You can write a usable AI use policy yourself. What you cannot do safely is write it the way the template pages tell you to, because the risk in a do-it-yourself policy is not the clause you left out, it is the sentence you made up. Most guides to AI policy clauses hand a non-lawyer a section list and a closing line about having counsel review the result. None says which of your own sentences will be read back to you later by a client’s security questionnaire, an auditor, or your insurer.
These eight rules are about that. They assume you are an operations lead, a chief of staff, or a head of people who got this as a second job, with no general counsel down the hall and a budget for an hour of help, not a month.
Two things people mean by “AI policy clauses”
The phrase splits two ways, and picking the wrong half costs a week.
Contract clauses are what you negotiate with a vendor or a client: no training on your data, who owns the output, who indemnifies whom when a generated asset infringes something, and what happens when the vendor swaps the model under your feet. For that, start from a published set rather than drafting cold. Bonterms’ standard forms include a free AI addendum, the European Commission publishes model clauses for AI procurement, and the UK Society for Computers and Law offers a dual-sided set.
Internal policy clauses are the rules your own staff follow. That is the rest of this article, and the version most readers need first.
One number frames your situation. In Littler’s 2026 Annual Employer Survey, fielded in March 2026 among 303 US in-house lawyers, legal chiefs, and HR leaders, 68% reported a formal policy on workplace AI use, up from 38% a year earlier. Only 55% had a review route for new tools, 54% limited what may be entered, and 25% ran risk-based training on legal and ethical use. The document is now the norm. The machinery behind it is not.
Rule 1: Decide whether you are writing instruction or evidence
Every AI policy does two jobs that pull against each other.
As instruction, it tells a marketing coordinator at 4pm on a Friday whether she can paste a client’s draft press release into a chat assistant. That version is short, in the second person, and names things: products, plan tiers, data types, people.
As evidence, the same document is what you hand a client who asks how you govern AI, or what an auditor reads before testing whether you do what it says. That version wants framework vocabulary and a review history.
Write the instruction version first and keep it to two pages. Then choose which framework language you will stand behind, because the evidence version commits you to proof. A policy quoting NIST or ISO vocabulary invites the question of who was trained and when, and a quarter of employers in Littler’s set can answer it.
Rule 2: Read what you already signed before you write a word
This is the highest-value hour in the exercise, and the one nobody schedules. Four documents you already have govern part of it.
| Document | What it already says about AI | The contradiction to avoid |
|---|---|---|
| Client NDAs and master agreements | Confidentiality duties covering material pasted into any outside system, usually without naming AI | Your policy permits a workflow an NDA forbids |
| Data processing agreement and privacy notice | Which processors may receive personal data, and what you told customers | A new AI vendor appears in practice but not in the notice you published |
| Employee handbook | Monitoring, discipline, device use, confidentiality | Two documents set different consequences for one act |
| Vendor terms for the tools in use | Whether your plan allows business use and keeps your data out of training | The policy approves a product your account is not on |
Read those four, then write. A clause that contradicts a confidentiality term you signed in 2023 creates an argument, not a control. We went through that side of it in what your NDAs already say about pasting client material into AI: most firms draft new rules on top of old rules they never reread.
Rule 3: Never write a promise you cannot show
Non-lawyers write absolutes because absolutes sound firm. An absolute is a standard you will be measured against by someone who did not write it.
| What people write | What it commits you to | What to write instead |
|---|---|---|
| “All AI output is reviewed before use” | Proving review happened, on every output, on request | “Work in these categories is checked by a named reviewer before it leaves the company” |
| “We never enter customer data into AI tools” | One pasted spreadsheet makes the statement false | “Customer records may be entered only in [tool] on [plan], never in a personal account” |
| “The company ensures all staff are trained on AI risks” | Evidence that every person was trained, not that you offered it | “Every new joiner completes the AI session in their first month; attendance is recorded” |
| “AI is never used for hiring decisions” | A recruiter’s résumé summary breaks it | “AI must not screen, rank, or score candidates; drafting a job ad is allowed” |
The right-hand column follows one pattern: name the actor, the scope, and the artifact. Statements about the world are promises. Statements about who does what are rules.
