How to Write an AI Policy for Your Company

A company AI policy is a short written rulebook that tells staff which AI tools they may use, what data may go into them, and who is accountable. The usual failure is an employee pasting a customer file or source code into a consumer chatbot plan. In the EU, Article 4 (AI literacy) and Article 5 (prohibited practices) of the AI Act have applied since 2 February 2025, while the NIST and ISO frameworks are voluntary.

The policy has to keep three things apart: law in force today, law with a later date, and frameworks you may adopt but need not. Sources were read on 20 October 2026.

What should an employee AI policy say?

An employee AI policy should say which tools are allowed, which data classes may go into which tier of tool, who to ask, and what happens after a breach. The data rule does most of the work: staff do not need to understand model training to follow a table that says customer files go in the business plan only.

Table A is a starting data-class matrix. The cells are defaults to edit, not requirements from any law.

The left column is about the plan a person is signed in to, not the brand: one vendor can treat a free account and a business account differently, as Step 4 shows. A blanket ban also fails predictably. With no approved tool, staff use a personal account on a consumer tier, which is the case Table A forbids. The practical answer is to approve a business-tier tool early so the ban has an alternative.

What goes in an AI acceptable use policy?

An AI acceptable use policy has six parts: a definition of AI, a register of approved tools, data rules, prohibited uses, a disclosure rule, and an incident rule with a named owner and a clock. The tool register goes stale first, so each row carries its own owner and review date.

Table B shows the register format. The two rows are layout examples, not recommendations of any product.

The incident rule needs one named person who receives reports and a clock for how fast staff must report. The 24 hours in the skeleton below is a placeholder, not a legal figure, and your contracts or regulators may require less. State that prompt reporting is expected and not punished on its own, because a penalty for reporting teaches people to stay silent.

How do you write a company AI policy, step by step?

You write it in eight steps, each ending with one owner and one output: name the owner, inventory the tools, set data rules, check what vendors do with inputs, list prohibited uses, plan literacy training, decide what is disclosed, and fix a review date. A first draft takes about three to four hours if the tool inventory is roughly known.

Step 1: Who owns the AI policy, and what counts as AI?

One named role owns the policy, and the policy defines AI in a sentence staff can apply. Without a compliance team, the owner is usually the operations lead or head of IT; a committee as owner means nobody answers a question on Tuesday.

A workable definition: an AI system is software that generates text, images, audio, video, code or a recommendation from the input you give it, including AI features built into tools you already use. That is our drafting wording, not a legal definition. If your company uses AI to decide things about people, the legal definitions matter more, which is why Step 5 sends hiring and monitoring to a lawyer.

The action: the managing director names the policy owner by role and writes the AI definition into the policy. Output: a named owner and a definition.

Step 2: Which AI tools is staff already using?

Start with what staff already use, because a policy that ignores it gets ignored. Ask each team to list the AI tools, browser extensions and built-in AI features in use, and which plan each person is signed in to: personal free, personal paid, or company-provided.

The first list will be incomplete. Expense claims, single sign-on records and software-install requests are places to cross-check, and a short amnesty (list it now, no penalty for past use) gets a more honest list than a threat. Sort each tool as approved, restricted to some data classes, or prohibited.

Treat four things as new tools that go through the register before use: a new AI feature switched on inside an approved tool, a meeting note-taker, a browser extension, and an agent that connects to company accounts. Give staff a simple way to ask, such as one shared mailbox or form that the policy owner answers within a stated number of working days.

The action: the policy owner collects team lists within a week and sets up the request route. Output: a draft tool register in the Table B format and a named place to ask for a new tool.

Step 3: Which data can go into which tool?

Use Table A and attach it to the policy. Choose four or five data classes, give each a one-line example staff recognise (a price list, a customer email, a database export), and decide whether a consumer tier, an approved business tier, or no AI tool at all is allowed for each.

Do not rely on staff to judge sensitivity case by case. If you handle special categories of personal data or work under customer confidentiality terms, put those in the never column until the owner approves a specific tool in writing.

If customer or employee personal data will go into a tool, check with your data protection lead or a lawyer whether a data processing agreement with the vendor is needed and whether a data protection impact assessment applies. Record who decided and the answer in the tool's register entry.

The action: the policy owner, with whoever handles data protection, signs off the data classes. Output: Table A completed for your company.

