A free AI policy template will get you a signed document by Thursday. It will not get you the two things the document exists to produce: a rule someone can apply at 4pm on a deadline, and a record that answers the question an auditor or regulator will put to you.
Templates are not the problem. The problem is that six clauses in a standard fifteen-section template are wrong the moment you inherit them, the advice attached to every template is “customize it,” and nobody tells you which six. SHRM’s 2026 State of AI in HR research found that among organizations with a policy, only about one in four believe it is clear and future-proof, and more than half describe their policy as too restrictive and too tightly tied to specific current tools. That is not a complaint about templates in the abstract. It is a report on which clauses failed.
What a template genuinely gives you
Concede the case for templates first. It is strong.
A credible template gives you completeness. You will not forget the incident-reporting clause, the acknowledgment log, or the review cadence, and those are exactly the clauses a first-time drafter omits. It also gives you vocabulary that lines up with the frameworks your buyers and auditors already read.
And it gives you speed, which matters more than drafting purists admit. Generative AI use at work among US adults aged 18 to 64 rose from 33.3% to 37.4% in the year to August 2025, according to the Federal Reserve Bank of St. Louis. Your staff are not waiting for your legal review to finish.
The better sources are free and non-commercial. SHRM publishes a customizable AI governance policy template for employers, and the NIST AI Risk Management Framework, published 26 January 2023, supplies a GOVERN function whose subcategories map almost clause-for-clause onto a policy outline. Neither is selling you monitoring software, which cannot be said for most of the template pages giving one away.
The honest answer to template versus drafted is not a side. It is a line drawn through the document.
The clause map: what survives inheritance
Every template presents its sections as equally fillable. They are not. Three tiers behave differently.
| Tier | Clauses | What inheritance does to them |
|---|---|---|
| Survives mostly intact | Purpose, scope, definitions, intellectual-property language, consequences for violation, acknowledgment mechanics | Generic is fine here. These clauses describe a posture, not a fact about your company. Swap the entity name and move on. |
| Needs translation | Human oversight and review, prohibited uses, incident reporting, review cadence, training commitment | The structure transfers; the thresholds do not. “Review AI output before external use” is true everywhere and actionable nowhere until you name which outputs, reviewed by whom, against what. |
| Wrong by construction | Approved tools and plan tiers, data classification tiers, named policy owner, escalation path, exception and approval workflow, monitoring disclosure | These are facts about your organization. A document written for a hypothetical company states them incorrectly by definition, and an incorrect statement here is worse than a blank. |
The third tier is where the SHRM finding lands. A policy that is “too tightly tied to specific current tools” is a policy whose tool list was inherited rather than inventoried, and the same drafting shortcut produces the other five failures in that row.
Two of them get skipped even by careful drafters. The exception workflow decides whether your policy is a control or a suggestion: with no route to approve a new tool and no stated turnaround on that route, people route around the policy instead. The monitoring disclosure decides whether staff trust the document at all. A policy silent on what is logged, and on whether those logs feed performance review, reads as a trap and gets treated as one.
We have argued the full version of this in the gap between what a policy says and what a team does. The clause map is the drafting-desk version of the same point.
Your approved-tools list names a product when it should name a plan
This single failure undoes more policies than any other, and almost no template page mentions it.
Templates list approved tools by product name: ChatGPT, Claude, Gemini, Microsoft Copilot. A product name is not a data-handling commitment. The plan tier is. Anthropic’s published consumer documentation, dated 16 March 2026, states that chats on consumer Claude plans are used for model improvement only where the user has opted in, and points commercial products such as Claude for Work and the API to a separate policy. The other major vendors draw a comparable line between their consumer and their business and enterprise agreements.
Read that against a clause approving “Claude” or “ChatGPT” and the problem is obvious. The person reading your policy has a personal account. Your policy approved the product, and procurement never bought the tier the policy assumed.
Nobody lied, nobody broke a rule, and your client material is now governed by consumer terms.
The fix is a sentence, not a project. Name the tool, the plan tier, and the account route in one row: Claude, Team plan, company SSO only; personal accounts are not approved for any company material. Then put the list in an annex with its own version date, so a quarterly tool change does not reopen the signed policy.
