AI Recruiting ROI: How to Build a Frontline Business Case on Your Own Numbers

Why are vendor figures a weak basis for AI recruiting ROI?

They describe someone else's funnel and rarely show their method. Most ROI figures you will find for AI in recruiting come from suppliers, without the sample, the roles or the starting point, so a finance director cannot check them.

Independent research is better, but it does not transfer either. A working paper by economists Brian Jabarian and Luca Henkel, revised in September 2026, reports a randomised field experiment with 67,056 applicants for entry level customer service jobs in the Philippines (Jabarian and Henkel, 2026). Applicants interviewed by an AI voice agent received a job offer 9.73% of the time, against 8.70% with a human recruiter: a relative rise of 12%. The difference carried through to job starts and to people still employed a month after starting. Human recruiters made the hiring decisions in both groups.

That study tells you which numbers to watch: offers, starts and people still working a month later. It is not an input: it covers one recruitment firm, one job type and one labour market, and an order picking role in Venlo may behave differently.

The same rule applies to us. Radius Hire is new, so this article explains the method and the product and claims no results.

What belongs in an AI recruiting business case?

Four groups of lines: gains, one time costs, running costs and risks. Most business cases for recruitment software list the gains in detail and the costs as a single licence fee. A case that survives a finance review needs every line, with a note on where each figure comes from.

Our model for the cost of a frontline hire shows how to build hourly labour cost and the cost of filling a seat from your own payroll, with official Dutch statistics as a reference. Both feed the recruiter hour and failed start lines.

How do you turn recruiter hours into euros?

Count only the hours that move to work you can name. The usual argument multiplies the screening hours the tool takes over by salary and calls the result a saving. It has two gaps.

First, the new process adds work. Someone still has to read every interview summary and check the evidence before a candidate moves on, because a person makes the decision. A review time close to zero usually means people have stopped reading.

Second, released hours only become money if something changes. If you plan to run with fewer recruiters, count only the payroll you will stop paying. If you keep the same team, the hours are worth what the team does with them instead, such as calling every new starter the day before their first shift or getting offers out the same day. If you cannot name the work, count the hours at zero.

Price the hours at the loaded cost: gross pay plus holiday allowance and employer contributions, divided by the hours people actually work rather than their contract hours.

What does a worked example look like?

A worked example shows which lines carry the case. Illustrative example: one site, one role family, one year. Every figure below is an assumption chosen to show the method, not a benchmark, a market average or a Radius result. Replace each one with your own.

Line B draws on the field study above, where 78% of applicants offered the choice picked the AI interviewer (Jabarian and Henkel, 2026). This example assumes about one in five candidates will want another route; they are screened as today and release no hours.

What is the most the plan and its running oversight could cost per year before the case loses money? That is the break even budget: K plus failed starts avoided × N.

Three things stand out. The recruiter hours on their own barely cover the internal cost of getting started. The case stands or falls on failed starts: each one avoided is worth more than 22 hours of recruiter time (€900 ÷ €40). And lines I and J fall away in year two, so show a three year view next to year one.

So the pilot's main job is to find out whether failed starts actually fall. If slow starts cost you overtime, add a line for each uncovered shift; agencies add placement lines, covered below.

Explore: see how the AI Interviewer collects answers and evidence for a recruiter to review.

Which costs and risks do business cases usually leave out?

Review time, legal preparation and candidates who will not use the AI step. These lines arrive after signing and are easy to miss.

Legal and governance work

Data protection. Using AI to evaluate candidates will in most cases require a data protection impact assessment before it starts. The GDPR requires one for a systematic and extensive evaluation of people based on automated processing, including profiling, where decisions based on it have legal or similarly significant effects on them (GDPR, Article 35), and the Dutch data protection authority's list of processing that always needs one includes profiling (Autoriteit Persoonsgegevens, 2019).

The AI Act. AI systems intended to analyse and filter job applications or evaluate candidates are listed as high risk in Annex III of the EU AI Act (Regulation (EU) 2024/1689). Their obligations apply from 2 December 2027, after the Digital Omnibus on AI (Regulation (EU) 2026/1744) moved the date from 2 August 2026, as the European Commission confirms. From then, an organisation using such a tool must, among other things, follow the provider's instructions, give oversight to trained people with the authority to act, monitor it, keep the logs under its control for at least six months and tell candidates that a high risk AI system is being used. That is recurring work, so it belongs in running costs (more on the EU AI Act and recruitment).

AI literacy. Since 2 February 2025 the AI Act has required organisations that use AI systems to take measures on AI literacy for the staff who operate them. Since the Digital Omnibus took effect on 27 July 2026, the duty is to take measures that support that literacy, without requiring a set level for each person (Regulation (EU) 2026/1744). Training therefore sits in the one time costs, with a smaller running cost for new staff.

The works council. In the Netherlands, an employer with a works council needs its consent before introducing systems suitable for monitoring staff behaviour or performance, and for rules on recruitment policy (Wet op de ondernemingsraden, article 27). Whether a specific AI recruitment tool needs consent depends on how it is used, so involve the works council early and get advice.

Emotion reading. Since 2 February 2025 the AI Act has banned AI systems that infer people's emotions from biometric data, such as facial expressions or voice, in the workplace, except for medical or safety reasons. The European Commission's guidelines, which are not binding, say this ban also covers candidates during recruitment (European Commission, 2025). Treat any tool that claims to read confidence or emotion from a voice as a legal risk.

Candidates who take another route

Not every candidate will take an AI interview, and some should not be asked to. Some prefer a person, some have a poor connection, a hearing or speech impairment, or little trust in automated systems. Keep a second route, such as a phone call or a conversation on site, and count its cost. Track how many people use it and whether they reach an offer at the same rate as everyone else. If they do not, you have a fairness problem to fix before you have a saving to count.

How do you test the case before you sign a long contract?

Run a time boxed pilot against a comparison group with agreed stop rules. A pilot is where your assumptions become your figures.

Two cautions. A before and after comparison at one site cannot separate the tool from the season. And small numbers mislead: with 50 starts in a group, two or three failed starts either way can be chance, so extend the pilot if the result sits close to your threshold.

How is the business case different for a staffing agency?

Agencies earn on placements, so count placements and margin as well as costs. For an employer, most gains are costs avoided: fewer failed starts, less overtime, fewer empty shifts. For a staffing agency the main gain is usually revenue. When a client sends the same order to more than one agency, speed to a suitable, available candidate can decide who places. So add these lines:

Do not count your client's savings as your own gain; they belong in the sales conversation. And run the pilot per client or per branch, because role mix, pay and travel distance can outweigh the effect of any tool.

What should the one page summary for finance say?

The decision, the break even budget, the evidence and the exit.

How Radius helps

Radius Hire is the AI hiring platform for frontline roles. Its AI Interviewer, Match and Radius Passport sit on one candidate record, and a recruiter makes every decision.

Radius is priced as an annual plan sized to your hiring volume and quoted after a demo, so you can set the quote against your own break even budget. There is no free trial. Radius is built for staffing agencies that recruit for several clients and sites (for agencies).

What makes an AI recruiting business case credible?

Your own numbers, visible assumptions and a test that could prove it wrong. A credible case starts from your applicants, your starts and your failed starts, not from a percentage in a brochure. It works out the most the tool can cost, then runs a pilot that is allowed to say no. If the numbers hold, the decision is easy to defend. If they do not, you found out on a pilot budget instead of a multi year contract.

General information, not legal advice. Position as of 22 October 2026.