Generative AI Acceptable Use Policy: How to Write One Employees Follow
A generative AI policy employees actually follow needs scope, data rules, prohibited uses and review duties — not a ban. Here is the structure and the template.
A generative ai acceptable use policy only works if people read it once and can apply it at 4pm on a Thursday. Most drafts fail in one of two directions: a blanket ban that guarantees shadow usage, or a principles document so abstract that nobody can tell whether pasting a customer list into a chatbot is allowed. The fix is structural — a short ai usage policy workplace guidelines document organised around the decisions employees actually make: which tools, what data, what they may not automate, and who reviews output before it leaves the building. This guide gives you the sections, the drafting rules, and the rules for using ai at work policy questions to answer explicitly.
Scope: be precise about what the policy covers
Open by naming the territory, because “AI” is not a policyable category:
- Covered: publicly available chatbots, image generators, AI features embedded in business tools (email drafting, meeting summaries, code assistants), and any AI tool processing company data or producing work products.
- Out of scope: deterministic automation and rule-based macros (say so — employees already know their macros are fine), and personal-use AI on personal devices with no company data — explicitly excluded, because employees will assume the worst otherwise.
- Approved tool list: attach or link the current list with owner and review date. This single artefact does more work than any paragraph: the natural question changes from “is AI allowed?” to “is this tool on the list?” — a question with an answer.
If your policy has no approved list, you do not yet have a policy — you have an opinion. Start the list with the five tools already in use (they are on the list whether you like it or not; adding them formally is not endorsement, it is visibility).
Data rules: the three-tier prompt test
The data paragraph decides whether the policy is real. Teach one test employees can run without asking anyone:
- Public tier — public marketing copy, published specifications, non-confidential text. Prompt freely in any approved tool.
- Internal tier — internal-only information: strategies, roadmaps, internal metrics, unreleased plans. Use only enterprise-tier tools with contractual no-training and retention commitments; never free consumer tiers.
- Restricted tier — personal data, customer records, employee data, secrets, source code under licence, anything regulated. Not into general-purpose AI tools at all, under any circumstances; use only purpose-built systems with a lawful basis and controls.
Name the three tiers in the policy itself with two or three concrete examples each — “customer export” and “the Q3 board deck” travel further than abstractions. Add the technical backstop: approved enterprise tools with SSO and admin controls are the enforcement; the tiers are the judgement call in between.
Prohibited uses: short, specific, defensible
Keep the ban list under ten items so it can be remembered, and make each item a scenario rather than a category:
- Making final decisions about people (hiring, firing, credit, disciplinary) on AI output without documented human review.
- Pasting restricted-tier data into any general-purpose AI tool.
- Publishing AI-generated content externally — claims, legal statements, regulated communications — without human verification and the applicable disclosure.
- Impersonating colleagues or customers with synthetic voice or images; creating deepfakes of any person.
- Circumventing controls: re-entering restricted data “as a story”, using unapproved tools through personal accounts to reach the same result.
- Relying on AI output as legal, medical, financial or safety advice without qualified human review.
- Coding practices that violate licence terms — including feeding licensed code into tools whose output ownership or training terms conflict with your licences.
Every prohibition should name the behaviour and the reason in one line. Employees follow rules they can explain to a colleague.
The policy skeleton: sections and what goes in each
Keep the document to one page of rules plus the reference material. The skeleton, with the drafting mistake each section invites:
| Section | What it must contain | The mistake to avoid |
|---|---|---|
| Scope | Covered tools/features, explicit exclusions, link to the approved tool list | Defining scope as “AI” and leaving macros, plugins and analytics to argument |
| Data tiers | Three tiers with two real company examples each | Using “confidential information” with no test employees can run unaided |
| Prohibited uses | Under ten scenarios, each with its reason | Category bans (“no AI for content”) that teams reinterpret by mid-week |
| Review & attribution | Who checks, who approves, when disclosure is mandatory | Requiring “human review” without naming where it sits in the workflow |
| Incidents | One-hour blameless reporting route for restricted data | No route at all, which guarantees silent cleanup attempts |
| Enforcement | Ladder from coaching to disciplinary route | Maximum-severity threats that make every breach a hidden breach |
| Governance | Owner, approved-list review cadence, version | No review date, so the tool list quietly rots |
Fill it top to bottom in a single drafting session, then cut a third of the words. Every sentence that does not change an employee’s behaviour is overhead.
