HR Tech Investment: A Recruitment Diligence Checklist

1. Is demand real?

Request evidence of an approved employer mandate, the hiring owner, an agreed role brief and current status. Separate active vacancies from talent pools, speculative demand and expired assignments. A profile registration is not an employer order.

2. Does the candidate data support the claim?

Sample actual records. Trace the original CV or supplied information, source, application date, permission, updates and duplicates. Distinguish searchable records from readable resumes and genuinely reachable candidates. Do not treat a large database count as verified supply.

3. Does screening improve the hiring decision?

Use employer-agreed criteria and compare the tool with an appropriate baseline on the same roles. Examine both unsuitable recommendations and suitable people missed. Aggregate accuracy can hide failures in scarce roles, career breaks or unfamiliar career paths.

4. Where does human judgement enter?

Ask who calibrates the role, checks inferred facts, challenges weak recommendations and decides an exception. Observe a real workflow with permission. A human approval button alone does not establish a meaningful review.

5. What does automation actually cost?

Separate software revenue, recruitment fees and pass-through costs. Include sourcing, recruiter review, implementation, model usage, support, replacement work and collection effort. Inspect realised margin after delivery, not only a presentation of theoretical automation savings.

6. Do outcomes have consistent denominators?

Reconcile applications, unique candidates, assessed shortlists, interviews, offers, accepted offers, joinings and retained hires. State the cohort and observation window. An accepted offer is not a joining, and a joining does not prove later performance.

7. Does reported revenue turn into cash?

Follow a sample from agreed fee terms to delivery evidence, invoice, replacement or refund exposure, and bank receipt. Ask whether customer concentration or long collection cycles are hidden by headline bookings. Synthetic tests and typed payment references are not received revenue.

8. Can the operating model be trusted?

Inspect access controls, audit trails, deletion and correction processes, incident ownership and model-change checks. NIST's voluntary AI Risk Management Framework highlights reliability, transparency, privacy and management of harmful bias. It is a risk lens, not a certification of a recruitment product. [2]

The evidence pack to request

• A permissioned sample linking the mandate, candidate evidence, assessment and hiring outcome.
• Customer cohorts showing repeat business, delivery effort, replacement exposure and cash collection.
• Documented failure cases and the changes made after them.
• A demonstration using permitted data, with accountable people present.