Step 4: Does the vendor train on what you enter?

It depends on the plan, and the answer sits in the vendor's own documents. In short, OpenAI, Anthropic and Microsoft say their business offerings are not used for training by default, Google says Workspace content is not used for training outside your domain without the customer's permission, and OpenAI says its individual plans (Free, Plus, Pro) may be used for training, with an opt-out setting. Cite the page and the date you read it: these are the vendors' statements as read on 20 October 2026, with the page's own date where it shows one.

OpenAI's enterprise privacy page (updated 8 January 2026) says "We do not train our models on your data by default" for ChatGPT Business, Enterprise, Healthcare, Edu, Teachers and the API Platform. Its help article says "When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models", with an opt-out called "Improve the model for everyone" under Settings, then Data controls.

Anthropic's commercial-products article (18 August 2026) says inputs and outputs from Claude for Work, the API and Claude Gov are not used for training by default, except reported feedback or an explicit opt-in. Its consumer article (16 March 2026) says training on consumer chats depends on a setting the user chooses; we could not confirm the default or any retention period.

Google's Workspace privacy hub (6 October 2026) says content is not used for generative AI training outside your domain, and not without the customer's permission. For the consumer Gemini app (19 October 2026), activity can be used for training when Keep Activity is on, and the page does not state the default.

Microsoft's enterprise data protection page for Copilot (updated 18 August 2026) says "Your data isn't used to train foundation models", with prompts and responses covered by its Data Protection Addendum.

These pages change without notice, which is why Table B has a date column. At purchase, put three questions on the order form and keep the answers: does the vendor train on our inputs, how long does it retain them, and which sub-processors can see them.

The action: the IT lead fills the trains-on-inputs column of Table B from each vendor page. Output: a register where every approved tool has a source URL and a read date.

Step 5: What uses of AI are prohibited?

Some prohibitions are law and some are your own choice, and the policy should say which is which. Article 5 of the EU AI Act has banned certain practices since 2 February 2025.

Article 5(1)(a) covers an AI system that "deploys subliminal techniques beyond a person's consciousness", and 5(1)(b) one that "exploits any of the vulnerabilities of a natural person", such as age or disability. Article 5(1)(c) covers social scoring: classifying people over time by social behaviour or personal characteristics, leading to unjustified or disproportionate treatment. The workplace clause is Article 5(1)(f): "the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions, except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons". Two further bans, on AI-generated non-consensual intimate imagery of an identifiable person and on AI-generated child sexual abuse material, apply from 2 December 2026.

For hiring and monitoring staff, do not draft the rules yourself. Annex III, point 4 classes AI used for recruitment and selection, and AI used for decisions on promotion, termination, task allocation or monitoring performance, as high-risk, from 2 December 2027. In California, SB 947 will bar employers from relying solely on an automated decision system to fire or discipline workers from 1 July 2027, as we reported in our news item on the No Robo Bosses Act. Colorado's law on automated decisions was repealed and replaced by SB26-189 on 14 May 2026, with some dates unconfirmed (see the foot of this page). For the wider argument over federal and state limits, see our analysis of who limits AI in America, which is argument, not a reference table.

The policy rule that follows: no AI output is the sole basis for hiring, firing, discipline or pay. AI-detector scores are probabilistic and produce false positives, as our comparison of AI detectors covers, so never make one the sole basis for discipline. Then add your own prohibitions: restricted data in an unapproved tool, personal accounts for company work, and presenting AI output as checked when it was not.

Before using AI on client work, check the client contract for AI clauses, and decide who checks AI output for copied or unlicensed material. Confirm both points with a lawyer; we state no rule on either.

The action: the policy owner writes the prohibited list in two groups (law, company choice) and sends the hiring, monitoring and client-contract questions to a lawyer. Output: the prohibited-uses clause.

Step 6: Do staff need AI literacy training?

Article 4 of the EU AI Act requires measures to support AI literacy, and it does not mandate a particular course or certificate. Training is the usual measure. As amended by Regulation (EU) 2026/1744, Article 4 reads: "Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used. This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual."

The earlier wording about ensuring a "sufficient level" has been replaced, so a policy repeating it describes the Act as it stood in 2024. The Commission's AI literacy Q&A (last update shown 20 October 2026), paraphrased here, says no mandatory training is imposed, no certificate is needed, and internal training records are allowed. National market surveillance authorities, not the AI Office, enforce Article 4, with enforcement rules applying from 2 August 2026, and a sanction is more likely where an incident shows a "lack of appropriate training and guidance of employees".

