AI in Hiring and Performance: Global Compliance

This is not legal advice. It is operational guidance for people who run operations, written to help you ask better questions of counsel. Requirements differ by country, state and sector, and they are moving quickly. Nothing here states what any specific law requires of you — verify your own obligations.

If your software helps decide who gets hired, what work gets assigned, or how someone's performance is assessed, you are probably in scope of a rule somewhere. Which rule depends on where you operate, and that answer is changing faster than most operations teams can track.

The useful move is not to memorise jurisdictions. It is to build a posture that holds across most of them — because the frameworks emerging in different places are converging on the same small set of demands.

The distinction every framework makes

Regulators are broadly converging on one line, and it is worth internalising because it predicts scrutiny better than anything about the model itself:

AI that assists a person is treated differently from AI that makes, or substantially informs, a decision about a person.

Summarising a document is the first kind. Ranking candidates, allocating work, or generating a performance assessment is the second — because the consequence lands on an individual who did not choose the tool and often does not know it was used.

Sophistication is not the trigger. Consequence to a person is. A simple scoring rule that filters CVs attracts more scrutiny than a large model that drafts meeting notes.

What the frameworks tend to ask for

The specifics vary. The categories are remarkably consistent:

That last row is the one operations teams miss most often. In several emerging frameworks the organisation using the system carries duties of its own — which means buying a compliant tool does not, by itself, make you compliant.

The four controls that hold anywhere

This is the practical core, and the reason this page is worth reading even though the legal detail is unsettled: the controls that satisfy most of these frameworks are the same controls, and they are good practice regardless.

1. Know where AI touches a decision about a person

Most companies cannot produce this list, which is itself the first finding.

Inventory every place software ranks, scores, filters or assesses a person — including features inside tools you bought for something else. Applicant tracking, performance software, workforce management and scheduling tools routinely ship AI features that are enabled by default and announced in a release note nobody read.

2. Keep the record

Every framework in this space either requires or assumes one: what the system produced, what informed it, who decided, what actually happened.

Retention periods differ. The requirement to have something worth retaining does not. And the practical consequence is blunt — if a vendor cannot export that record, you cannot meet an obligation to keep it. See how to audit what your AI assistant actually did for the twelve questions that establish whether a vendor can.

3. Name a human, and give them enough to judge with

Not "the team" — a named person accountable for each system, with authority to override and the information needed to exercise it.

Oversight that cannot see the reasoning is not oversight, it is sign-off. If the person responsible cannot say why an output was produced, you have a rubber stamp with a job title.

4. Test for disparate impact, and keep the results

Whether or not a formal audit is currently mandated where you operate, testing whether a tool produces different outcomes across protected groups is what protects you if anyone asks — a regulator, a candidate, or a journalist.

Keeping the results matters as much as running the test. An untested system and a tested system with no records look identical from outside.

Where to check, by region

This is a research checklist, not a statement of law. It names instruments people in each region commonly need to check. It does not describe what any of them requires, when it takes effect, or whether it applies to you — all of which need confirming with counsel for your own circumstances.

If you operate in more than one of these, build to the strictest applicable standard. Maintaining different postures per jurisdiction is more expensive than meeting the highest bar once, and it fails at the seams — usually at the moment a candidate in one country applies for a role posted in another.

Questions for your vendors

Beyond the twelve audit questions:

1. Which of your features make or materially inform decisions about individuals?
2. Have you conducted bias testing? Can we see the methodology and the results?
3. Can we disable AI on person-affecting decisions specifically, without losing the rest?
4. What documentation do you provide to support our compliance obligations?
5. If a candidate or employee asks how a decision was made, what can we tell them?
6. Which jurisdictions have you built for, and what did you change to do so?

Question 4 separates vendors treating this as a product requirement from vendors treating it as a disclaimer. Question 6 is the fastest way to find out whether they have thought about anywhere other than their home market.

What we do

Alygn scores work against organisational priorities, and that scoring informs performance evidence. That places us squarely in the category above, so to be explicit about our own posture:

• Nothing acts autonomously. The system proposes; a human decides.
• Every AI action is logged with its grounding, the human decision, and the model and prompt version — and the record is exportable, not just viewable.
• Scores are explainable. Four separate pillar scores rather than one opaque number, so a score someone disagrees with is arguable rather than mysterious. See the four-pillar method.

We think that is close to the minimum bar for anything in this category, whatever your jurisdiction ends up requiring of you.

Related

• How to audit what your AI assistant actually did
• Evaluating strategy execution software — where these questions belong in a purchase decision