Generative AI in HR: A Source-to-Review Workflow

Short answer

Generative AI in HR is most useful as a controlled drafting workflow: start with approved source material, specify the audience and output, check each factual claim against the source, and have an accountable person approve the result. It can make a first draft faster. It cannot establish that an old policy is current, that a summary is complete or that a recommendation is suitable for an employment decision.

Choose a deliverable that can be checked

Start with one reusable deliverable: a manager briefing based on an approved policy, a job-description draft based on a reviewed role analysis, a learning outline based on a validated practice, or a set of conversation prompts. Avoid an open request to “improve HR.” A narrow deliverable reveals what sources, reviewers and checks are actually needed.

The NIST AI resources advocate a risk-based approach to AI. In practice, the more an output could affect people, the more carefully its source, limitation and approval route should be defined. A public-facing policy explanation and a private outline for a workshop are not the same use case.

Use this source-to-review workflow

Try this drafting instruction

Use this original template with approved material in a tool your organisation permits. Replace the brackets before running it.

Prepare a draft briefing for [audience] about [specific task]. Use only the supplied [document names and versions]. For each factual statement, identify the supporting document and section. Preserve eligibility conditions, dates and exceptions. If a source is missing, unclear or contradictory, put the question in an unresolved-items list. Do not invent a rule or decide an individual case. Return: a short briefing, a source table, and questions for [named approving role]. Label the output as a draft.

A source reference in the output is a claim to check. Open the cited section and confirm that it supports the sentence; a model can produce a plausible but incorrect reference.

For a fictional quality test, imagine the approved source says an onboarding check-in is for new starters after their first completed week. A draft says “All employees must complete a check-in every week.” Reject the changed population, mandatory wording and repeated schedule. Correct the draft from the source and add that failure to the next test. A shorter briefing is useful only if it preserves the rule it explains.

A practical output-check sheet

Copy these questions beside any generated HR output:

1. Which source supports each material claim, number, rule or date?
2. Is the source current for this workforce, country and audience?
3. Did the output omit an exception, eligibility condition or required escalation route?
4. Does it describe a person, group or risk more confidently than the source allows?
5. Could a reader mistake a draft for a policy, employment decision or legal advice?
6. Who must review it before it is shared, and where will the approved version live?

The CIPD's AI skills-planning guidance is practical guidance for people professionals responding to AI-driven change. It does not substitute for an organisation's current policy, role analysis or local decision process.

A fictional example: manager briefing

This example is fictional. An HR team wants a short manager briefing explaining a revised parental-leave process. It provides the approved policy, version date, eligible populations and contact route. The instruction asks for a plain-language outline with source references and a list of questions the policy does not answer.

The policy owner checks every eligibility statement, date and escalation route. A manager reviews whether the sequence makes sense in a real conversation. The final brief links back to the official policy rather than presenting the generated text as the authority. The team records an ambiguous phrase for the next policy update.

This workflow improves preparation. It does not let a model determine an employee's entitlement, infer a personal circumstance or replace the policy owner.

Use generative output alongside, not instead of, employee context

Generated content can explain a process, but it cannot know whether the process works in daily practice unless relevant people supply and review that context. A manager briefing about onboarding may need employee accounts and operational evidence before it is changed. Keep those sources separate from generative text and state what remains uncertain.

The SHRM State of AI in HR 2026 overview stresses the combination of technology and human judgement in HR. That is a useful direction, not evidence that any generated output is correct without review.

For selecting a use case and review level, see AI and HR in 2026. Lontra's product can support a bounded employee-conversation question and human-reviewed manager brief. It is not an HR policy generator, system of record or automated employment decision-maker.

Sources

• NIST artificial intelligence and risk-management resources
• CIPD AI skills planning guidance
• SHRM State of AI in HR 2026 overview