AI can take an employee handbook from blank page to solid draft in an afternoon. It can also hand you an unlawful work rule you will not notice until someone files a complaint. Here is where the line sits.
AI is genuinely good at the boring 70% of an employee handbook: structure, plain English, tone, gap analysis against a checklist. It is dangerous on the other 30%, because the legally loaded sections are the ones where a single overbroad sentence creates real liability. This guide gives you a three-pile sorting method, a five-step workflow, the prompts that actually work, and the short list of sections that must never leave your building without a lawyer reading them. It also covers the section almost every handbook now needs and most get wrong: your AI-use policy.
Most handbooks are not out of date because HR is lazy. They are out of date because employment law moves faster than any one person can track. Brightmine surveyed HR teams and found that 62% of companies review their employee handbook for compliance about once a year or less, and fewer than half feel very confident it reflects current law.[1] That is not negligence. That is arithmetic.
Then AI arrived in the middle of all that, from two directions at once.
The first direction is your staff. Gallup found that 45% of US employees now use AI at work at least occasionally, and only 37% said their organisation had actually implemented it.[2] Read those two numbers together and you get the real situation: people are using tools nobody approved, on documents nobody vetted. Your handbook is probably silent on all of it.
The second direction is regulation. SHRM’s 2026 research found that 19 of the most populous US states have enacted laws or regulations covering employer use of AI, and 57% of HR professionals working in those states did not know they existed.[3] Employers have noticed the risk even where they have not fixed it: Littler’s annual survey found 68% now have a formal AI policy, up from 38% a year earlier, and 79% expect AI-related litigation in the next twelve months.[4]
So you are being asked to rewrite a document you barely have time to review, in a legal environment that changed twice while you were reading this, and to add a brand new section about the exact technology you are considering using to write it. I understand the appeal of opening ChatGPT and typing “write me an employee handbook.” Please do not do that. Do this instead.
The most useful thing I teach HR teams in workshops is not a prompt. It is a sorting question, and it takes about twenty minutes with a printed table of contents and three highlighters.
The question is not “can AI write this section?” AI can write anything. The question is what happens if this section is subtly wrong and nobody catches it for eighteen months? Sort every section by that answer and the whole project gets simpler.
| Pile | Sections | Who does what | If it is wrong |
|---|---|---|---|
| Green | Welcome letter, mission and values, dress code, expenses, travel, equipment care, communication norms, benefits summaries that point to plan documents, glossary, table of contents | AI drafts, HR edits | Someone is mildly confused. You fix it in the next version. |
| Amber | Attendance, remote and hybrid working, performance management framing, IT and acceptable use, social media conduct, the AI-use policy | AI drafts, HR edits, lawyer reviews the finished text | You have created an expectation you did not mean to create, or a rule that quietly chills protected activity. |
| Red | At-will statement and disclaimer, conduct and confidentiality rules, all leave policies, pay and classification, discipline and termination, arbitration, anti-harassment complaint procedure | Lawyer drafts or rewrites; AI is used only to check plain English after sign-off | Money. Sometimes a lot of it. |
How to allocate handbook sections between AI drafting and legal review, based on the consequence of an undetected error. Red-pile categories reflect NLRB, SHRM and Brightmine guidance cited throughout this article.
Here is the part that surprises people: the green pile is usually most of the document by page count. Welcome letters, office norms, expense rules and benefits summaries take up an enormous amount of space and carry almost no legal weight. That is where AI earns its keep, and it earns it honestly.
The red pile is short. Six or seven headings. Which means “get a lawyer to review the handbook” stops being a vague, expensive, open-ended request and becomes a specific, quotable, one-afternoon piece of work. If you have ever tried to get budget approval for legal review, that reframing is worth more than any prompt in this article.
I want to be specific about why these sections are dangerous, because “ask a lawyer” is useless advice if you cannot explain to your CFO what you are protecting against.
This is the one almost nobody expects. Under the NLRB’s Stericycle standard, adopted in 2023, a workplace rule is presumptively unlawful if it has a reasonable tendency to chill employees from exercising their rights. The employer then has to prove the rule serves a legitimate business interest that could not be served by narrower language.[6] This applies to non-union private employers too, which is the part that catches people out.
Now think about what an AI drafts when you ask for a professional conduct policy. It writes something tidy and broad: employees should maintain a positive working environment, avoid discussing confidential matters including compensation, refrain from disparaging the company online. Every one of those clauses is exactly the kind of overbroad rule Stericycle was built to catch. The AI is not being reckless. It is producing the average of every handbook it has ever seen, and a lot of those handbooks are wrong.
