AI for Writing Policies
Policies are where writing meets law: documents that create obligations, get quoted in disputes, and must be understood by every employee on their worst day. AI helps with the two hardest writing problems policies have — getting a structured first draft out of a set of decisions, and translating legalistic prose into plain English people can actually follow. What it must never touch is the decision layer: what the policy requires, and whether it’s approved.
Why this task matters
Every organisation runs on policies, and most organisations’ policies are somewhere between stale and contradictory — written years ago, amended by email, existing in three versions of which two are ‘final’. Writing them is slow because the stakes are high and the blank page is intimidating; updating them is slower because nobody owns it. The result: staff follow folklore instead, and the gap between the written rule and the real rule widens every year.
The traditional workflow
- Someone (eventually) is assigned the draft
- Research: old policy, other organisations’ versions, the actual decisions made
- Draft in the template; wrestle the structure
- Consultation rounds; legal or HR review where it applies
- Approval, publication — and the old version lingers somewhere findable
How AI can help
Draft
- A structured first draft from the actual decisions: scope, requirements, responsibilities, exceptions
- Supporting material: the summary for the intranet, the FAQ, the announcement
Summarise
- A long or legalistic existing policy into a plain-English summary for staff
- Consultation feedback into themes for the review meeting
Analyse
- Plain-language review: reading level, jargon, ambiguity in the requirements
- Consistency: does the draft contradict the related procedures or the code of conduct?
- Gap-check against your policy template: definitions, review date, ownership, related documents
What must stay human
What the policy requires is a leadership decision, not a drafting choice — every ‘must’, ‘should’ and exception is judgement. Legal effect demands human authority: policies with employment, safety, privacy or financial consequences need qualified review, and AI drafting doesn’t reduce that need by one minute. Approval is accountability. And lifecycle is human discipline: one current version, superseded versions properly retired — the plain-English draft means nothing if staff (or an AI assistant) can still find the 2019 version marked final.
Traffic light assessment
Plain-English summaries of existing approved policies; consultation-feedback synthesis. The source is settled and checkable; you’re translating, not deciding.
First drafts of new or revised policies from documented decisions. Genuinely useful, but every requirement gets verified against what was actually decided, and consultation and review proceed exactly as before.
Deciding requirements; anything with legal effect going out without qualified review; policies touching individual employees’ specific circumstances. Requirements are governance decisions. Legal-effect documents are human-authored with AI assistance at most — never the reverse.
Example prompt
For turning decisions into a first draft:
Draft a Working From Home policy from the decision notes below, using this structure: Purpose, Scope, Policy statements, Responsibilities, Exceptions, Definitions, Review. Constraints: plain English an employee reads once and understands, requirements as ‘must’ and discretions as ‘may’ — exactly as the notes specify, don’t invent requirements the notes don’t contain, and flag any decision that’s ambiguous or missing (like who approves exceptions) rather than filling the gap. Audience: all staff. Length: under 900 words. [paste decision notes]
The risks
The legal-effect risk leads: policies create obligations, and an AI-invented requirement — or a softened one — has real consequences; qualified review is non-negotiable for anything touching employment, safety, privacy or money. Confidentiality applies to the inputs: decision notes and consultation feedback often contain sensitive specifics. And the quiet killer is version chaos: publishing a new policy without retiring the old one means every assistant, search engine and anxious employee may still find the wrong answer — lifecycle discipline is part of the writing task, not an afterthought.
A better workflow
The current way
- The draft waits months for a volunteer
- Legalistic prose staff can’t apply
- Old versions linger; folklore fills the gap
The AI-assisted way
- Decisions documented first (they were the hard part all along)
- AI drafts the structured policy and its plain-English summary the same week
- Human consultation, qualified review and approval as always; publication includes retiring the superseded version
What improves
- Policy debt actually shrinks — the drafting bottleneck was the queue’s cause
- Staff-facing language improves organisation-wide
- Reviews focus on substance because the prose arrives clean
- One current version, findable by humans and AI alike — which was the entire point
Key takeaways
- AI drafts structure and plain English; requirements and approval stay human
- Legal-effect documents always get qualified review — AI changes nothing there
- Draft from documented decisions; flag gaps, never fill them silently
- Consultation is a relationship process no tool replaces
- Retiring the old version is part of publishing the new one