer can only score what it can check. Notice period, salary expectation, right to work, years with a technology: these have answers a model can grade. Whether the person is any good, whether they will show up, whether the client will like them: these do not, and the scorer either ignores them or, worse, invents a number for them from the tone of the transcript. So the pile that comes out marked qualified is the pile of people who gave gradeable answers, which is nearly everyone, and the finding out has not happened.

The third is where it goes. That pile of qualified candidates goes to the client, because the tool says they are qualified and the consultant, who did not speak to them, has no basis to disagree. The hiring manager, who sees two people a week, now does the real screening in the interview room, and after a few rounds of that gives you fewer slots. The call was automated, and the work it was supposed to do moved one step to the right, to the most expensive person in the chain.

What a screening call is actually for

This is the part the demo misunderstands. On a good recruiter's screening call the consultant is listening for how the person talks about their last job, whether they hesitate on why they are leaving, what they ask about the client, whether they sound like someone this particular hiring manager will warm to in the first five minutes. Half the questions are not on any list, because they come from the previous answer. The call is adaptive, and the signal is in the adaptation.

A scripted agent cannot do that, and a candidate with an assistant open makes sure it never will. So the call, which was already the only place left where a recruiter learned anything true about a person, would be handed to a machine that learns nothing from it. We looked hard at building this. The pieces were on the shelf. We decided the product would make our customers worse at the one thing their clients pay them for.

What we built instead

The screening call in Recruitly is the recruiter's. What we put the AI on is everything around it, because the around-it is where the hours were and the judgement was not.

Before the call, the pre-call summary reads the candidate's record, the job, and every previous conversation and note, and gives the consultant a short brief so they walk into the call knowing what was said last time and what still needs finding out. That is a few minutes of reading done in a second, and it is the difference between a call that repeats last week and one that moves the person forward.

After the call, the recording is transcribed and analysed: a summary, the sentiment, the key moments, the action items and the people and companies mentioned, all written to the record without the consultant typing a word. Whether the call came through our own phone system or through Aircall or Quo, the summary lands in the same place, and the next person who opens the record sees what was said. The scorecard for the role is drafted from that and the CV, and the consultant corrects it.

The effect on a desk is that the screening call gets shorter and better, because the consultant stops asking what they already know and stops writing it up afterwards, and the number of real conversations a person can have in a day goes up. That is the automation that pays, and the one that replaces the conversation pays the vendor.

Where an automated check is fine

None of this means the facts have to be gathered by a human. Notice period, right to work, salary range, availability: these are form fields, and they should be collected as form fields, through the application form, the candidate portal or a quick reply, with the answers landing on the record. A candidate answering a form is not being screened, and nobody pretends they are. The mistake is dressing a form up as a conversation, scoring it, and calling the result qualified.

How to decide for your desk

1. Separate the facts from the judgement. Facts go on forms and land on the record. Judgement is a call a person makes, and the AI's job is to make that call shorter.
2. Put the summary on every call before you consider replacing any. If your consultants' calls are not being transcribed and written up automatically today, that is the first thing to switch on, and on most desks it returns more hours than an automated screener ever would.
3. Read the transcripts a voice agent produces before you trust its scores. Ask how many answers were read off a screen. On the desks that have tried it, the answer changes the decision.
4. Measure the client's yes rate on the people you send, before and after any screening automation. That number is the only one that tells you whether the screening found anything out.

AI can do the call, and the screening was never the questions, and the candidates have AI too. We built the parts that make the recruiter's call faster on both sides and left the call itself with the recruiter, and I would make that decision again with everything that has shipped since.

Written for agency owners who have seen the voice-agent demo and want to know what is inside it.