Watch the verbs too. “Ensure,” “guarantee,” and “all” are the words most likely to turn an internal memo into a claim about your controls that a customer relies on.
Rule 4: Build the data rule from your own last two weeks of work
Template data sections classify information in the abstract: confidential, sensitive, restricted, public. Staff do not think in those labels, and the person least able to apply them is the one holding a file at speed.
Do it the other way around. Open the last two weeks of real work and sort your ten most common artifacts by name.
A professional-services firm ends up with a list a person can hold in their head: client work in progress, anything naming a person, pricing and unreleased financials, credentials and keys, HR matters. Five types from your own files beat twenty tiers written for a company that does not exist.
Then write the rule as a sentence about accounts rather than qualities. The trap worth knowing is that a product name is not a data commitment; the plan tier is. We take that apart in where policy templates break, and it is the most common defect in a policy drafted from a download.
Rule 5: Put the facts that change monthly in a dated annex
Three things in an AI policy go stale fast: the approved-tool list, the plan tiers, and anything naming a named assistant or feature. The rest, the data rule, the review duty, the escalation route, holds for a year or more.
Split them. The signed policy carries the stable rules. A one-page annex carries the tool list under its own version date, and the owner updates it without reopening the document everyone signed. A policy still listing a product you dropped tells every reader nobody maintains it.
Rule 6: Name one owner and one route, not a committee
Littler put the share of employers with a standing AI oversight committee at 45%. At 50 to 500 people with no compliance function, a committee is usually what stops a policy from working, because nobody asks a committee a question on a Friday afternoon.
Name a person. Give them one route for the two questions that come up most: can I use this tool, and can I put this material into it. State the answer time, in days, in the document.
The exception route is what templates leave blank and what decides whether you have a control or a suggestion. With no stated way to get a yes, people take the yes themselves and stop telling you.
Rule 7: Say what you log, in plain words
A policy silent on monitoring reads as a trap, and staff treat it as one. Say which admin logs exist, who can see them, what triggers a look, and whether any of it reaches performance review.
If the honest answer is that you have a console nobody opens, write that. Overstating your visibility is the mirror image of Rule 3 and fails the same way.
Writing it down also keeps you out of the enforcement posture that destroys what you want, as enforcing an AI policy without surveillance sets out.
Rule 8: Write the consequence your handbook can deliver
The consequences section is where a non-lawyer is most likely to create a real employment problem, because discipline is already governed by documents you did not write.
Three guardrails. Match the language your handbook already uses. Keep the first-offense response proportionate, since most breaches will be a well-meaning person pasting the wrong file. Route anything touching pay, standing, or firing through whoever owns HR.
One clause belongs here and gets left out: the no-blame reporting line. Tell people what to do in the first hour after a mistake, and say plainly that reporting it fast is the behavior you want. A policy that makes self-reporting dangerous buys silence, the one outcome you cannot manage around.
Where a non-lawyer should stop
Four triggers mean you buy an hour of counsel instead of reading another template.
- Regulated data. Health records, financial account data, children’s data, or anything under a sector rule.
- Client contracts that reach your tools. Flow-down duties, audit rights, or a client who must approve subprocessors.
- Anything near an employment decision. AI in hiring, promotion, discipline, or firing carries bias and state-law exposure that no drafting fixes.
- A disclosure duty. Any duty to tell a client, a regulator, or a court that AI was involved in work product.
Absent those four, a careful two-page document plus the annex is a sound control, and an employment lawyer’s review at normal hourly rates is a few hundred to a couple of thousand dollars of market cost rather than a project. What drafting and rollout cost end to end is covered in what it costs to get an AI use policy written and rolled out, and the choice of who to hire is in lawyer, consultant, or enablement partner.
Pay for the hour that reviews your sentences, not the one that writes a draft you could do yourself.
The document is the cheap half
Littler’s split says it plainly: 68% have the policy, 25% have risk-based training. Writing was never the bottleneck.