A filed PDF nobody read is a weak answer; a short session with a record is a reasonable one. At joining and whenever a tool is added, cover Table A, the prohibited list, how to check AI output, and who to tell about an incident, and keep a dated attendance list.

The action: the policy owner, with HR, schedules the first session and starts the record. Output: a training clause and a dated attendance list.

Step 7: What must be disclosed to customers?

Under the EU AI Act a company's own disclosure duties as a deployer are narrow; the broad ones sit with providers. Article 50 applies from 2 August 2026. Article 50(1) says "Providers shall ensure that AI systems intended to interact directly with natural persons are designed" so people are told they deal with an AI system, and 50(2) requires outputs "marked in a machine-readable format and detectable as artificially generated or manipulated". Both are provider duties, so a company that ships its own chatbot is the provider of that system. Whether Article 111(3) delays 50(1) for systems already on the market is an open reading, and the EU AI Act Transparency Audit on this site reports both. The audit covers what providers publish; this guide covers what a company does.

The deployer duties are 50(3) and 50(4). Under 50(3), deployers of emotion recognition or biometric categorisation systems "shall inform the natural persons exposed thereto of the operation of the system". Under 50(4), deployers of deepfakes "shall disclose that the content has been artificially generated or manipulated"; the same paragraph reaches AI-generated text published to inform the public on matters of public interest, unless it had human review or editorial control and a person holds editorial responsibility.

For the policy: label AI-generated images, audio and video that show real or realistic people or events, and tell customers when they are talking to an AI assistant. The second half is a company choice, since the Act places 50(1) on the provider; if you buy the assistant, ask the vendor to confirm its 50(1) position in writing. On the US side, our item on the FTC investigation and the White House accord reports that an FTC spokesperson confirmed on 30 September 2026 an investigation of OpenAI, Anthropic and other AI companies over safety. That concerns the model makers, and we found no FTC instrument on a company's disclosure of its own AI use that we could cite as current.

The action: the communications lead drafts the label and chat-notice wording and the policy owner approves it. Output: a disclosure clause with approved wording.

Step 8: Who reviews the policy, and when?

Pick one review cadence with one owner, and add event triggers. Six months is our suggested default, not a requirement of any source we read. Triggers for an early review: a vendor changes its data terms, a new tool is approved, an incident happens, or a legal date arrives.

The legal dates to calendar, as of 20 October 2026: 2 December 2026 (new Article 5 bans, and the Article 111(4) deadline for providers of existing generative systems to comply with Article 50(2) marking), 1 July 2027 (California SB 947 operative), 2 December 2027 (Annex III high-risk, including employment) and 2 August 2028 (Annex I high-risk).

Keep consequences proportionate and consistent with the rule that prompt reporting is not punished: an accidental breach that is reported promptly is a coaching conversation, a repeat breach goes through your normal disciplinary process, and a deliberate breach, such as hiding the use of a banned tool, can be treated as misconduct.

The action: the policy owner puts the review date, the triggers and the consequences ladder in the policy and a review date in a calendar invitation. Output: a dated review line with a named owner.

What does an AI policy template look like?

The template has 12 clauses: purpose and scope, what counts as AI, owner, tools, data, prohibited uses, human review, disclosure, training, incidents, consequences and review. A one-page minimum version follows it, and both are ungated.

Paste either into your own document and edit at the square brackets. The wording is ours, drafted for this guide; it is not a form from any regulator. The label and chat-notice lines in clause 8 are drafting aids, not wording any law prescribes.

[COMPANY NAME] AI USE POLICY
Version [x]. Owner: [role]. Approved by: [role]. Date: [date]. Next review: [date].

1. Purpose and scope
This policy sets out which AI tools staff, contractors and temporary workers may use for company work, what data may go into them, and who is responsible. It applies to all work for [company], on any device.

2. What counts as AI
An AI system is software that generates text, images, audio, video, code or a recommendation from the input you give it, including AI features built into tools we already use.

