Short answer
A good AI HR use case starts with a specific work decision and a small, reviewable test. Define who participates, what material is permitted, what the output may support and who checks it. Start with a reversible workflow where people can correct context, then expand only if the team can show that the process helped without creating unowned work or unsupported conclusions.
Do not begin with a list of departments
“Recruiting,” “engagement” and “performance” are broad areas, not use cases. A testable use case names the work moment and decision. For example:
| Broad area | Testable use case | Human decision |
|---|---|---|
| Onboarding | Gather accounts of a first-week handover | Whether to clarify a process step |
| Manager development | Prepare questions for a development discussion | What support or practice to agree |
| Knowledge sharing | Capture a specific local practice | Whether the practice is ready for validation |
| Employee feedback | Organise accounts around a defined operating issue | What the responsible team should investigate |
Each has an output that a named person can review. None requires an AI system to assess an individual, decide pay or make an employment decision.
The first-test canvas
Use these fields before inviting participants:
- Work question: What should the organisation understand or decide?
- Population: Which roles, location and period are in scope?
- Input boundary: What information is permitted, and what is excluded?
- Output boundary: Is the output a question, a brief, a draft practice or a record for review?
- Reviewer: Who checks accuracy, context and the next step?
- Stop condition: What failure, ambiguity or burden means the test should pause?
- Expansion condition: What observed evidence would justify another cohort?
This is not paperwork for its own sake. It reveals whether a proposal has an operating owner or only an appealing title.
Fictional example: first-week handovers
Fictional example: a US distribution site wants to understand why new starters arrive at a shift without a clear handover. It invites a limited group to describe one recent handover, its timing and what information was missing. A process owner reviews the accounts alongside the current instruction and prepares one question for the next manager discussion.
The test does not claim to diagnose performance or prove the handover caused later outcomes. It asks whether the accounts reveal a specific, checkable process gap. If they do, the team tries a small clarified handover step and sets a point to review it. If access, language or review effort makes the process unworkable, that is a useful test result too.
Choose a use case with appropriate controls
The EEOC and DOJ warn that AI and software used in employment decisions can create disability-discrimination risks, including where workers are screened out or disability-related inquiries result. Read the official notice. This does not mean every HR use case is the same; it means the decision boundary should be explicit.
For a first test, prefer a workflow that:
- supports a conversation, process check or approved knowledge capture;
- has a human reviewer able to inspect the source context;
- can be paused or corrected without affecting an employment outcome;
- has a defined participant experience and response route;
- avoids treating a generated pattern as a conclusion about a person.
NIST's AI Risk Management Framework provides a voluntary frame for considering those risks in design, use and evaluation.
Where Lontra fits
Lontra can support a focused employee conversation and a manager brief around one defined work question. It is useful where context will help a responsible person prepare a human follow-up. Managers receive briefs, not raw conversations, and the responsible people decide what to investigate or change. Explore the product.
For a full implementation sequence after a use case has been selected, see the AI HR implementation guide. For a broader map of AI roles in the people function, see AI and HR in 2026.
Frequently asked questions
What is a good first AI HR use case?
A good first use case has one named work decision, a limited participant group, permitted inputs, a human reviewer and a way to tell whether the workflow helped.
Should an AI HR pilot start with a high-stakes decision?
No. Start with a reversible, bounded workflow where people can inspect and correct the output before any wider use.