Both frameworks your buyers and insurers read point the same way. EU AI Act Article 4 has applied since 2 February 2025 and, after the AI Omnibus amendment, asks providers and deployers to “take measures to support the development of AI literacy” among their staff and anyone operating AI systems on their behalf. It draws no line by company size. NIST’s AI Risk Management Framework, the reference auditors reach for, asks in its GOVERN function that staff and partners get AI risk training suited to their duties.
Neither asks for a better document. Both ask whether your people can apply judgment to the case your clauses did not anticipate, which is the argument in the gap between policy and practice.
That gap is why AI enablement, rather than another round of drafting, is where the remaining risk sits, and why buying licenses is not the same as building capability. A ninety-minute applied session on your own documents, twice a year, does more for your exposure than a third revision of the policy. Sessions like that run roughly $2,000 to $15,000 in the open market.
Write the two pages. Then teach them.
Frequently asked questions
Can a non-lawyer write an AI use policy?
Yes, for most companies with no regulated data and no flow-down duties in client contracts. The parts that matter most, the tool list, the data types, the owner, the escalation route, are facts about your company no lawyer can supply. Buy an hour of counsel to review what you wrote rather than hours to draft it, and send it to an employment lawyer if the consequences section has teeth.
Does “AI policy clauses” mean contract clauses or internal policy clauses?
Both, and which one you mean decides where you start. Contract clauses cover training rights, output ownership, indemnity, and model changes, and free sets from Bonterms, the European Commission, and the Society for Computers and Law save you drafting them cold. Internal clauses must be written in house, because they name your tools, your data, and your people.
Which sentences in an AI policy create legal exposure?
Absolutes and promises about outcomes. “All output is reviewed,” “we never enter customer data,” and “the company ensures all staff are trained” each commit you to proof you cannot produce, and each becomes a claim about your controls once a client questionnaire or an auditor sees it. Rewrite them as rules about who does what, in which tool, with which material.
Which documents do we have to read before drafting?
Four: your client NDAs and master agreements, your data processing agreement and privacy notice, the employee handbook, and the terms of the AI plans you are on. Each already governs part of this. The common failure is a new rule that contradicts a confidentiality or monitoring term signed years ago.
How long should the policy be, and what belongs in an annex?
Two pages for the policy, one for the annex. The policy carries the stable rules: scope, data types, the check before work goes out, the owner, the escalation route, reporting, and consequences. The annex carries approved tools with plan tiers and account routes under its own version date, so a tool change does not reopen a signed document.
Do we have to tell staff what we log?
You should, and in most cases you will want to. A policy silent on monitoring gets read as a trap and quietly ignored, while a plain statement of which logs exist and who sees them raises compliance with no new tooling. Check your handbook’s monitoring language before you write this clause.
Should the policy name a committee or a person?
A person, below a few hundred employees. Littler found 45% of employers have a designated AI oversight committee, but a committee cannot answer a question on a Friday afternoon, and unanswered questions are how staff learn to route around the document. Name one owner, one route, one turnaround time.
Is a signed policy enough for a client security questionnaire or a SOC 2 audit?
It is the first artifact anyone asks for and rarely the last. Expect follow-up on how new tools get approved, which logs exist, how access is cut off, and whether the people in scope were taught anything. The policy proves intent; the approval record and the attendance list prove it works.
Key takeaways
- The risk in a do-it-yourself AI policy is an invented sentence, not a missing clause.
- Read your NDAs, your data processing agreement and privacy notice, your handbook, and your real plan terms first.
- Replace every absolute with a rule naming an actor, a tool, and a material. “Ensure,” “guarantee,” and “all” are the words to strike.
- Draw data categories from the last two weeks of real work, not a template’s sensitivity tiers.
- Keep the signed policy to two pages; move tools and plan tiers into a dated annex.
- One named owner with a stated answer time beats a committee at this size.
- Say what you log. Silence reads as a trap; overstatement fails at the first incident.
- Buy counsel for regulated data, flow-down duties, employment decisions, or a disclosure duty. Not for the first draft.
If you want a second pair of eyes on your draft, or a view of what an applied session covers once it is signed, book a free AI-readiness call. We will read it against how your team works and say which sentences will not survive a client questionnaire.
Dirk Jan van Veen, PhD