Three regimes want three different artifacts
“Aligned with NIST and the EU AI Act” appears on most template pages and means almost nothing, because the three things buyers most often conflate ask for three different deliverables.
| What you are answering to | What it asks for | What you must be able to hand over |
|---|---|---|
| EU AI Act, Article 4 | Measures supporting AI literacy among staff and anyone operating AI systems on your behalf, scaled to role and context. Applicable since 2 February 2025 | An internal record of the measures you took. No certificate exists or is required |
| NIST AI RMF, GOVERN 1.1 and 1.2 | Legal and regulatory requirements understood, managed and documented; trustworthy-AI characteristics integrated into organizational policies and procedures | The policy, mapped to named roles, plus evidence the mapping is maintained. Voluntary guidance, and not something you can be certified against |
| ISO/IEC 42001:2023 | A functioning AI management system under clauses 4 to 10, including internal audit and management review | An auditable system, independently certifiable. The only one of the three a template cannot approximate |
A drafted policy becomes necessary when one of these stops being a reference and starts being a commitment. Certify to ISO/IEC 42001 and no template survives the auditor, because the standard asks for a management system rather than a document. For a US employer with no certification ambition, a well-customized template plus a counsel read is proportionate and defensible. Who signs the result and answers for it is a question we take apart in what leaders and board directors need to know about AI governance, and the per-regime artifact question in more depth in the regulated-team view of compliance training options.
Why a correctly customized policy still changes nothing
Here is the uncomfortable part, and it applies equally to the $29 template and the $4,000 drafted version.
Your policy is a set of decisions written down. The decisions themselves get made at a desk, under deadline, by somebody who read the policy once, eight weeks ago, during onboarding.
The client deck is due at five. The contract clause needs summarizing. The question in front of them is not “what does section 5.2 of our acceptable use policy say.” It is “can I paste this.” Those are two different cognitive events, and a document only answers the first.
So a rollout made of an all-hands slide and an attestation checkbox produces an immaculate record and no behavior change. The record proves the document was received. It proves nothing about whether anyone can apply it to a specific paragraph of a specific contract on a specific Tuesday. Security awareness programs learned this a decade ago and most AI policy rollouts are repeating it from scratch, a pattern we unpack in what AI risk training leaves out that security awareness also missed.
What closes the gap is practice on your own material. Put eight people in a room with real documents from their own desks, ten of them deliberately borderline, and have them decide out loud: in, out, or ask.
Three things happen that no document achieves. People calibrate against each other, so the rule converges instead of fragmenting. The ambiguous cases surface while the policy is still editable, which usually sends two or three clauses back for rewriting. And the decision becomes something they have done before, which is the only reliable predictor of what they will do under time pressure.
That is the argument this whole program rests on, and it is the same argument as buying licenses is not the same as building capability. A policy is a license to a rule. Enablement is whether anyone can use it.
What each layer costs
Three cost lines, and they are not substitutes for one another. Market ranges, not quotes.
The document. Free to roughly $50. Credible free templates exist from SHRM, NIST’s framework material, and several security vendors; paid versions mostly add a Word file and a generator. Nothing here is expensive enough to agonize over.
Counsel review of your customized draft. A few hours of employment or technology counsel, on a narrow scope: your jurisdiction and sector exposure, your monitoring disclosure, and anything touching hiring, lending, insurance, or benefits decisions. Send a customized draft, not a blank template, or you are paying hourly rates for a tool inventory.
Facilitated practice. Workshop engagements sit in a $2,000 to $15,000 market range depending on scope and headcount. Most teams skip this layer, and it is the only one that touches what people do rather than what they signed.
A budget spent entirely on drafting and review is a budget spent entirely on paper.
The order that works
Sequence matters more than source, and most teams run it backwards.
- Inventory before you draft. Pull the real list from expense reports, your identity provider, and a no-blame amnesty week. You cannot approve tools you have not found.
- Decide the six wrong-by-construction clauses. Tool list with plan tiers, data tiers, named owner, escalation path, exception route with a stated turnaround, monitoring disclosure. Decisions, not drafting.
- Now pick a template and pour those decisions into it. This is where a template earns its keep.
- Counsel reads the customized draft, on the narrow scope above.
- Practice before you attest. Run the borderline-document session before the acknowledgment goes out, and expect to rewrite two or three clauses because of it.
- Re-measure at 30 days. Ask three people in different functions what never goes into an assistant.
