How to Draft HR Policies With AI (and Keep Counsel in the Loop)

TL;DR: Policy work is slow because drafting is slow, assembling structure, finding precedent language, keeping fifty documents consistent. AI collapses that part: a competent first draft in minutes, a plain-language rewrite of a lawyer-dense section in seconds, a consistency sweep across your whole handbook in an afternoon. What it does not collapse is the part that makes a policy safe to publish: knowing what the law in your jurisdictions actually requires. A large language model will state legal requirements with total confidence and no warranty, some real, some outdated, some invented. This guide gives you the split: AI owns the drafting labor, humans own every factual claim, and employment counsel owns the sign-off.

This guide is part of the AI for HR hub. The documents you produce here are also the corpus for a grounded employee Q&A assistant, see AI for employee onboarding.

Why policies punish AI shortcuts

Three properties make policy drafting medium-high risk in the hub’s ranking:

1. Policies are read literally, later, by adversaries. A handbook clause gets quoted back in disputes, unemployment hearings, and litigation. Ambiguous or accidentally generous language becomes a commitment. Some clauses (at-will disclaimers, arbitration provisions, PTO accrual terms) have jurisdiction-specific magic words, and their absence or mangling has consequences.
2. The law is jurisdictional and moving. Leave laws, pay transparency, non-compete enforceability, drug-testing rules, and required postings differ by country, state, and sometimes city, and change every year. Model training data lags reality; a model’s confident summary of “the law” is a snapshot of the internet at some past date, averaged.
3. Hallucinated law looks exactly like real law. In most writing, a hallucination is embarrassing. In a policy, an invented “requirement” gets institutionalized, and a missed real requirement becomes a violation with your logo on it. You cannot proofread your way out of this without legal knowledge, which is why review is a role, not a step.

The conclusion isn’t “don’t use AI.” Counsel reviewing a clean, complete, well-structured draft costs a fraction of counsel drafting from scratch, and AI is what gets you the clean draft.

The drafting workflow

Step 1: Decide positions before prompting

A policy encodes decisions: How much leave? What’s the remote-work rule? Who approves exceptions? AI cannot make these, if you don’t supply them, it silently defaults to internet-average positions, and you end up debating a phantom policy nobody chose. Write the decisions as bullet points first; the drafting session turns decisions into documents, not the reverse.

Step 2: Draft against your decisions and your context

Example prompt (first draft): “Draft a [remote work] policy for a company of [size] operating in [jurisdictions]. Use ONLY the positions I’ve decided: [paste decisions]. Structure: purpose, scope (who’s covered), the policy itself, procedures, exceptions process, and effective date. Plain language at an 8th, 10th grade reading level. Where a position I haven’t decided is needed, insert [DECISION NEEDED: question], do not choose for me. Where legal requirements may apply, insert [LEGAL REVIEW: topic] rather than stating what the law is.”

The two bracket conventions do the safety work. [DECISION NEEDED] stops silent defaults; [LEGAL REVIEW] stops the model from playing lawyer while still flagging where the real one should look. A draft that arrives at counsel with its legal questions pre-flagged is dramatically cheaper to review.

Step 3: Use AI where it beats humans, the editing passes

Drafting gets the attention, but editing is where AI is most reliably excellent, because every task below is grounded in text you provide rather than knowledge the model claims:

The ambiguity hunt deserves special mention: adversarial reading is tedious for humans and effortless for a model, and it routinely finds the sentence that would have cost you.

Step 4: Human verification of every claim

Before counsel ever sees it, a named HR owner verifies: numbers (accrual rates, notice periods, thresholds) against the decided positions; names, systems, and processes against reality; and that every [DECISION NEEDED] has been resolved by a human with authority to decide. AI-drafted specifics that nobody decided are the most common way phantom policy ships.

Step 5: Legal review, scoped, not skipped

Employment counsel reviews before publication. To keep it efficient rather than performative:

• Send the decided-positions list alongside the draft, so counsel reviews intent as well as text.
• Point them at the [LEGAL REVIEW] flags first, but don’t limit them to it, the flags are the model’s guesses about where law lives, not a warranty.
• Batch related policies into one review cycle.
• Record what counsel changed. Over time this teaches your team where the legal landmines cluster, and it’s your evidence of diligence.

For lightweight internal process docs that create no entitlements (meeting norms, tooling guidelines), a defined lighter path, HR owner review only, is reasonable. Write down which path each document class takes; the classification itself should get counsel’s blessing once.

Step 6: Publish with versioning, then keep it alive

Version number, effective date, owner, and next-review date on every policy. Feed the approved versions into your grounded Q&A assistant, and only approved versions; a draft in the corpus becomes confidently cited fake policy. When laws change, AI helps you find the blast radius (“which of these policies mention parental leave, notice periods, or [topic]?”) but counsel defines what the change must be.

Governance: put the rules in writing

Two meta-policies make this workflow safe at scale. First, your AI acceptable use policy should name which tools may touch policy work, what data may be pasted (no employee names or case details in policy prompts, policies are general documents), and who may publish. Second, your rollout sequencing: policy drafting belongs after your team has built review discipline on lower-stakes content, per the AI adoption roadmap. A team that skips review steps on job descriptions will skip them on the termination policy too, and only one of those mistakes is cheap.

FAQ

Can AI write a legally compliant HR policy? Not reliably, it hallucinates requirements, misses jurisdictional rules, and trains on outdated law, all while sounding certain. Use it for drafting and editing labor; employment counsel confirms compliance before publication. The economics still work: counsel reviewing a clean draft costs far less than counsel drafting.

Do small companies without in-house counsel still need legal review? For binding policies, yes. Keep costs sane by batching AI-prepped drafts for outside counsel, starting from employer-association templates, or using an HR compliance service, but don’t publish unreviewed AI output as policy. A wrong policy creates commitments; no policy at least doesn’t.

Which policies are safest to start with? Internal process documents that create no legal entitlements, meeting norms, equipment guidelines, how-tos. Build the drafting-and-review habit there; bring counsel in before touching leave, accommodation, discipline, or termination policies.

Can we use AI to answer employees’ questions about policies? Yes, via a grounded assistant that answers only from approved policy documents with citations, covered in the onboarding guide. Never let a general chatbot improvise policy answers.

Not sure whether your policy stack, or your review process, is ready for AI-assisted drafting? The free AI readiness assessment gives you a prioritized starting point in about ten minutes.