3. Owner
[Role] owns this policy and the tool register. Questions go to [role / address].

4. Tools
The tool register lists every approved tool, the plan we use, the data classes it may handle, and the date we last checked the vendor's data terms. A tool not on the register is not approved. Personal accounts on consumer plans may not be used for company work. A new AI feature switched on inside an approved tool, a meeting note-taker, a browser extension, or an agent that connects to company accounts counts as a new tool: ask [role / address] and wait for approval.

5. Data
Public information may go into any approved tool. Internal information may go only into business-tier tools on the register. Customer personal data may go only into a tool whose register entry allows it, after [role] has recorded whether a data processing agreement and an impact assessment are needed. Never enter into any AI tool: unannounced deals or layoff plans; special-category personal data without [role]'s sign-off; passwords, API keys or private keys. Anything not covered is not allowed until [role] says otherwise.

6. Prohibited uses
Staff may not: use AI to infer the emotions of colleagues or candidates; use AI to score or rank people on behaviour or personal characteristics; make a hiring, firing, discipline or pay decision on AI output alone; generate realistic images, audio or video of a real person without that person's written consent; or use AI to generate sexual imagery of an identifiable person, or any sexual imagery of children.

7. Human review
A person is responsible for everything sent under their name. Check facts, figures, names, quotations and code before use. AI output is not a source. Before using AI on client work, check the client contract for AI clauses and check output for copied or unlicensed material; [role] decides who does the check.

8. Disclosure
Label AI-generated or AI-edited images, audio and video that show real or realistic people, places or events (for example: This image was created or altered with AI). Tell customers when they are talking to an AI assistant (for example: You are chatting with an AI assistant. A person can take over on request).

9. Training
Everyone covered attends a short AI session at joining and when a new tool is added. [Role] keeps a dated attendance list.

10. Incidents
Tell [role / address] within [24 hours] if you put restricted data into an unapproved tool, or an AI tool does something unexpected with company or customer data. Prompt reporting is expected. [Role] records the incident, decides who else must be told, and reviews the register.

11. Consequences
Consequences are proportionate. An accidental breach reported promptly leads to coaching and is not punished for the reporting. A repeat breach is handled under [existing disciplinary process]. A deliberate breach, such as hiding use of a banned tool, is treated as misconduct. [Role] decides.

12. Review
[Role] reviews this policy every [6 months], and sooner if a vendor changes its data terms, a new tool is approved, an incident occurs, or a legal date on the calendar at [location] arrives.

If you only have time for a page, use this minimum and add the rest later.

[COMPANY NAME] AI USE: ONE-PAGE MINIMUM. Owner: [role]. Next review: [date].
1. Use only the AI tools on the register at [location]. No personal accounts for company work. Ask [role / address] before switching on a new AI feature, note-taker, extension or agent.
2. Public information may go into any approved tool. Internal and customer data only into tools whose entry allows it. Never enter passwords, keys or secrets.
3. Do not use AI to judge people's emotions, rank staff or candidates, or decide hiring, firing or discipline on its own.
4. You are responsible for everything you send. Check AI output before use.
5. Label AI-generated images, audio and video of real or realistic people. Tell customers when they are talking to an AI.
6. Report a mistake to [role] within [24 hours]. Prompt reporting is expected.
7. Everyone attends a short AI session at joining and when a tool is added.

What is a simple AI policy for a small business (under 50 staff)?

A business under 50 staff can run on the one-page minimum plus three decisions: who owns it, which one or two tools are approved on a business plan, and a calendar date for review. Fifty is our rule of thumb for when one page stops being enough, not a legal threshold.

The EU AI Act names small and medium-sized enterprises directly: Article 4(2) says the Commission and the Member States shall support and facilitate the efforts of providers and deployers, in particular SMEs, and the Commission is to publish practical examples of how to comply. That support does not remove the Article 4 duty to take measures. For a small company, the measure can be one 30-minute session and a signed attendance list. Write the customer-data rule from Table A first, because that is where a small business has most to lose.

Is an AI ethics policy the same as an AI use policy?

No: an AI ethics policy states the values a company holds, while an AI use policy states the rules staff follow, and only the use policy changes behaviour. An ethics statement with no tool register, data rule or owner is a document nobody can apply on a Tuesday afternoon.

