What does CV screening cost you today?
The cost is rarely the hours. It is the tail of the pile.
A vacancy closes, the applications land in a burst, and the first pass is the job that gets squeezed by everything else on the desk that week. The first forty get read properly, the next forty get skimmed, and the last forty get judged on a job title. That is what happens when volume arrives faster than attention, and it is worst on the roles that attract the most applicants. The price is the candidate you never called.
The hours are real too. Totaljobs surveyed 748 HR leaders and found recruiters spend an average of 17.7 hours of administrative work per vacancy, 3.6 of them reviewing applications. That research is British, published on 19 August 2025, and we found no Irish equivalent, so treat it as an order of magnitude. The number worth having is your own. On your last five roles, how long was it from advert close to shortlist, and how many applications did somebody read to the end?
What exactly gets automated?
Every CV and cover letter is read against the brief for that specific role, so the hundredth application gets the same attention as the first. Each is turned into facts: years in a named discipline, the technologies actually used rather than listed, notice period, location, work authorisation, gaps and moves. You cannot rank consistently against a document that is free text on one page and a table on the next.
Then your knockout criteria, written in plain words before the role goes live. Must hold a current right to work. Must have three years hands-on with the named system. They are applied the same way every time.
Every candidate comes back with a position, a score against each criterion and the line from the CV that justifies it, so you can see why a name sits at number three, and so can the client. The near-misses, who fail one criterion and clear the rest, come back as their own list. They are the first people a manual sift loses.
What still needs a person?
Every decision. That is not a hedge, it is the design.
The system ranks and explains. You decide who gets called, who gets rejected, and who gets pushed at the client despite the ranking. We do not build automatic rejection, and nobody is filtered out of your view: a candidate who fails a knockout is flagged as failing it and stays on the list, because the most common way these systems go wrong is by quietly deleting people.
The brief is yours too. You write the criteria and the weighting at the start. A system that infers its own criteria from your past hiring will reproduce your past hiring, including the parts you would not defend.
The law points the same way. Annex III of the EU AI Act classes as high risk any AI system "intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". You cannot escape that by calling the tool a preparatory step, because Article 6 says a system in Annex III "shall always be considered to be high-risk where the AI system performs profiling of natural persons". As the deployer you have to use it in line with the provider's instructions and assign human oversight to people "who have the necessary competence, training and authority". Those obligations currently apply from 2 December 2027. Article 4, on AI literacy for the staff who operate these systems, has applied since 2 February 2025. Already in force, Article 22 of the GDPR gives a person the right not to be subject to a decision based solely on automated processing where it significantly affects them, with safeguards including human intervention and the right to contest, and the Employment Equality Acts cover selection across nine protected grounds.
How is the result measured?
Against a baseline you agree before anything is built. Four numbers.
Days from advert close to shortlist delivered. The proportion of applications read in full rather than skimmed, which is close to 100% after and worth measuring honestly before. The proportion of shortlisted candidates the client takes to interview, which is the real quality signal because it is the client's judgement rather than ours. And placements originating from candidates who ranked outside the first twenty, which tells you whether the tail of the pile was ever worth reading.
Take those four for the three months before we start. If they do not move, the build did not work, and we would rather find that out on your desk than argue about it.
What could go wrong?
It learns your history and calls it a standard. Train a ranker on who you hired before and it ranks for who you hired before. This is why the criteria are written by you rather than inferred.
A criterion that is not about a protected ground behaves like one. An employment gap, a graduation year, an address, the name of a school. None is one of the nine grounds and each can stand in for one. The defence is that every criterion is explicit and reviewable, so a proxy can be found and removed rather than sitting inside a score.
Ranking gets mistaken for deciding. The moment a shortlist is treated as an output rather than an input, you have a solely automated decision, and both the AI Act oversight duty and Article 22 point at it. Oversight has to be real: the person doing it needs the time and the authority to overturn the ranking.
No record of why somebody was rejected. If a candidate, a client or the WRC asks, the answer cannot be that the system scored them low. Keep the reasoning with the decision. A sift done in somebody's head leaves no record at all.
How long does it take to put in?
Two to four weeks for a single desk.
A session with you to write the brief structure and the knockout criteria in your words. A build tested against roles you have already filled, so you can see whether it would have surfaced the person you actually hired. That is the only honest test available before it goes live. A parallel run on two or three real vacancies, where you sift as normal and compare. Then handover, with the documentation of what it does and who oversees it, which is the record the AI Act expects you to hold anyway.
Then we leave. No retainer, you own what was built, and you can change the criteria without calling us. If you want to work out whether this is the right thing to automate first, that conversation is the opportunity review, and it starts with your last five vacancies rather than with software.