What does it cost to build a kit for every role?
Two things, and only one of them is hours.
The hours are the part that gets counted. Totaljobs surveyed 748 HR leaders and found that recruiters spend an average of 17.7 hours per vacancy on administrative work, including 2.5 hours scheduling interviews and another three hours processing post-interview notes. That research is UK data, published on 19 August 2025, and we found no Irish equivalent, so treat it as an order of magnitude rather than a fact about your desk. The number worth having is your own: on the last five roles you filled, how long went into preparing to interview, and how long went into writing it up afterwards?
The second cost is the one nobody puts a figure on. When the questions are improvised, the second candidate is not asked what the first candidate was asked, so the comparison at the end is not a comparison. Three people interview well and the notes say three different things about three different topics. The client then asks why this one and not that one, and the honest answer is a feeling. The feeling is often right, which is exactly why it is worth being able to show the working.
The write-up carries the same problem in a different shape. It is written from memory a day later, in whatever order the memory arrives, and the parts that were never asked look identical on the page to the parts that were asked and answered badly.
What exactly gets automated?
The blank page, the format, and the sorting afterwards.
The kit itself. The job spec goes in and a draft kit comes back: four to six competencies for that specific role, questions under each one with follow-up probes, and a scoring rubric with a written descriptor for each level, so that a 3 means the same thing to you on Tuesday as it did to your colleague on Monday. Technical roles get the technical competencies named in the spec rather than a generic list with the job title dropped into it.
In your format. Your template, your rating scale, your section headings, your logo on the scorecard the client sees. An agency that has spent years building a house style should not have to adopt somebody else's because a tool exports one shape.
Reuse. Roles you fill often stop starting from nothing. The kit you approved for the last warehouse supervisor is the starting draft for the next one, adjusted for what this client asked for differently, which is usually two or three things rather than twenty.
The write-up. After the interview, your notes go in and the write-up comes back arranged against the competencies the kit used, each piece of evidence sitting under the competency it belongs to, in the format the client receives. Where a competency was not covered, it says so rather than quietly leaving a gap that reads like a pass.
That last point matters more than it sounds. The write-up structure is where most of the three hours goes, and it is the piece a recruiter is most likely to do at half past six on a Friday.
What still needs a person?
The editing, the asking, and the scoring. All three, every time.
Every kit is a draft until a recruiter has read it. Questions that are technically fine and useless on a live desk are the normal failure mode: a competency the client did not ask about, a probe that does not survive contact with someone who has done the job for fifteen years. The interviewer cuts, adds, and reorders before it goes anywhere near a candidate.
The interviewer asks the questions and the interviewer scores the candidate. Nothing scores a person in these builds.
That is a design choice and it is also the legal line. 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". Read the input side of it. Generating a question set from a job spec evaluates nobody, because there is no candidate in the input. A system that reads a person and returns a score is doing the thing Annex III names. So we build the kit generator and the write-up structurer, and the number on the scorecard is written by the interviewer.
If you do want the screening side, that is a different page and a different build with obligations attached: screening applications against the brief is high risk under the same Annex III point, and it is built and documented accordingly. We would rather keep the two apart on your desk than blur them and then argue about which one you bought.
Does a generated question set help or hurt on the nine grounds?
It helps if somebody reviews it, and it hurts at scale if nobody does. Automation makes a good question set consistent and a bad one systematic.
A recruitment agency is named directly in Irish equality law, not caught by it indirectly. Section 11 of the Employment Equality Act 1998 says that "[without] prejudice to its obligations as an employer, an employment agency shall not discriminate against any person who seeks the services of the agency to obtain employment with another person". The nine protected grounds are gender, civil status, family status, sexual orientation, age, religious belief, membership of the Traveller community, race and disability.
The evidential argument is the practical one. Section 85A of the same Act, inserted by the Equality Act 2004, puts it plainly: "Where in any proceedings facts are established by or on behalf of a complainant from which it may be presumed that there has been discrimination in relation to him or her, it is for the respondent to prove the contrary." Proving the contrary is much easier when there is a written question set that every candidate for that role was asked, and a completed scorecard with a note against each score, than when there is a diary entry and a recollection.
