AI in HR: controls, decisions and fairness

Decide where AI may help in HR, see every AI result and the person who decided on it, tell people when AI is used, and check that AI that scores people treats groups fairly.

Contents: Who can use it · The rules AI follows · Turning AI on · Country rules · Data and training · Pausing all AI · Notices · Opting out · Decision log · Fairness reports · Monitoring · FAQ

Who can use it

Everything is on every plan. Permissions on the Roles screen, under AI in HR:

• AI settings (List, Update): see and change the switches, notices and settings, pause AI.
• AI decision logs (List, Show, Export, Decide): read the log, download it, and record the decision on an AI result.
• AI fairness reports (List, Run): read reports, run one, mark a failed one as reviewed.

Roles that could read the privacy settings get the List permissions; roles that could change them get the rest. Employees need no permission to read their own AI notices or to opt out.

Where: HR → Settings → AI in HR (tabs Features, Settings, Notices, Decision log, Fairness, Monitoring). Employees: My HR → AI and my data.

AI uses the AI chosen for the whole workspace in AI settings: 9bx AI, or your own provider. Usage counts toward your AI allowance like every other AI feature.

The rules AI follows

• AI only suggests, ranks, summarizes, drafts, scores, flags, reads documents or proposes. It never rejects a candidate, rates an employee, changes pay, ends employment or approves anything.
• When someone acts on an AI result, a named person records the decision and the reason. For hiring, performance and termination the person also confirms that they reviewed it themselves.
• Every AI result is kept in the decision log with the engine, its version and what it was based on (references to records, not the personal data itself).
• Personal details such as email addresses and phone numbers are hidden from the AI engine unless you turn that off.

Turning AI on

Every AI feature starts off. On the Features tab:

Area	Features
Self-service	HR assistant
Documents and writing	Writing help (job descriptions, letters, policies, announcements), reading ID documents and receipts
Performance	Review drafts and feedback summaries
Hiring	Reading resumes, candidate match scores, interview notes and summaries
People analytics	Payroll, attendance and data quality checks, attrition and absence predictions, questions about HR numbers
Engagement	Survey comment themes
Expenses	Duplicate and fraud checks

1. Turn a feature on for All countries, or open it and turn it on or off for one country of your legal entities. A country setting wins over the all-countries one, so you can turn a feature on everywhere except one country.
2. Turn on the whole area switches every feature of an area at once.
3. Each switch shows who changed it last and when.

Features marked High risk (hiring, performance, predictions) need the confirmation step for every decision. Features marked Fairness tested can be held by a failed fairness report.

If you already used AI resume reading in your hiring settings, Reading resumes is on for you. It still needs the hiring setting on as well.

Country rules

Some countries have special rules for AI in employment decisions. When the country pack of one of your legal entities says so, turning that feature on for the country (or for all countries) shows the rules first. Tick I have read and accept these rules to continue. The text you accepted, who accepted it and when are kept with the switch.

Your data and training

On the Settings tab:

• Allow AI providers to train on our data: off by default. Leave it off unless you have agreed otherwise with your provider.
• Our own AI engine is contracted without training rights: tick it to record that your own engine's contract does not let the provider train on your data. The list of engines in use (your own or the 9bx engine) is shown beside it.
• Hide personal details from AI: on by default.
• Fairness threshold: the lowest acceptable ratio between groups in a fairness report. The default is 0.8 (the four-fifths rule).
• Keep the decision log for (days): 730 by default. A country pack can set its own period for entries about people in that country.

Pausing all AI

If something goes wrong, click Pause all AI, give a reason and confirm. Every HR AI feature stops at once, whatever its switch says. The page shows who paused it, when and why. Click Resume to start again; the switches are as you left them.

Notices

People are told when AI is used:

Notice	Who sees it	Default wording
AI-assisted draft	The person writing	AI-assisted draft. Review it before you use it.
Candidate notice	Candidates	We use AI to help read applications. A person reviews every application and makes every decision. You can ask for your application to be reviewed without AI.
HR assistant	Employees	You are talking to an AI assistant. Its answers can be wrong; check important answers with HR.
HR data checks	Employees	Your organization uses AI to help spot errors and trends in HR data. People review every result, and AI never makes decisions about you.

On the Notices tab you can reword each notice per language. When you leave a language empty, the wording of the country pack is used, then the default. The record of who saw or acknowledged a notice keeps the exact wording shown.

Opting out

• Employees open My HR → AI and my data, read the notices, click Got it on each, and turn on Do not use AI on my data. This is saved as a consent record (purpose "AI processing of HR data"), so it also shows under Privacy → Consents. AI features then skip that person.
• Candidates who ask (for example by email) are recorded by HR: on the candidate, choose Record AI notice, pick Opted out and the source. AI features then skip that candidate.

The decision log

The Decision log tab lists every AI result: feature, area, who or what it was about, engine and version, status and who decided.

Status	Meaning
Waiting for a person	Someone has to accept, change or reject it before anything happens
Accepted	Used as it was
Changed	A person changed it; the version kept is shown next to the AI's
Rejected	Not used
No decision needed	Shown for information only

To decide, open an entry, choose Accept, Change (and give the version you keep) or Reject, and write the reason. For high-risk features tick I reviewed this myself. A decision cannot be changed afterwards.

Filter by feature, area, status, person or dates, and click Export to download the filtered log as a spreadsheet file for an audit. AI results about an employee are also included in their personal data export.

Fairness reports

For features that score or rank people (candidate match scores, predictions), the Fairness tab shows how often each group is recommended:

• Run a report, choose the feature and Past results (the last 90 days, or dates you choose) or Test people (made-up people who differ only in gender or age, where the feature supports it).
• Past results use only what is lawfully held: a candidate's voluntary diversity answers, an employee's gender and age band. Groups with fewer than five people are left out so nobody can be singled out.
• Each group shows its rate and its impact ratio to the best group. Below the threshold, the report fails and the feature is held: it stops until someone reads the report and clicks Mark as reviewed with a note.
• Drift is flagged when the worst ratio fell by 0.1 or more since the last report, or the engine or prompt version changed.
• A report runs every week on its own for each fairness-tested feature that is on.

Monitoring

The Monitoring tab shows, for the last 30 days (or the period you choose), per feature: whether it is on (and why not), how many results, how many wait for a person, how many were accepted, changed or rejected, and the override rate (changed plus rejected, out of all decided). It also counts employees and candidates who opted out.

Troubleshooting and FAQ

A feature says it is off although I turned it on. Check, in order: AI is not paused; the country of the person has no switch of its own that is off; the country's rules were accepted; no failed fairness report is waiting for review; an AI is available in AI settings.

Why can't I decide an entry? It is already decided, or it needed no decision.

Can AI reject a candidate or change pay by itself? No. AI can only produce a result; a person decides.

Does 9bx train AI on our data? No. Training is off by default and the 9bx engine is contracted without training rights.