Every recruitment subprocess (sourcing, screen, assess, interview, decide, offer, onboard) rated on uplift potential and legal exposure under EU AI Act high-risk classification + US state laws. The safe-and-uplifting band is sourcing, scheduling, onboarding. The danger band is screening and decisioning.
Ask ten vendors whether AI belongs in recruitment and you will get ten versions of yes. Ask an employment lawyer and you will get a more useful answer: it depends which part of the process the AI touches. Recruitment is not one function but seven, sitting at very different points on two axes: how much recruiter time the automation gives back, and how much legal exposure it creates when it goes wrong.
We build recruitment automation for companies of 50 to 5,000 people, and the map below is the one we draw on a whiteboard in the first 45 minutes with a head of talent. It rates each subprocess for uplift and for exposure, and the pattern is consistent enough to state plainly: the safe and uplifting band is sourcing, scheduling and onboarding. The danger band is screening and decisioning. Most of what vendors are selling sits in the second band, because that is where the demo looks most impressive.
Which recruitment steps are safe to automate?
The safe band shares one property: nothing in it makes or shapes a decision about a person. Sourcing finds candidates and writes them into the ATS; nobody is rejected by it. Scheduling moves calendars, panels and reminders; the interview is still a conversation between people. Onboarding runs after the person has said yes; paperwork, provisioning and the first-week plan are administration, not judgement.
These three are also where the recruiter hours actually go: sourcing spreadsheets, reschedule emails and chasing hiring managers, not weighing candidates. Automating them gives the week back without asking anyone to defend a decision they did not make.
Why are screening and decisioning the danger band?
Because they are the decision. The EU AI Act treats AI systems used to recruit, select, evaluate or make decisions about people in employment as high-risk, with obligations around risk management, documentation, human oversight and transparency that a mid-market talent team has no appetite to own. A growing list of US state laws on automated employment decisions adds candidate notice, bias auditing and record-keeping on top. The details vary by jurisdiction, so take legal advice on where you sit. The direction does not vary: the more a system decides about a person, the more you must be able to explain and defend.
Screening is the trap because it looks like administration. 'Just filter the pile down to the ones worth reading' sounds like scheduling. It is not. A filter that never shows a recruiter a candidate has rejected that candidate, whatever the vendor calls the feature, on criteria nobody in the room can fully state. If it skews against a protected group, the company that deployed it answers for it, and 'the model did it' has never helped anyone.
Decisioning is the same problem without the disguise. Automating the call to advance, reject or offer buys very little time, because that call was never where the hours went, and all of the exposure.
Screening is a rejection with a nicer name. The law treats it as one, and so should you.
What does the function-by-function map look like?
Seven subprocesses, rated for uplift (recruiter time returned) and exposure (legal risk under the EU AI Act and US state laws on automated employment decisions). The ratings are ours, from building these systems. We are engineers, not lawyers, and the legal reading of your own process is for your counsel.
- Sourcing. Uplift high, exposure low. Finding and enriching candidates from channels you already pay for decides nothing about anyone; it widens the pool rather than narrowing it.
- Screening. Uplift high, exposure high. Filtering who a recruiter sees is a selection decision under the EU AI Act and US state laws on automated employment decisions, whatever the demo calls it.
- Assessment. Uplift medium, exposure high. Tests, scored video, inferred traits and rankings evaluate a person directly; the obligations attach to the evaluation, and the uplift is smaller than the deck claims.
- Interviewing. Uplift high, exposure low. Panels, time zones, reschedules, preparation and structured notes are administration; the conversation and the judgement stay in the room.
- Deciding. Uplift low, exposure high. The call to advance, reject or offer was never where the hours went, and it is precisely what the law is about.
- Offer. Uplift medium, exposure medium. Drafting the letter and paperwork is safe; letting a system set level or pay touches equal pay and discrimination law, so a person sets the terms and the system produces the documents.
- Onboarding. Uplift high, exposure low. The decision is made and accepted; provisioning, forms, introductions and the first-week plan are administration with a grateful audience.
What does 'a human decides' actually mean in practice?
It does not mean a person clicks 'approve' on a ranked list. If the system has sorted the pile and the recruiter approves the top of it in the time it takes to read a name, the system decided and the person rubber-stamped it. Regulators are not naive about this; neither should you be.
In the builds we ship, a human decision has three properties. The system never scores, ranks or rejects a candidate on its own; it prepares a structured summary of each application against the stated requirements of the role (what is evidenced, what is missing, what to ask) and presents it beside the original. The recruiter or hiring manager reads and decides, and the decision is recorded as theirs. And the boundary is documented, which steps are automated, which are human and why, so the answer exists before anyone asks. That third property costs almost nothing and is the one most often skipped: it is the difference between 'we use AI in hiring' and 'every decision about a candidate is made by a named person'.
How do you get the uplift without the exposure?
Automate the safe band completely and leave the danger band alone. Look at the map again: the safe band is where the uplift is. Sourcing, scheduling, communication, structured summaries, job description drafting, onboarding and reporting fill a recruiter's week and should never have needed a recruiter. Screening and decisioning were a small share of the hours and a large share of the risk.
Two further rules. Keep the automation on top of the ATS you already run rather than in a new tool that wants its own copy of your candidate data; one system of record is easier to document. And put a human review on any message that carries weight, an offer or a rejection after a final round, because the candidate reads it as a decision.
If a vendor's pitch depends on the model reducing the pile for you, ask who signs the bias audit and who answers the candidate asking why they were never seen. If the answer is 'you', the exposure has been priced into your side of the contract.
What to do next
Draw the map for your own process: seven rows, two ratings each, and a note on who decides today. Most teams find the safe band is where the recruiter week already goes, and the danger band is the part they were about to buy a tool for.
That is the work we do. J Labs automates the safe band on top of the ATS you already run (sourcing, scheduling, candidate communication and structured summaries) with the human decision preserved and documented at every stage. It starts with a 20 or 45-minute discovery call and, where it makes sense, a four-day Spec from €5,000 that writes the brief, the acceptance criteria and the compliance boundary before any build is priced. The average build ships in eight weeks. Take legal advice on where your process sits; we will build the part that gives your recruiters the week back.