An AI agent screens job applications in seconds. A built-in guardian checks every decision, blocks unfair criteria, and hands every judgment call to a person. If outcomes ever turn unfair, it switches the AI off by itself.
applications a year at our example company. Recruiters spend most of their time just reading CVs.
Women were shortlisted at only 63% of men's rate by a real-world-style vendor AI we tested. Nobody noticed.
applicants were hidden from recruiters by that same tool and auto-rejected after 14 days unless someone happened to open them.
The AI does the repetitive work. The guardian watches the AI. People stay in charge.
An approval agent reviews the AI against company policy and the law. A human committee signs off and decides how much the AI may do on its own.
Names, age, gender, address and college are hidden before scoring. The guardian checks each decision before it reaches a candidate, and blocks unfair job criteria like "top colleges only".
The guardian keeps measuring outcomes for women, older applicants, regions and disability. If a group is treated unfairly, it switches the AI to assist-only mode instantly.
A bank is hiring four relationship managers. Press play to run the full system on the server, live. All applicants are synthetic.
applications in
We fed in 1,200 biased past decisions to test the safety switch.
| Applicant | Outcome | Why |
|---|
Pick a proposal. The agent reads it, runs the company's checks for risk, fairness and the law, and writes a recommendation for the approval committee in a few minutes. The rules decide the verdict, and people make the final call.
63 automated checks must pass before any part of the system can be released, including 9 deliberate attempts to break the guardian.