Most people assume a boilerplate “this handbook is not a contract” line does the job. In Hall v. City of Plainview, the Minnesota Supreme Court disagreed, holding that a general disclaimer did not prevent the PTO policy from creating a contractual right to 1,778 hours of accrued time. Individual portions of a handbook can create contractual rights even where other portions do not.[7] The practical lesson from that case is that disclaimers need to be specific and consistent with every clause around them, which is precisely the kind of cross-document reasoning AI is worst at.
Accrual, carryover and payout-on-termination language is where handbook drafting turns into real money. It varies by state, it interacts with attendance policies in non-obvious ways, and getting it wrong is expensive. The Department of Labor’s Wage and Hour Division recovered more than $259 million in back wages for nearly 177,000 workers in fiscal year 2025, the highest figure since 2019.[8]
Exempt versus non-exempt, meal and rest breaks, overtime, final pay. Jurisdiction-specific, heavily litigated, and an area where a confident-sounding paragraph is worse than no paragraph at all.
“Progressive discipline” language is the classic accidental-contract trap: describe a four-step process and you may have promised a four-step process. Arbitration agreements and class-action waivers vary sharply in enforceability by state and by how consent was obtained.
This is the sequence I use with HR teams. It takes a focused day for a first draft, plus whatever your lawyer needs. The order matters, because most of the pain comes from people starting at step three.
The five-step sequence described in this section. Steps 1, 2 and 5 involve no AI drafting at all.
Step one is an inventory, not a draft. Before you write a word, find every policy that already exists: the old handbook, the offer letter template, the intranet page nobody updated, the email from 2023 that became the remote work rule. Half the value of a handbook project is discovering that you have three contradictory versions of the same policy. AI cannot find those for you. Walking around and asking can.
Step two is the sort. Twenty minutes, three highlighters, as above.
Step three is drafting the green pile. Feed the AI your existing material and ask it to rewrite, not invent. This distinction is the single biggest quality lever in the whole process. “Rewrite our expense policy at an eighth-grade reading level, keeping every threshold and approval step exactly as written” produces something usable. “Write an expense policy” produces something generic that quietly contradicts what your finance team actually does.
Step four is the gap check, and it is the most underrated step. Paste your table of contents into the AI and ask what a handbook for a company of your size, in your states, in your industry, would typically contain that yours does not. Treat the output as a list of questions for your lawyer, never as a list of sections to go and write. It is very good at spotting that you have no lactation accommodation policy. It is not qualified to write one.
Step five is legal review of the red pile plus state addenda. If you employ people in more than one state, the practitioner rule is that your handbook must be written for the most demanding jurisdiction you operate in, or carry state-specific supplements. A single national handbook with no state addendum is the classic multi-state failure. Related workflows like writing SOPs with AI follow the same logic, though the stakes there are lower.
Every prompt below assumes you have pasted in real source material. A prompt without your content attached is a request for fiction.
“Below is our current [policy name]. Rewrite it for clarity at roughly an eighth-grade reading level. Keep every number, threshold, deadline and approval step exactly as written. Do not add any new rule, entitlement or obligation. If anything in the original is ambiguous, do not resolve it: list it separately at the end under ‘Needs a decision’.”
That last clause is the important one. Left to itself, AI resolves ambiguity by inventing a plausible answer. Forcing it to surface ambiguity instead of smoothing it over turns a risk into a to-do list.
“Here is the table of contents of our employee handbook. We are a [size] company in [industry] with employees in [states]. List policies commonly found in handbooks like ours that are missing here. For each one, say in one line why it typically appears. Do not draft any of them. Flag which ones are usually legally required somewhere in our states so I can raise them with counsel.”
“Read this handbook draft as an employment lawyer preparing to cross-examine us. Find every place where two sections could be read as contradicting each other, and every place where a rule promises a process we would then be held to. Quote the exact sentences. Do not rewrite anything.”
This one is genuinely excellent and costs you nothing. It is also the prompt most likely to make you slightly sick when you run it on a handbook you inherited.
“You are a new employee on your first day, with no HR background. Read this section and tell me: what am I allowed to do, what am I not allowed to do, and what do I still not know?”
Role-inversion prompts like this consistently outperform “make this clearer,” because they force the model to check comprehension rather than polish sentences. If you want more on why that works, our guide to using AI for employee onboarding covers the same technique applied to first-week materials.
