AI at Work

A One-Page AI Usage Policy Template for Small Teams

Copy a one-page AI usage policy for small teams: approved tools, four data tiers, review levels, incident steps, and EU AI Act and GDPR notes.

By Editorial team

Published 12 min read

Ask this guide

Answers come straight from this guide’s text. Nothing is generated, and nothing you type leaves your browser.

Say a six-person design agency has no AI rules. One designer pastes a client's customer list into a free chatbot to fix its formatting. A project manager connects an assistant to the shared drive. A new hire asks whether a client-call transcript can go into a summariser and gets three answers from three colleagues. Nobody acted in bad faith; nobody knew where the line was.

The Dutch data protection authority (Autoriteit Persoonsgegevens, AP) described this pattern in August 2024: breach notifications after employees put patients' or customers' personal data into AI chatbots, among them a GP practice employee's medical data and a telecom employee's file of customer addresses. Its advice was to make clear agreements about whether chatbots may be used and which data may go in.

This guide puts those agreements on one page: four data tiers, review levels, a tool checklist, an incident routine, and EU notes for teams that only use AI tools. It is a starting point from a software studio, not legal advice.

Sort data into four tiers before writing any other rule

Most policy arguments are really arguments about data, so settle the tiers first.

Tier Covers May go into
0, Open Already public: published posts, public docs Any approved tool
1, Internal Drafts and process notes without client or personal data Approved tools only
2, Confidential Client material under NDA, unreleased numbers, source code, and personal data such as names, emails and customer lists Only tools marked tier 2; paste the minimum
3, Never Passwords and keys; special-category data (health, racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic data, biometric data used to identify a person, sex life or sexual orientation); anything a contract bars No AI tool

Personal data is "any information relating to an identified or identifiable natural person" (GDPR Article 4(1)); the special categories come from Article 9(1), which prohibits processing them by default. GDPR links here go to gdpr-info.eu, an unofficial copy of Regulation (EU) 2016/679. Now sort seven pastes:

  1. A published blog draft: tier 0.
  2. Last week's meeting agenda: tier 1.
  3. A client's strategy deck under NDA: tier 2.
  4. A CSV of customer names and emails to deduplicate: tier 2, in a tool marked for personal data, only the columns needed.
  5. A client-call transcript: tier 2 at minimum (names, voices, NDA content). Check where the recorder stores it.
  6. A .env file to debug a config error: tier 3. Swap in placeholders first.
  7. A colleague's sick-leave email to summarise: tier 3 (health data).

Swapping names for IDs lowers exposure but does not anonymise: Recital 26 treats pseudonymised data that can be linked back to a person as personal data. When unsure, treat a paste as one tier higher and ask.

The one-page template, ready to copy

A policy is a decision aid for the moment of pasting, so it has to be short enough to read then. Every rule below is one a colleague could break and someone could notice. Copy the block and fill the brackets.

# AI usage policy: [Company name]

Owner: [name, role] | Deputy: [name] | Version 1.0 | Effective [date] | Next review [date]

## 1. Scope
Everyone who works for [Company], including freelancers and interns.
"AI tool" means any generative AI product or feature: chatbots,
assistants inside other software, extensions, meeting recorders.

## 2. Approved tools
Use only tools on this register, with your work account.

| Tool and plan | Highest tier | Training on our data | DPA | Client-approved | Checked on |
|---|---|---|---|---|---|
| [Tool, plan] | [0-2] | [Off] | [Yes/No] | [Yes/No/n.a.] | [date] |

To request a tool, tell [owner] what it is for, which data it will see
and where its data terms are. Answer within five working days.

## 3. What may go in
- Tier 0, Open (already public): any approved tool.
- Tier 1, Internal (no client or personal data): approved tools.
- Tier 2, Confidential and personal data: only tools marked tier 2;
  client data only if the client approved the tool. Paste the minimum.
  No feedback buttons on these chats.
- Tier 3, Never (passwords, keys, special-category data, anything a
  contract forbids): no AI tool.
When unsure, treat it as the higher tier and ask [owner].

## 4. Check before you rely on or share output
- A: only you use it. Read it first.
- B: colleagues rely on it, or it enters an internal document or code
  review. Verify every claim a decision depends on; a colleague reviews.
- C: it leaves the company, or involves money, law, health, security or
  a person. A named reviewer who did not write the prompt signs off;
  log it in [shared sheet] with date, check and reviewer.
AI makes no final decision about hiring, pay, promotion or dismissal,
and is not used to infer employees' emotions.

## 5. Disclosure
Tell clients about AI use when their contract requires it or they ask.
Tell people when they are talking to an AI system. Label realistic
AI-generated images, audio or video of real people or events.