The questions that cause trouble are rarely the obvious ones. Nobody writes "are you planning a family" into a template. The ones to watch are the questions that sound like small talk and land on a ground anyway: what year you finished college, whether the commute suits with the kids, how you would fit with a young team. A generated bank can produce those, which is precisely why the review step is not optional.
What we do not do is certify the bank. We are not your employment lawyer and a workflow is not legal advice. Your own adviser signs off the question set once, the same way they sign off a contract template, and after that the workflow's job is to make sure the approved questions are the ones that actually get asked.
What could go wrong?
Questions that sound right and test nothing. A generated competency question can be fluent and hollow. The test is whether an experienced person could answer it well without ever having done the job.
The rubric quietly becomes the decision. Scoring stays human, and a number on a page still exerts a pull. If a scorecard total is being read as an answer rather than as a summary of what was said, the structure has stopped helping.
Kits going out unedited. The whole argument for this build assumes the edit happens. If the desk gets busy and the kit ships as generated, you have automated the drafting and lost the reason it was safe.
A write-up assembled before the interview. The structure exists to arrange notes that were actually taken. Pre-filling it, or letting it infer what a candidate probably said, turns a record into a story.
Kits and scorecards that live nowhere. If the approved kit and the completed scorecard are not filed against the vacancy, the evidential benefit is gone and only the time saving remains. That filing is the same discipline as keeping the pipeline current, and it is worth building at the same time.
How is it measured, and how long does it take to put in?
Four numbers first, taken before anything is built.
Hours per vacancy spent preparing to interview and writing up afterwards, your own version of the UK figures above. The proportion of interviews with a completed scorecard on file, read from records rather than from memory, which is usually the number that surprises people. Days from interview to write-up delivered to the client. And the client's interview-to-offer rate, because a better-structured interview should show up in their decisions, not only in yours.
Take three months of history on those 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 later.
Two to four weeks for one desk. A session to write your competency framework and scorecard template in your own words. A build tested against roles you have already filled, so you can hold the generated kit beside the one you actually used and judge it. A parallel run on two or three live vacancies. Then handover, with documentation of what runs and who owns each step.
Then we leave. No retainer, you own what was built, and you can change a competency without calling us. This is the cheapest of the recruitment builds to put in and the easiest to see working, which is why it is often the one to do first, before the database you already paid for or the paperwork after a placement. If you want to work out which of them is right for your desk, that conversation is the opportunity review, and it starts with your last five vacancies rather than with software. Who you would be working with is the two founders, and the recruitment side is led by someone who ran an IT and technical desk for eight years.
Questions we get asked
Can AI write interview questions for a specific role? It can draft them. Give the workflow the job spec and it returns competencies for that role, questions and follow-up probes under each one, and a scoring rubric in your own template. What comes back is a draft. The interviewer reads it, cuts what does not fit the client, adds what the spec left out, and only then does it go near a candidate.
Is using AI to generate interview questions high risk under the EU AI Act? Not on its own, and the distinction matters. Annex III of the EU AI Act classes systems intended to analyse and filter job applications and to evaluate candidates as high risk. Generating a question set from a job spec evaluates nobody, because there is no candidate in the input. A system that scores the candidate afterwards is squarely inside Annex III, which is why we leave scoring with the interviewer.
Does a standard interview question set help with Irish equality law? It helps as evidence, and it is not a legal opinion. Under section 85A of the Employment Equality Act 1998, as amended, once a complainant establishes facts from which discrimination may be presumed, it is for the respondent to prove the contrary. A written question set every candidate was asked, and a completed scorecard with notes, is the kind of record that answers that. Have your own adviser review the bank.
Can AI write the post-interview write-up? It can structure one. The interviewer's notes go in, and the write-up comes back arranged against the competencies the kit used, with each piece of evidence under the competency it belongs to, in your client-facing format. It does not add observations nobody made and it does not award the scores. If a competency was not covered, it says so rather than filling the gap.
How long does it take to set up interview kits on a recruitment desk? Two to four weeks for one desk. A session to write your competency framework and scorecard template in your own words, a build tested against roles you have already filled so you can compare the generated kit against the one you used, a parallel run on two or three live vacancies, then handover with the documentation. There is no retainer and you own what was built.