The pragmatic route is one document: open the use policy with an ethics preamble of three to five lines and tie each line to a rule in the body. The OECD AI Principles, adopted in 2019 and updated in May 2024 per the OECD's page, give a ready list of values: inclusive growth, sustainable development and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; and accountability. Where a law such as Article 5 already prohibits something, the ethics line and the prohibited-use clause can say it once. Where a value has no rule behind it, write the rule or drop the claim.

Which EU AI Act duties apply to a company, and from when?

For a company using AI tools at work, Article 4 literacy and Article 5 prohibitions have applied since 2 February 2025 and Article 50 transparency since 2 August 2026, while high-risk employment systems move to 2 December 2027. Regulation (EU) 2026/1744, the Digital Omnibus on AI (Official Journal, 24 July 2026), changed Article 4 and the high-risk dates, so text written from the 2024 original is out of date in places.

The dates in Table C come from the Commission AI Act Service Desk consolidation as at 27 July 2026, checked against the Article pages on 20 October 2026. Article 113 sets 2 August 2026 as the general date, with Articles 4 and 5 from 2 February 2025, new Article 5 provisions (ba), (bb), (1a) and (1b) from 2 December 2026, Annex III high-risk from 2 December 2027 and Annex I from 2 August 2028. Article 50 is not in the delayed list.

Article 111(4) gives providers of systems generating synthetic audio, image, video or text, placed on the market before 2 August 2026, until 2 December 2026 to comply with Article 50(2). That is a provider transition for machine-readable marking, not a pause on deployer duties. Penalty amounts are not stated here because we did not read Article 99, and each duty binds a company only if the company is within the Act's scope, which this guide does not decide.

Which frameworks should an AI policy map to?

Map the policy to NIST AI RMF 1.0 and its Generative AI Profile for a recognised structure, and to ISO/IEC 42001 only if you want a certifiable management system. Both are voluntary. NIST says the framework "is intended to be voluntary, rights-preserving, non-sector-specific, and use-case agnostic", and ISO says "Certification for ISO/IEC 42001 is voluntary."

The AI RMF (NIST AI 100-1, January 2023) has four functions: GOVERN, which "cultivates and implements a culture of risk management within organizations"; MAP, which "establishes the context to frame risks related to an AI system"; MEASURE, which covers analysing, assessing, benchmarking and monitoring AI risk; and MANAGE, which "entails allocating risk resources to mapped and measured risks on a regular basis". NIST AI 600-1 (July 2024) is a companion for generative AI and names 12 risks, including Confabulation, Data Privacy and Information Security. ISO describes ISO/IEC 42001:2023 as specifying "requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS) within organizations." The standard costs CHF 225 as shown on 20 October 2026, so Table D uses only that public description. ISO does not certify organizations itself; independent certification bodies do.

Table D is our mapping, a reading aid and not NIST's, ISO's or the OECD's own table.

Nothing in the sources we read says following NIST or ISO discharges a duty under the AI Act. Use Table C as the legal checklist and Table D as the structure for the rest.

How do you check that your AI policy works?

Give the finished text to someone who did not write it and ask five questions: which tools may I use, may I paste this customer file into this tool, who do I tell if I get it wrong, who owns this policy, and when is the next review. If any answer takes more than a minute to find, rewrite that clause.

Then run four checks. Register: every approved tool has a plan, data classes, an owner, a trains-on-inputs entry with a URL and a read date, and a review date; re-open each vendor page once and update any row whose wording changed. Citations: for each Article you cite, open the text on the EU AI Act Service Desk or EUR-Lex and confirm the paragraph and date, and confirm the policy says "take measures" for Article 4 and not the superseded "sufficient level" wording. Training: a dated attendance list exists. Incident path: ask a colleague to send a test report to the named person and note how long the reply takes. Put the review date, the legal dates from Step 8 and the vendor-term reviews in one shared calendar owned by the policy owner.

Is this guide legal advice, and what does it leave out?

No: this guide is general information, not legal advice, and we are not a law firm. We read the documents cited, on the dates stated, and we did not test or implement any policy or tool. Whether the EU AI Act, a US state law or any other rule applies to your company depends on facts about your business and location that we do not know, and this guide does not decide the Act's territorial scope.

It leaves out hiring and monitoring rules, which need a lawyer; the Act's risk categories in full; penalty amounts; sector rules for finance, health or education; and data-processing contracts. Items we could not confirm against a primary source are listed in the "What we could not confirm" section at the foot of this page. Have a qualified person read the skeleton before it binds your staff.