A useful first AI task for HR is one you can check against a reliable source. Drafting an explanation of a PTO policy, for example, gives the reviewer something concrete to inspect: the answer should match the approved policy, preserve its conditions and identify any question the policy does not resolve.
That makes a stronger training exercise than asking a tool to "help with HR." It teaches the practitioner to supply the right information, test the answer and decide what requires an HR judgment.
Ajaia's HR work includes training and implementation. Ari describes a team that completed training and then wanted an agent to answer HR questions; our implementation team built a new front door for those questions. He also describes work on a recruiting agent to help source candidates. These are anonymous practice examples, with no quantified outcomes claimed. The exercises below are fictional and contain no employee or candidate records.
Start with policy explanation before personal eligibility advice
PTO and benefits questions often have a source document, a recurring audience and a person who owns the answer. That gives a team a place to start. It also reveals an important boundary: explaining a policy is different from deciding an individual's eligibility.
Use a current, approved document. Record its effective date, relevant population and policy owner. If the answer depends on location, employment status or another policy, decide how the workflow will collect that information or escalate the question. A clear answer based on the wrong policy version is still wrong.
For an introductory lab, use the short fictional policy below. It is teaching material, not a recommended employment policy:
- Full-time employees receive 15 PTO days per calendar year.
- Employees request PTO in the HR portal. Their manager approves the request.
- Up to five unused days may carry into the next calendar year and must be used by March 31.
- The excerpt does not explain part-time eligibility, cash payouts or exceptions.
Give the tool a bounded task: "Using only this fictional excerpt, draft an answer to: 'Can I carry over eight unused PTO days, and can I get paid for the other three?' Cite the relevant sentence. Identify any part of the question the excerpt does not answer. Do not infer a payout rule."
This exercise tests whether the learner can recognize a missing answer. A response such as "You can carry over five days and receive payment for the other three" invents a benefit. It should fail review even if it sounds clear and reassuring.
A source-grounded answer would say: "The excerpt allows up to five unused days to carry into the next calendar year, to be used by March 31. It does not state whether unused days can be paid out. HR needs to confirm that part of your question."
That sample answer is written for this exercise; it is not a recorded model result. Have participants generate their own answer, inspect it and explain why they would accept or revise it.
Teach the review before teaching a longer prompt
Review the answer against a few explicit checks. Does every policy claim have support in the supplied text? Are the conditions intact? Has the model filled a gap? Can the reader see what to do next?
Ask a second question that the excerpt cannot answer, such as whether a part-time employee receives PTO. The correct outcome is an identified information gap and a route to the responsible person. A confident guess should not count as successful completion.
Asking for a citation does not guarantee that the citation exists or supports the answer. Open the source and compare the relevant text. NIST's Generative AI Profile describes the risk of false content presented confidently. The review step remains necessary when the output includes a convincing explanation.
Once the learner can do this with a short excerpt, try a longer approved source with more conditions. Increase the task's difficulty while keeping the review criteria visible.
Payroll variance analysis needs reliable calculations
AI can help organize questions about a payroll variance. The underlying calculation should come from a checked spreadsheet or payroll system, with access appropriate to the people involved. Keep real payroll records out of an introductory lab.
Consider a fictional aggregate: overtime cost increased from $8,000 to $10,000 between comparable periods. The verified difference is $2,000, or 25% of the earlier amount. Give the tool those checked figures and ask it to draft an investigation checklist.
A useful checklist might ask whether overtime hours, pay rates, staffing, period length or adjustments changed. Those are possible explanations to investigate. The figures alone do not establish the cause. An answer claiming "understaffing caused the increase" has gone beyond the evidence.
Have the learner separate the checked calculation from hypotheses and missing data. The payroll owner should confirm the records, explanations and any proposed correction. A model-generated anomaly does not authorize changing pay.
For a later approved pilot, measure the time needed to produce a reliable variance review, including validation and follow-up. Faster commentary is of limited value if it sends the payroll team to investigate unsupported causes.
Candidate sourcing should produce leads a recruiter can assess
A sourcing exercise can help a recruiter describe a role, draft search terms and organize evidence about possible candidates. Start by having the role owner confirm the actual job requirements. An ambiguous job brief can become a more polished search for the wrong person.
Use fictional profiles or approved public sources in training. Ask the tool to map each relevant requirement to evidence, keep unknowns visible and distinguish an explicit qualification from an inference. A profile that mentions a project does not necessarily establish the person's exact responsibility, results or availability.
Do not turn a sourcing exercise into an automatic rejection or hiring decision. Keep the recruiting owner responsible for the criteria, source permissions and outreach. Before adopting a workflow that evaluates or screens people, involve the organization's employment, privacy and accessibility owners. A human clicking approval does not, by itself, establish that a selection method is fair or appropriate.
The EEOC's guidance on AI and disability discrimination describes risks when software and algorithms assess applicants and employees. Treat employment assessment as a separate, consequential use case rather than an incidental extension of a sourcing demo.
An HR agent needs an owner for the answers it serves
A question-answering agent can become a convenient entry point for HR information. Before building one, inspect how a person currently finds an answer and what HR does when the question is ambiguous.
Define the source library, access boundaries, update process and escalation path. Decide which questions the agent can answer directly and which need a person. Test outdated documents, conflicting sources, missing information and requests for another employee's details. These are proposed evaluation cases, not claims about the anonymous Ajaia implementation.
The team also needs a way to correct an answer and replace stale source material. An agent that was accurate during its launch test can become unreliable when a policy changes. The AI agent evaluation guide helps an implementation owner plan those checks.
Turn the exercise into a repeatable HR habit
Choose one recurring question, one approved source and one reviewer for the first pilot. Have participants complete the task during training and bring a later attempt to office hours. Record what made the answer difficult to produce or verify.
Track whether people are comfortable trying the task, whether answers meet the source and escalation rules, and whether the workflow is used again. Include review time. A high volume of generated answers is not evidence that employees received better HR support.
The HR training program can be scoped around those tasks. Bring a description of the recurring work and your approved tools to the first conversation; confidential employee records are not needed to explain the problem.