Step 5 is what gets cut when the calendar tightens, and cutting it turns the exercise into paperwork.
If you want a second opinion on which of these layers you are missing before you commission anything, book a free AI-readiness call. We run hands-on fluency sessions against a team’s own documents and daily work, and we will tell you plainly when a free template and two hours of counsel time is the right answer.
Frequently asked questions
Is an AI use policy legally required?
It depends on your footprint. In the EU, Article 4 of the AI Act has required providers and deployers to support AI literacy among staff and anyone operating AI systems on their behalf since 2 February 2025, which in practice means documented measures. In the US there is no single federal mandate, but sector rules and state statutes reach the same place indirectly, and SOC 2 examiners increasingly ask for an AI-specific acceptable-use policy with version history. Take the specific question to counsel.
Can a template be enough on its own?
For a small, single-jurisdiction company with no certification ambition and low-risk use, a well-customized template plus a counsel read is proportionate. What is never enough is a template with only the company name changed. If you could remove your company name and the document would fit any organization, you have a document about AI rather than a policy for your company.
Which clauses should never be inherited from a template?
Six: the approved-tools list with plan tiers, your data classification tiers, the named policy owner, the escalation path, the exception and approval workflow with its stated turnaround, and the monitoring disclosure. Each is a fact about your organization, so an inherited version states it wrongly.
Can we extend our existing acceptable use policy instead?
Partly. Device use, credentials, and confidentiality carry over. What does not is the material question: which categories of information may be entered into a third-party model, under which account, and what happens to it afterward. That needs new text, whether in a new document or an appendix to the old one.
Do we need a lawyer to review it?
Yes, and keep the scope tight: jurisdiction and sector exposure, monitoring and employee-privacy disclosure, and anything touching employment decisions. Send a customized draft. Paying hourly rates to fill in bracketed placeholders is the most expensive way to run a tool inventory.
How often does an AI policy need rewriting?
The policy body, roughly annually. The tool annex, quarterly, which is why it belongs in an annex with its own version date rather than a clause in the signed document. SHRM’s 2026 finding that over half of existing policies are too tightly bound to current tools describes a structural problem with where the tool list lives, not a drafting failure.
What is the difference between an AI use policy, NIST AI RMF, and ISO/IEC 42001?
The policy is the document telling staff what they may and may not do. NIST AI RMF is voluntary guidance that supplies a governance structure to organize that policy around, and you cannot be certified against it. ISO/IEC 42001:2023 is a certifiable management-system standard covering clauses 4 to 10, including internal audit and management review. One is a rule set, one is a reference, one is an auditable system.
How do we know whether the policy changed anything?
Four checks, none of which needs new tooling. Ask three people in different functions what never goes into an assistant and see whether the answers match. Compare the sanctioned-account roster against what finance reimburses. Look for a verification trace on real external output from last month. And watch whether questions and self-reported mistakes went up, which looks like bad news and is the healthiest signal available.
Should we block unapproved tools outright?
Blocking without approving is the pattern that reliably fails. If the only compliant option is to do less work than last week, the policy will be ignored quietly by the people you most need on side. Approve at least one good tool per genuine use case, make the exception route fast and visible, and treat enforcement as the backstop rather than the strategy.
Key takeaways
- Templates are strong on completeness, vocabulary, and speed, and weak on exactly the clauses that describe your company. That is where policies fail in practice.
- Six clauses are wrong by construction when inherited: tool list with plan tiers, data tiers, named owner, escalation path, exception workflow, monitoring disclosure. Decide those before you pick a template.
- Approve plan tiers, not product names. Anthropic’s published consumer documentation separates consumer plans, where chats feed model improvement on opt-in, from commercial products under a separate policy, and a clause approving “Claude” does not distinguish the two.
- Three regimes want three artifacts: Article 4 wants an internal record and no certificate, NIST AI RMF GOVERN wants a documented policy mapped to roles, ISO/IEC 42001 wants an auditable management system. Only the last rules templates out.
- A correctly customized policy still changes nothing by itself. Its decisions get made at a desk under deadline, and practice on your own borderline documents is what makes them repeatable.
- Spend proportionally: the document is near free, counsel review is a few hours, facilitated sessions sit in a $2,000 to $15,000 market range. Most budgets fund only the first two.
Dirk Jan van Veen, PhD
