AI sourcing is only useful if it explains itself

Ranking without reasoning is not progress

Keyword search produced poor shortlists quickly. Unexplained AI produces different shortlists just as quickly, and nobody in the process can say why a candidate sits at position three rather than position thirty.

For a regulated employer this is unusable on its own terms. Decisions must be defensible after the fact, sometimes years later, to an auditor who was not in the room and has no interest in the sophistication of the model.

There is a practical cost too. When recruiters cannot see why the tool ranked as it did, they either accept the list uncritically or ignore it entirely. Both outcomes waste the investment.

What explainability looks like in practice

Each match should show the evidence behind it: the responsibilities, the scale, the domain and the outcomes that lifted the score, alongside the specific gaps that reduced it. 'Strong on payments platform scale, no regulated-entity experience' is a sentence a hiring manager can act on.

Gaps matter as much as strengths. A shortlist that only advertises its confidence teaches the recruiter nothing, whereas a visible gap tells them whether the brief is wrong or the market is thin.

Recruiters then correct the brief instead of arguing with the tool. That is the loop that matters: the next shortlist improves, rather than the current argument being won.

Explainability and fairness are the same project

You cannot audit for adverse impact in a system whose reasoning is unavailable. Once the criteria behind each score are recorded, you can ask whether a criterion is genuinely predictive of performance or merely correlated with a background.

This is also how proxy criteria get caught. A requirement for a specific certification, or for continuous employment, can quietly exclude entire groups; visible reasoning makes that pattern reviewable rather than invisible.

What to ask a vendor

Ask to see a shortlist for a role you know well, and ask why candidate five is below candidate four. If the answer is a score with no attached evidence, you are buying a faster black box.

Ask what happens when the hiring manager disagrees. A defensible system captures the correction and applies it to the brief, producing an auditable record of the standard changing.

Finally, ask what leaves the system. An exportable decision record — brief, criteria, reasoning, outcome — is what turns a sourcing tool into something your compliance function will actually approve.