## 6. Ownership and security
AI output is a draft; add human work before delivery. Use single
sign-on and MFA where offered. No unapproved extensions, plugins or
connectors; connectors get the least access that works.

## 7. When something goes wrong
Tell [owner], or [deputy], the same working day. Good-faith reports
are never punished.
1. Stop. Note the tool, the time and what was entered. Rotate any
   pasted password, key or token now; do not wait for a decision.
2. Do not try to fix it by pasting more.
3. The owner requests deletion where the vendor allows it, informs any
   client whose data was involved, decides on any regulator
   notification, and records each decision.

## 8. Training and review
Everyone attends the 45-minute session within 30 days of joining and
when this policy changes, and confirms in writing that they have read
it. We keep a dated record. The owner reviews this policy, the register
and all connectors and extensions quarterly.

If you keep a shared prompt library, add each template's data tier as a field.

Pick a review level by who acts on the output and how hard it is to undo

Section 4 of the template names three levels. The skill is assigning them quickly.

Example Level Why
Ten name ideas for your own brainstorm A Only you use it
Code merged to main B, or C if it touches authentication, payments or personal data Run it, test it, read the diff
Client proposal with market figures C It leaves the company, and the figures can be checked
Shortlisting CVs or rating staff Not delegated See below

Our risk-tiered method for checking AI output covers how to run the checks; its tiers 1, 2 and 3 correspond to levels A, B and C here.

The last row is our policy choice, informed by the Act. Annex III lists as high-risk the AI used to recruit or select people (including filtering applications), decide on promotion or dismissal, allocate tasks based on personal traits, and monitor or evaluate performance. High-risk is not a ban: it brings extra duties for the deploying organisation, such as assigning competent human oversight and informing affected workers (Article 26), applying from 2 December 2027. Article 5(1)(f) is stricter: it prohibits inferring emotions in the workplace, except for medical or safety reasons. Keep people decisions with people, and take legal advice before a tool touches them.

What to check before a tool goes on the register

Approve the plan, not the brand: the same product can carry different data terms on different plans. Two vendors show the pattern.

  • Anthropic. On consumer plans (Free, Pro, Max), chats may be used to improve Claude when the "Model Improvement" setting is on; Incognito chats are excluded. Turning it off does not cover everything: feedback and conversations flagged for safety review can still be used. For commercial products such as the API, Anthropic says inputs and outputs are not used for training by default unless you opt in or send feedback.
  • Microsoft. Its Copilot documentation says prompts, responses and Microsoft Graph data are not used to train foundation models, that Copilot only surfaces data a user already has permission to view, and that admins decide whether Anthropic and OpenAI models power it. The product is now named Microsoft Copilot (formerly Microsoft 365 Copilot).

An assistant connected to your drive does not break permissions; it makes over-sharing easy to find. Fix the over-shared folder first.

Check Look for
Training on inputs The default for your plan, who can change it, and routes the setting does not cover
Feedback buttons Whether feedback stores or reuses the conversation
Retention and deletion How long chats are kept; whether an admin can delete them
Processor terms A data processing agreement; Article 28(3) requires processing by a processor to be governed by a contract
Client authorisation If you process a client's personal data, the AI vendor is your sub-processor, and Article 28(2) says a processor shall not engage another without the controller's prior written authorisation, specific or general. Check the client's contract or DPA before tier 2 client data goes in
Location and sub-processors Where data is processed and by whom; Microsoft's page notes Anthropic models are currently outside its EU Data Boundary
Admin and connectors SSO and MFA; who sees chat history; what a connector can reach
Output indemnity Whether the vendor defends customers against copyright claims over output, and on what conditions; Microsoft ties its commitment to using the built-in guardrails and content filters

Tip: Treat feedback buttons as sending the whole conversation to the vendor. Anthropic's consumer documentation says a thumbs-up or thumbs-down stores the entire related conversation for up to five years, and the feedback may be used for training. On Team and Enterprise plans, owners can switch member feedback off with the "Rate chats" setting.

Vendor terms change; we checked these pages in October 2026, so record each check date in the register.

EU notes: the AI literacy duty, key dates and GDPR basics

Warning: This is general information for EU teams, not legal advice. Laws and enforcement change; check what you rely on with counsel or a data protection officer.

Article 4 of the AI Act, as amended in July 2026

Article 4 has applied since 2 February 2025 to providers and deployers. A deployer is anyone using an AI system under their authority, except for personal non-professional activity (Article 3(4)), so a team using a chatbot for work is one.

Regulation (EU) 2026/1744, the "Digital Omnibus on AI", entered into force on 27 July 2026. Article 4 originally asked providers and deployers to "take measures to ensure, to their best extent, a sufficient level of AI literacy" of their staff. It now asks them to "take measures to support the development of AI literacy" and "does not require providers or deployers to guarantee any specific level of AI literacy of any individual." The Commission's 2025 proposal would have moved the duty to the Commission and Member States; the adopted text keeps it with companies and adds, in Article 4(2), that the Commission publishes practical examples of compliance.

