Policy drafting looks like ideal AI work. The language is formal. The documents are repetitive. The output often needs summaries, explanations, and manager guidance. That surface logic is why teams move too fast. In HR, policy text does more than inform. It shapes manager behavior, employee expectations, escalation paths, and internal records. A fluent draft can create false confidence if the source material, audience, and approval line are weak.
AI can still help here. It helps most when the task is framed around approved source text and a defined audience.
Start with a source hierarchy
Every policy prompt should begin with a source hierarchy. This tells the model what outranks what and tells the reviewer what to defend.
For most HR teams, the source hierarchy should look like this:
- approved policy text or approved draft text
- legal or compliance comments that are cleared for use
- internal owner notes about audience, rollout, or training needs
- prior communication examples, only if they do not override the approved text
If that hierarchy is missing, the model can blend old language, informal notes, and half-settled ideas into one smooth answer. That is a fast way to create confusion.
Workflow 1: compare old and new policy language
One of the most useful policy prompts is simple comparison. Policy owners often need a fast way to see what changed and what a manager needs to notice.
Prompt template:
Compare these two policy versions.
Source priority: the new version is the controlling text.
Return:
1. material changes,
2. wording changes with no practical effect,
3. places where a manager may misread the new language,
4. open questions for the policy owner.
Do not state that a change is approved unless the source text says it is approved.
Old version:
[paste old text]
New version:
[paste new text]How to use it:
- Input: the previous approved version and the new approved draft
- Next step: confirm that the comparison did not miss a high-consequence change
- Expected outcome: a faster first pass for review and communication planning
This use is strong because it keeps the source visible. The model is reading, not guessing.
Workflow 2: approved policy to manager FAQ
Managers rarely need the full policy text in the first pass. They need a plain-language FAQ that stays faithful to the source.
Prompt template:
Create a manager FAQ from this approved policy text.
Audience: frontline managers.
Return:
1. the top questions a manager is likely to ask,
2. the approved answer in plain language,
3. when the manager should escalate to HR,
4. which phrases from the policy must stay unchanged.
Use only the source text provided.
If the source does not answer a question, say that escalation is required.How to use it:
- Input: approved policy text and any owner notes about audience
- Next step: check that the FAQ did not convert policy silence into a confident answer
- Expected outcome: a more usable manager aid with fewer wording errors
The important control is the escalation rule. An honest “escalate to HR” is better than a smooth unsupported answer.
Workflow 3: approved language to rollout communication
After policy text is approved, teams often need rollout communication for employees or managers. AI can help draft this, especially when the team wants one short explanation plus a few action steps.
Prompt template:
Draft a rollout message from this approved policy language.
Audience: [employees or managers].
Return:
1. what is changing,
2. when it takes effect,
3. what the audience needs to do,
4. where questions should go.
Keep the tone clear and direct.
Do not add promises, exceptions, or interpretations that are not in the approved source.How to use it:
- Input: approved policy language and rollout details
- Next step: compare the message against the source and the intended audience
- Expected outcome: a faster draft that the owner can correct quickly
This works because the drafting job is narrow and the source text is already defined.
Where policy prompts should stop
An HR policy library should support drafting around approved text. It should not normalize these uses:
- writing final legal language from a blank prompt
- creating jurisdiction-specific claims from memory
- implying that a FAQ answer overrides the policy
- turning unresolved discussion notes into official guidance
These uses fail for the same reason: they shift authority from the owner and reviewer to the model.
The sign-off line must stay visible
Every saved policy prompt should name the sign-off line directly:
- who owns the source text
- who reviews legal or compliance meaning
- who approves the audience-facing draft
- what version of the source text the prompt depends on
If that line is invisible, the prompt becomes detached from authority. Detached prompts create policy drift.
Save the prompt with its source rule
For policy work, the reusable unit is:
- the prompt
- the source hierarchy
- the sign-off line
- the last approved source version
Saving all four together keeps the asset honest. It also makes later updates easier, because the reviewer can see exactly what the draft depended on.
Policy prompting works when the model stays downstream of approved text and upstream of human sign-off. Comparison drafts, manager FAQs, and rollout messages are useful because they reduce rewrite work without changing who owns the meaning. If the source version, escalation path, and approval line stay visible, prompts can accelerate communication without weakening policy control.