AI Training for HR: Keep Health Inference Out of Absence Summaries
Train HR teams to produce a minimal operational absence summary without exposing health details, inventing causes or letting AI make a people decision.

AI training for HR should include a difficult but common review skill: summarising an absence note without turning limited information into a health story. A fluent draft may repeat details that the recipient does not need, infer a cause that the employee never stated or make a management recommendation that belongs to an authorised person.
The training goal is not to teach an AI tool to process live health records. Use synthetic or explicitly approved material in an approved environment. Teach HR staff to minimise the input, separate stated facts from inference and keep support, attendance and employment decisions outside the model.
Why AI training for HR needs an inference boundary
The European Commission’s current AI-literacy guidance says learning should reflect the system, the user’s role, the context and the people affected. It explicitly notes that staff using tools for writing or translation need to understand risks such as hallucination. The UK Government AI Playbook adds data minimisation, accuracy, clear responsibility and meaningful human control for work involving personal data or consequential decisions.
For an HR exercise, that means the pass condition is not elegant prose. It is a small, necessary record whose facts can be traced to the source, whose unknowns stay unknown and whose next decision is routed to the right person.
The four-part absence-summary boundary
Purpose
Name the operational need before using a tool: scheduling, a follow-up date or a request for an authorised HR conversation. Do not collect detail merely because it appears in the source.
Minimum facts
Keep only facts needed for that purpose, such as dates, availability, the employee’s stated request and an agreed contact route.
No inference
Separate the employee’s words from assumptions. Do not add a diagnosis, cause, prognosis, attitude or capability judgement that the source does not state.
Human route
Name who may decide the next action and what needs specialist review. The AI may structure a draft; it does not approve leave, fitness, adjustment or performance action.
Do not confuse a tidy summary with a permitted record
| Unsafe draft move | Safer training response | |
|---|---|---|
| Input | Paste the complete live case into any familiar tool | Use a synthetic case or an explicitly approved route with only necessary fields |
| Cause | Turn 'not feeling well' into a named condition | Write only what the employee stated, or omit the cause when it is not needed |
| Forecast | Predict the return date from previous absences | Record the stated date and mark confirmation as pending |
| Judgement | Recommend approval, discipline or capability action | Route the decision to the authorised HR or management process |
| Audience | Copy medical detail into a manager update | Give the recipient only the operational information they are allowed and need to act on |
| Record | Save the whole conversation as evidence | Keep the approved summary and required decision record under the organisation’s rules |
This workflow is a learning pattern, not legal advice or permission to process employee information. Organisations must follow the law, contracts, policies, approved tools and specialist guidance that apply to them. If the permitted purpose, account or reviewer is unclear, the correct training outcome is to stop and route the question.
Worked example: an unsupported health inference
Reading is a start. Practice makes it stick.
Start learningA fictional HR exercise contains this synthetic note: an employee says they will be absent on Monday, hope to return on Wednesday and would like to discuss a temporary change to their morning schedule. The note gives no diagnosis and makes no statement about long-term capability.
The AI draft says the employee is experiencing stress, is likely to need a phased return and may struggle with their normal workload. The wording sounds considerate, but every health and capability statement is invented. It also moves from summarising a request to advising on a people decision.
A safer operational draft says: 'The employee reported absence for Monday, expects to update their return status by Wednesday and requested a conversation about a temporary morning-schedule change. HR follow-up is required; no cause or adjustment decision is recorded in this summary.' The reviewer then checks whether even this information is necessary for the intended recipient and approved record.
Run the review as a training exercise
- 1
Set the approved environment
Use a synthetic case and name the tool, account, audience and data rule before the exercise starts.
- 2
Write the purpose
State the single operational action the summary should support. Remove details that do not serve it.
- 3
Mark source statements
Underline dates, requests and quotations that are explicitly present. Label every other statement as inference or decision.
- 4
Draft the minimum
Keep only necessary facts, visible unknowns and the authorised next route. Do not ask the model to fill gaps.
- 5
Apply human review
Have an HR or policy owner check necessity, accuracy, audience and decision authority before any use.
- 6
Repeat with a changed case
Use a new example where the dates are unclear or the request crosses an authority boundary. Test whether the learner pauses instead of guessing.
Measure the behaviour, not the polish
Score four observable behaviours: the learner uses the approved route, removes unnecessary detail, catches unsupported inference and sends the decision to the authorised person. Do not grade whether the wording sounds warm or professional until those gates pass. A beautifully phrased privacy or authority error is still an error.
Build the wider path with Bokili’s safe-refusal exercise (https://bokili.com/en/learn/ai-training-employees-safe-refusal), the affected-person check (https://bokili.com/en/learn/ai-training-employees-affected-person-check) and the four-part HR policy answer (https://bokili.com/en/learn/ai-training-hr-four-part-policy-answer). Bokili’s HR and L&D programme is available at https://bokili.com/en/for-hr.
- Write a fictional absence note with dates, one employee request and one deliberately missing fact.
- Name the operational purpose and authorised recipient.
- Highlight every health detail or inference that is unnecessary for that purpose.
- Draft a summary using only stated, necessary facts and visible unknowns.
- Add the human owner for the next decision.
- Ask a reviewer to identify one detail to remove and one boundary that needs escalation.
Keep unknown health information unknown
The safest HR summary is not the most complete-looking one. It is the smallest approved record that supports a legitimate next step without inventing a person’s condition, capability or need. Practise the boundary on synthetic cases, make the reviewer visible and stop when the route is unclear. Bokili turns that judgement into short, role-specific practice rather than another general warning.
Sources
- AI Literacy — Questions & Answers — European Commission
- Artificial Intelligence Playbook for the UK Government — UK Government
Reading is a start. Practice makes it stick.
Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.
Start learningKeep reading

Allocate AI Licences to Task-Ready Work, Not Waiting Lists
Use four readiness gates to allocate scarce AI access to work with an approved input, clear outcome, review owner and support route.

AI Skills for Employees: Calibrate Confidence Before Checking
Train employees to record confidence before AI advice, verify the decisive claim and learn from the gap between prediction and evidence.

ChatGPT Projects: Save a Verified Response as a Reusable Source
Turn a checked ChatGPT answer into reusable project context without losing its sources, owner, review date or replacement trigger.