Review duties and attribution: who owns the output
Every policy needs one unambiguous rule: the human who sends it owns it. AI output enters your organisation as a draft, never as a deliverable. Structure the duty by stakes:
- External content (marketing claims, public statements): human verification of every factual claim, plus the applicable Art. 50 disclosure when content is synthetic and publicly published.
- Operational content (client emails, reports): author reviews for accuracy and tone; a named approver for anything contractual.
- Code: licence and security review before merge — AI-generated code carries the same obligations as any other contribution.
- Analysis and summaries: check against source documents; summarisation errors compound quietly.
Add attribution guidance calibrated to your culture: when disclosure is legally required (synthetic images, AI-assisted text published to inform the public on matters of public interest), it is mandatory; everywhere else, encourage honest labelling rather than policing it. Employees should never be penalised for revealing they used an approved tool — hidden use is the risk, and penalising disclosure guarantees it.
Rollout: policy, training, and the enforcement ladder
A policy lands through three mechanisms, in order:
- Acknowledgement. One page, read-and-acknowledge, stored with date — your evidence of measures under the AI Act’s AI literacy obligation and reasonable notice for any later disciplinary reliance.
- Ninety-minute enablement. Live examples: a good prompt with internal-tier data in the approved tool, a bad prompt with restricted data, a deepfake request refused. Include the tier test as an exercise. Attendance recorded — this session doubles as Article 4 literacy evidence for generative tools.
- Graduated response to breaches. First instance: coaching and a logged note (most early breaches are ignorance, not malice). Repeat or careless: manager escalation and re-training. Deliberate restricted-data exfiltration: the existing information-security disciplinary route — the AI policy should reference it, not reinvent it.
Review the policy quarterly against the tool list; generative tooling changes faster than annual policy cycles tolerate.
How to check your own position
Take five recent real situations from your company — a marketing draft, a support reply, a spreadsheet analysis, a piece of code, someone’s meeting summary — and test them against the policy as written. If the policy cannot answer three of the five in under a minute, it is written for lawyers, not employees. Then ask a non-technical employee the two questions that matter: which tool would you use for X, and what data can go into it? If the answers are wrong, the fix is not more policy — it is a shorter one, an approved list with real names, and the tier test with your own examples. The measure of success is not acknowledgement rates; it is whether someone resolves a borderline case without escalating it, correctly.
Frequently asked questions
Should a company ban ChatGPT and generative AI at work?
No — bans reliably produce shadow usage and lose the audit trail. An effective ai acceptable use policy for companies names approved tools, sets the data tiers, prohibits the genuinely unacceptable uses and requires human review of output. Controlled access with enterprise terms gives you visibility, contractual no-training protections and a defensible record; a ban gives you none of the three.
What should an AI acceptable use policy include?
Eight elements: scope (which tools and features are covered), the approved tool list with owner, a data-tier test with examples, prohibited uses as scenarios, review and attribution rules, incident reporting for suspected data leakage, the enforcement ladder, and review cadence. One page of rules plus the tool list beats ten pages of principles — employees apply what they can remember.
How do I introduce an AI usage policy to employees?
Acknowledgement plus a ninety-minute session built on your own examples: a good prompt, a bad prompt, a refused request, and the tier test as an exercise. Record attendance — under the EU AI Act it doubles as AI literacy evidence — and position the policy as enabling safe use, not restricting it. Policies introduced as permission outlive policies introduced as prohibition.
What are the most common mistakes in generative AI policies?
Banning instead of channeling; no approved tool list (so nothing is actually actionable); vague data rules (“confidential information” means nothing at 4pm); missing review duties, so AI drafts ship unchecked; omitting an enforcement ladder, which makes the policy unenforceable; and never being updated — a two-year-old tool list is a map of a country that no longer exists.
Does an AI policy satisfy the EU AI Act AI literacy requirement?
Partially. Article 4 requires measures to ensure sufficient AI literacy among staff operating AI on their behalf — a policy provides the rules, but literacy implies understanding, so pair it with recorded training on real examples. The policy document plus attendance records together make the defensible position: measures defined, measures delivered, evidence retained.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.