Why decision-support, never automated screening-out

Building hiring software in 2026 means starting from a different floor than building it in 2020.

In 2024, Mobley v. Workday moved from "novel theory" to "active class certification": a plaintiff successfully argued that an AI-driven applicant-tracking system can itself be the agent of discrimination, distinct from the employer who deploys it. The implication is unambiguous — vendors who sell automated workflows that take candidates out of pipelines can be jointly liable for the disparate impact those workflows produce. The court did not require proof of intent. The court required proof that the system, operating as designed, produced a disparate-impact outcome.

In parallel, the EU AI Act — high-risk-system provisions enforced from August 2026 — codifies Article 22 of the GDPR into hiring-specific obligations: a candidate should never be subject to a decision based solely on automated processing that produces a legal or similarly significant effect. The recruitment platform that flips a candidate from applied to not progressing without a human in the loop is, by construction, a thing the EU AI Act is asking employers not to deploy.

Both signals converge on the same answer. The legal posture of an AI hiring platform is not how accurate is the model; it is what role does the model play in the outcome. The platforms that try to be the decision-maker assume the legal risk of the decision-maker. The platforms that produce decision-support evidence — and let the human recruiter make the call — do not. The shorthand for this distinction in policy circles is "decision-support, never automated rejection."

KimonRecruit is built on the second posture, by construction.

Concretely:

• There is no code path in KimonRecruit that transitions a candidate to a no-progress pipeline stage without a recruiter user_id on the audit row. The database enforces it: a system actor cannot land on the no-progress stage by construction.
• Every assessment score is replayable from the prompt + the model version + the parameters. A recruiter who flags a result as anomalous can re-run the score themselves. The platform does not absorb the decision into a black box.
• Adverse-impact monitoring runs continuously, not retroactively. Four-fifths-rule breaches surface as flags in the recruiter dashboard the moment they cross threshold. The platform does not wait for a tribunal to ask.
• The audit trail is hash-chained and Ed25519-signed. Tribunal-grade evidence does not depend on us — it can be verified off-platform via the published audit-anchor.

None of this is theatre. The platform genuinely cannot send a candidate a no without a recruiter making that call. It is a constraint, not a setting.

That constraint has a cost. KimonRecruit is slower per-pipeline than a platform that auto-screens at the first failed assessment. It moves human attention to the candidate, every time. That is the trade.

The reason for the trade is the same reason this post is the launch essay rather than a marketing one-liner. If you build a tool that can author a no-progress outcome without a human, you have built a tool that can be the subject of Mobley. If you build a tool that cannot, by construction, author that outcome alone, you have moved the legal exposure back to the human decision-maker — where the Equality Act 2010 and the EU AI Act both, separately, locate it.

Decision-support, never automated screening-out. That is the floor. Everything else is detail.