autonomy should be a dial, not a switch

Most AI governance conversations in HR are conducted as a series of yes or no decisions. Do we allow AI in screening. Do we allow it in performance calibration. Do we allow it to draft a development plan.

The framing loses on both sides. A no blocks value where the risk was negligible. A yes grants independence nobody specified, because the question never asked what the system may do without a person in the path.

Autonomy is not a property of a system. It is a setting, chosen per use case, and it should be adjustable in both directions. The useful governance question is not whether to allow AI in a process. It is how far up the autonomy scale a particular use case has earned the right to sit, and what evidence would move it.

what the regulation asks of a deployer

The regulation supports this framing more than most compliance summaries suggest.

The EU AI Act classifies employment AI as high risk in Annex III, point 4. That covers systems used for recruitment or selection, including placing targeted job advertisements, analysing and filtering applications and evaluating candidates. It also covers systems used to make decisions affecting terms of work-related relationships, promotion or termination, to allocate tasks based on individual behaviour or personal traits, and to monitor and evaluate performance and behaviour.

Two provisions matter for the argument here. Article 14 requires human oversight measures that are commensurate with the risks, the level of autonomy and the context of use. The regulation already assumes autonomy is a variable. Article 26 places specific duties on deployers: use the system according to instructions, assign human oversight to people with the necessary competence, training and authority, monitor operation, and keep the logs the system generates for at least six months where those logs are under your control.

Article 86 gives a person adversely affected by a decision based on an Annex III system's output the right to clear and meaningful explanations of the system's role in that decision and of the decision's main elements.

On timing, the Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It moved the application date for standalone Annex III high-risk systems from 2 August 2026 to 2 December 2027. The transparency obligations in Article 50 stayed on their original schedule. The AI literacy obligation in Article 4 has applied since 2 February 2025, in a softened form.

The deferral is sixteen months of preparation time, not a change of direction. Deployer duties around oversight, competence and logging are unchanged in substance.

the enforcement record points at the logs

New York City's Local Law 144 shows what happens when an employment AI rule meets practice. It took effect on 1 January 2023, with enforcement from 5 July 2023. Employers using automated employment decision tools must commission an independent bias audit annually, publish a summary of the results, and give candidates at least ten business days of notice. Penalties reach 1,500 dollars per violation per day.

Researchers presented a study of the law's first year at the ACM Conference on Fairness, Accountability and Transparency in June 2024. Of 391 employers reviewed, 267 had active New York City job listings. Eighteen had posted a bias audit report. Thirteen had posted the required transparency notice. Of 386 published impact ratios, 377 exceeded the 0.8 threshold used in US discrimination analysis.

The authors named the pattern null compliance: when a rule leaves scope for the regulated party to decide whether it applies, non-compliance becomes indistinguishable from non-applicability. The lesson for a governance design is not that disclosure rules fail. It is that governance which depends on a discretionary act of publication will produce very little evidence, and evidence is the thing you will be asked for.

four levels, not two

Here is a maturity model you can apply per use case. Every level is a legitimate resting place.

level 1: observe

The system reads and describes. It surfaces a pattern, flags an anomaly, answers a question about existing data. It produces no proposal about any individual. Attrition risk surfacing at team level sits here, as does answering a question about headcount movement.

level 2: draft

The system produces content a person will edit and own. A job description first draft, a summary of a performance cycle, a suggested set of interview questions. The person who signs is the author, and the system's output has no standing until they do.

level 3: recommend

The system produces a proposal about a specific person with a stated rationale, and a person decides. Shortlist ranking, a suggested internal match, a flagged promotion candidate. This is the level where most employment AI risk concentrates, because the recommendation shapes the decision even when the human retains formal authority.

level 4: act

The system takes an action without prior human approval, inside pre-agreed boundaries. Scheduling an interview, releasing a routine notification, moving a candidate between stages on an objective criterion. Very little employment decision-making belongs here, and what does belong is usually procedural rather than evaluative.

Most organisations should hold evaluative use cases at levels one and two, run a small number at level three under close observation, and reserve level four for reversible, low-consequence actions.

what to log at each level

The log is the governance. Without it, every other control is an assertion. Each level adds to the level below it.

at observe: the query or trigger, the data sources and populations touched, the definitions applied, the timestamp, and the identity of the requester. This is the minimum that lets you reconstruct what was seen and by whom.
at draft: everything above, plus the generated output as produced, the final version as published, and the delta between them. The delta is the single most useful governance signal you can collect. Drafts accepted without edit are the ones to sample.
at recommend: everything above, plus the candidate set considered, the features or factors that drove the ranking, the rationale shown to the decision maker, the decision they took, and whether they followed the recommendation. Override rate by decision maker and by affected group is the metric that tells you whether human oversight is real or ceremonial.
at act: everything above, plus the boundary conditions that authorised the action, the identity under which the system acted, the action taken, its reversibility status, and the escalation path if it was outside the boundary.

Six months is the retention floor under Article 26. Set it longer where local employment claim windows run longer, and align it with local data protection obligations.

three gates for moving a use case up

Promotion between levels should require evidence, not confidence. Three gates, all of them mandatory.

the accuracy gate. At least one full cycle of shadow operation at the level below, with measured agreement between system output and the decisions humans made independently. Disagreements reviewed individually, not in aggregate.
the disparity gate. Outcome rates compared across protected groups at the proposed level, with a stated threshold agreed before the test rather than after. Sub-threshold results block promotion and trigger investigation.
the reversibility gate. A documented answer to one question: if this system is wrong about one person, how does that person's situation get corrected, who is accountable for correcting it, and how long does it take. If the answer is unclear, the use case is not ready.

Demotion should be as routine as promotion. A use case that fails a quarterly review moves down a level and stays there until it passes again. Governance that only moves in one direction is not a dial.

the regional picture

Enterprises across Southeast Asia, the Middle East and India are not waiting on Brussels.

Singapore's Infocomm Media Development Authority launched a Model AI Governance Framework for Agentic AI on 22 January 2026. It is voluntary, and it is built around four dimensions: bounding risk upfront, making humans meaningfully accountable, implementing technical controls, and enabling end-user responsibility. Its guidance maps closely to the model above: define checkpoints requiring human approval before high-stakes, irreversible or atypical actions, and maintain traceability from an agent's action back to the human authority under which it acted.

India's Digital Personal Data Protection Rules were notified by the Ministry of Electronics and Information Technology on 14 November 2025, giving operational shape to the 2023 Act. Saudi Arabia's Data and Artificial Intelligence Authority published its Principles and Controls of AI Ethics in September 2023, organised around seven principles including transparency and explainability and accountability.

The details differ. The common demand does not: show what the system did, show who was accountable, and show that a person could have intervened.

what a dial buys you

An on-off switch forces a single decision under maximum uncertainty. A dial lets you deploy at a level you can defend today, collect the evidence that would justify moving up, and move back down without declaring failure.

It also changes the conversation with risk and legal functions, because the question shifts from whether to trust a system to what the system will be permitted to do and what record it will leave. That question has answers. entomo built enzo to be answerable in exactly that sense, with a citation attached to every response rather than available on request.

Start by listing your current AI use cases in HR and assigning each one a level. Most organisations discover that at least one use case is operating a level higher than anyone intended, and that nobody is logging enough to prove otherwise.