Here is the irony at the centre of this whole project. The section your handbook most urgently needs is the one about AI, and it is also the section most likely to be badly written, because everyone is drafting it in a hurry.
SHRM’s data makes this uncomfortably clear. About half of organisations using or piloting AI have a policy governing workforce use of it. Of those that do, only around a quarter feel their policy is clear and future-proof. More than half say theirs is too restrictive and tied too tightly to whichever tools happen to exist right now.[3]
That last finding tells you exactly how to write yours: principle-based, not tool-based. A policy that names ChatGPT and Copilot will be wrong within a year. A policy that describes categories of use and categories of data will survive.
What a workable AI section covers:
And one warning that almost nobody mentions: an AI policy can itself be an unlawful work rule. If your policy bans employees from discussing the company’s AI tools, or from posting about workplace AI online, you have written exactly the kind of overbroad restriction the Stericycle standard targets.[6] An AI-drafted AI policy is a very efficient way to create that problem. Put this section in the amber pile at minimum, and in the red pile if you are in a unionised or union-adjacent environment.
Four failure modes, in the order I see them.
Generating instead of rewriting. The team asks for a handbook, gets 40 pages of confident prose, and adopts most of it because it reads well. Six months later somebody asks where the bereavement leave entitlement of five days came from and the honest answer is that a language model picked a plausible number. Every entitlement in your handbook should trace to a decision a human made.
Treating a national template as compliance. The handbook reads beautifully and contains nothing that satisfies your California, New York or Massachusetts obligations. Brightmine notes that California alone often has more than 50 employment law changes taking effect in a single year.[1] One document, no addenda, employees in six states is not a handbook. It is a liability with a cover page.
Publishing and forgetting. You have just done the hardest version of this work. Set a recurring quarterly calendar entry to review it, and put the review date in the document itself. Otherwise you are back in the 62% within a year.[1]
Skipping the acknowledgement trail. A handbook nobody has confirmed reading is difficult to rely on. Whatever your distribution method, keep the record.
Honestly, the thing I would most like you to take from this is the sort, not the prompts. Prompts change every few months. The realisation that your handbook splits into a large harmless pile and a short dangerous one is what makes the whole project affordable, reviewable and finishable. Print your table of contents this week and mark it up. That is the entire first step, and it costs nothing.
If you want to see the same consequence-based sorting applied elsewhere in the people function, our guides on AI for performance reviews and AI for engagement surveys use the same logic on higher-stakes documents.
No, and the failure mode is subtle rather than obvious. AI produces a fluent handbook that averages every handbook it has seen, which means it will confidently include entitlements you never agreed to and conduct rules that are drafted too broadly to be lawful. Use it to rewrite policies you already have and to audit for gaps and contradictions. Keep a human decision behind every number and a lawyer on the conduct, at-will, leave, pay and arbitration sections.
Policy text on its own is usually fine, because it contains no personal data. What is not fine is pasting employee names, salary figures, disciplinary records, health information or anything from a live investigation. Check your organisation’s own rules first: Littler found only 54% of employers restrict what information can be entered into AI tools, so there may simply be no guidance yet.[4] If you are on a paid business tier with training disabled, your exposure is lower, but the rule about people data still holds.
No law says a lawyer must draft any of it. The practical list, based on where liability concentrates, is: the at-will statement and disclaimer, conduct and confidentiality rules, all leave policies, pay and classification, discipline and termination procedures, arbitration clauses, and the anti-harassment complaint procedure. Conduct rules are the least obvious and most important, because under the NLRB’s Stericycle standard an overbroad rule is presumptively unlawful even at a non-union employer.[6]
If your employees use AI at work, yes, and they almost certainly do: Gallup found 45% of US employees use it at least occasionally while only 37% said their employer had implemented it.[2] Write it around principles rather than named tools. SHRM found that more than half of existing AI policies are already considered too restrictive and too tied to current tools, which is what happens when you name products instead of describing categories of use and data.[3]
More often than most organisations manage. Brightmine found 62% review theirs about once a year or less, and fewer than half of HR respondents felt very confident their handbook reflected current law.[1] A quarterly light review with a full annual pass is realistic for most teams. Put the review date inside the document so the next person inherits the cadence rather than the guesswork.
This guide draws on published research from SHRM, Littler Mendelson, Brightmine and Gallup, plus primary sources from the NLRB, EEOC and US Department of Labor. Every statistic was verified against the original publication before inclusion. Nothing here is legal advice: it is a workflow for deciding which parts of your handbook need a lawyer and which do not.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.