A New Tool, Assessed

In practice

The vendor pitch covers two features, so the team assesses them separately. First, the type of AI. Both features run on a foundation model licensed from another company, embedded in the vendor's recruiting software. Second, the risk scan. Drafting job descriptions carries confabulation and some bias risk, with moderate consequences, because a person reviews every posting. Screening applications is different: it touches data privacy, harmful bias and human–AI configuration, and an error means a qualified person is turned away without anyone noticing. The consequence is high. Third, the data terms. The vendor keeps application data for a year and uses customer inputs to improve its model unless the customer opts out; the team requires the opt-out and a shorter retention period in the contract. Fourth, the evidence. The vendor supplies an independent audit report, but nothing on how screening performs for different groups of applicants, so the team asks for it and plans its own test on past applications. Fifth, the roles. The vendor develops the tool, the company's integrator connects it to the applicant system, and the recruiting team operates it, so the contract names who monitors performance and who handles complaints. The decision: drafting is approved with the standard review rule. Screening is held until the bias evidence and the team's own test are in, and even then the tool may only suggest; recruiters make every screening decision.