The Commission's AI literacy Q&A adds that "There is no need for a certificate", that organisations "can keep an internal record" of training, and that contractors acting on your behalf are in scope. National market surveillance authorities supervise it from August 2026 and could impose penalties. Our reading, not a guarantee: a written policy, a short session, a dated attendance record and a quarterly review make a reasonable set of measures for a small team.

Dates that matter if you only use AI tools

Date What starts For your team
2 August 2026 Most of the Act, including Article 50 transparency Article 50(4): disclose deepfakes, and AI-written text published to inform the public on matters of public interest unless a human reviewed it and a person holds editorial responsibility. Marking output as machine-readable is the provider's job
2 December 2027 High-risk duties for Annex III uses, including employment Applies if AI screens applicants or evaluates staff

The Omnibus moved the Annex III date from 2 August 2026 to 2 December 2027.

GDPR basics for what goes into a prompt

  • Breaches have a clock, and whose clock depends on your role. The controller (the organisation deciding why and how data is used) must notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of a breach, unless it is unlikely to risk people's rights and freedoms (Article 33(1)); if the risk is high, it must also tell the people affected (Article 34). A processor, such as a vendor or an agency handling a client's customer data, must tell the controller without undue delay (Article 33(2)); the contract may set a shorter deadline. The AP says personal data entered into a chatbot against agreed rules is a breach; chatbot use that is part of an organisation's policy is not a breach, it adds, but "often not permitted by law".
  • Proposed changes are not law yet. A separate Commission proposal of 19 November 2025 would amend the GDPR, including a longer breach-notification deadline and a higher risk threshold (EDPB and EDPS comment). In October 2026 Parliament's procedure file still showed it awaiting committee decision, so the rules above are the ones in force.

Rolling it out: a 45-minute session and a quarterly review

  1. Build the register (about 90 minutes). Ask in one message which AI tools people used last month, add each, and check it against the table above.
  2. Run the 45-minute session. Five minutes on why, starting from the Commission's own example: staff using ChatGPT for advertising text or translation, who "should be informed about the specific risks, for example hallucination". Then 15 sorting real pastes into tiers (reuse the seven above, then add your own); 10 on review levels, tailored to roles; 5 on an incident drill (who you tell, and when); 10 on questions and tool requests. Record the date and attendees as your internal record of literacy measures, and add each question to a short FAQ under the policy.
  3. Turn the tier check into an if-then plan. "If I am about to paste a table with an email column, then I check its tier first." Pin four questions beside the screen: What tier is this? Does the tool's register entry allow it? Can I remove names, keys or columns first? Who will act on the output, and at what review level? Our guide to if-then planning covers how to write these.
  4. Review quarterly (30 minutes). Re-check the register and settings, remove any connector or browser extension with access to mail, drive or code that is not on the register, then go through incidents, near misses and requests for new uses. More near-miss reports after the session is a good sign.

Where to start this week

Today, build the register. Tomorrow, confirm each tool's training setting on the plan you actually use, and cap unclear ones at tier 0 (public material only) or take them off. By Friday, fill the template, book the session and calendar the first review.

If you do only one thing, publish the tier 3 line. A pasted key has to be treated as leaked and rotated, and health data is hard to take back.

Frequently asked questions

We are three people. Does an AI policy apply to us?

The Commission's Q&A applies Article 4 to "all providers and deployers of AI systems" and mentions no size exemption. A three-person policy can be half a page and one conversation; the data tiers matter most.

Why not simply ban AI tools?

The AP notes that employees often use chatbots on their own initiative, contrary to what their employer agreed, and it leaves not allowing them as an option. A ban only works if people follow it. In our view, a short approved list with clear data rules gives people a compliant route and makes problems easier to report.

That depends on jurisdiction, and we have verified only the US position. In its January 2025 copyrightability report, the U.S. Copyright Office concluded that purely AI-generated material is not copyrightable and that, with current generally available technology, prompts alone do not give sufficient control, while human-authored expression and creative selection, coordination, arrangement or modification are protected. Sufficiency is judged case by case. In practice, keep drafts and edits as a record of human contribution, and check who warrants originality in client contracts.

What if someone already pasted client data into a free chatbot?

Follow section 7 of the template: note the tool, time and contents, rotate any pasted credential, and tell the owner the same day. The owner checks whether personal data was involved. For your own staff or customer data you are the controller, and the 72-hour clock in Article 33(1) applies if the breach is likely to risk people's rights. For a client's data you are probably a processor and must tell the client without undue delay (Article 33(2)); the contract may set a shorter deadline.

Sources and further reading