From Annual Audits to Continuous Compliance

The problem

AI hiring systems are high-risk under the EU AI Act. They also sit under NYC Local Law 144, CCPA and CPRA, Illinois BIPA, EEOC Title VII guidance, and WCAG 2.2. That is six frameworks for one product category, and the industry's answer is an annual bias audit.

The research on those audits is not kind. A 2025 review of all 116 public Local Law 144 audits found 83 percent were missing demographic data. A platform that ships code three times a week can introduce a privacy violation on Tuesday that no one sees until next year's audit.

What we built

A compliance pipeline that runs inside CI/CD instead of once a year. Fifteen AI agents in two teams. A nine-agent code review team, plus a supervisor, gates every promotion from development to test to production. A six-agent red team tests the live environment on a schedule. The privacy agent and the accessibility agent hold veto power: if either one says no, the deployment does not happen. Everything runs in GitHub Actions.

1. Gate, don't advise.The pipeline can block a release. An audit that produces a report nobody has to act on is the problem we were trying to fix.
2. Two vetoes, no exceptions.Privacy and accessibility findings rated Critical stop the promotion. Everything else goes to a human.
3. Static and adversarial.The review team reads the code. The red team hits the running system and tries to break it.
4. Six frameworks, one pass.Each agent carries the mapping for its domain. Cross-domain findings go to the supervisor, who can challenge any single agent's verdict.
5. Tiered cost.Routine changes get the relevant specialist. The full fifteen-agent sweep runs weekly. The full pipeline costs about nine times a single review session in tokens.

What it found

We pointed it at our own production platform. Two coordinated audits found 90 issues. The compliance sweep found 56: 14 Critical, 14 High, 24 Medium, 4 Low. The WCAG 2.2 AA audit found 34: 6 Critical, 18 High, 10 Medium. Five of the compliance findings were legal blockers across GDPR, the EU AI Act, BIPA and CCPA. Six accessibility findings made the platform's core AI feature invisible to screen reader users. The estimated WCAG conformance was 45 to 55 percent. The system issued REJECT against the codebase of the people who built it.

The paper reports every finding, including the ones we would have preferred not to discover. The compound risks were the useful part: AI feature work and accessibility work had proceeded on separate tracks with no awareness of each other, and no single-framework audit would have caught that.

Why it matters

The hours and the findings are the headline. The point is cadence. Compliance that runs on every commit is a control. Compliance that runs once a year is a snapshot of a system that no longer exists by the time the report lands. This system does not replace counsel, assistive-technology testing or human judgment. It makes real oversight possible by doing the correlation and the bookkeeping, and putting people in front of the decisions that need people. That is the same principle behind Security Scan Mapper, applied to AI governance.

On the paper

The biggest limitation is the one you would expect: one platform, one stack, and recall was never formally measured. I serve on the AIES 2026 program committee and expect to revise this for another venue.