An AI policy template for employees is a short internal rule set that says which AI tools staff may use, through which accounts, what data may go into them, who checks the output and how to report mistakes. The copy-ready template below covers this in 11 clauses. Fill in the bracketed parts, have management approve it and train staff before it takes effect.
Key points
- The rule employees use most is the data classification: public, internal, confidential and personal data, each with a plain answer to "may this go into the tool?"
- The account type decides more than the brand. OpenAI, Microsoft, Anthropic and Google all state that their business plans do not train models on customer data by default.
- Assistants built into Microsoft 365 or Google Workspace can reach everything the user can open. Datatilsynet's sandbox report with NTNU told organisations to "get your own house in order" on access rights first.
- Datatilsynet considers a data protection impact assessment (DPIA) required as a general rule when generative AI processes personal data.
- A policy needs training, a named owner and a review date to work. Plan about four weeks from first draft to the date it applies.
What should an AI policy for employees cover?
An AI policy for employees should cover eight things: scope, approved tools and accounts, what data may be entered, human review of output, disclosure, approval of new tools, logging, and training, plus an incident routine and a review cycle. Two pages is enough for most companies.
The policy is a set of working rules for people who already use ChatGPT, Copilot, Claude or Gemini at work, often on personal accounts their employer has never seen. It does not replace the privacy work the GDPR requires (records of processing, data processing agreements, DPIAs). It tells employees the result of that work, in a form they can apply in ten seconds: which tools have passed the checks, and for which kinds of data.
Which data may employees paste into ChatGPT, Copilot, Claude or Gemini?
Employees may paste public information into any approved tool, internal information only into company business accounts, and confidential or personal data only into tools approved for that class and for a recorded purpose. Passwords, access keys and special categories of personal data stay out entirely unless a specific solution has been assessed for them.
Personal data is any information about an identifiable person (GDPR Article 4): a name, an e-mail address, a CV, a customer number, a voice recording. Special categories are listed in Article 9 and include health, ethnic origin, religion, trade union membership and biometric data.
| Class | Examples | Personal or free account | Approved company account |
|---|---|---|---|
| Public | Published web pages, press releases, public reports, your own job ad after it is live | Allowed, but use the company account for work anyway | Allowed |
| Internal | Process descriptions, meeting notes without names, draft marketing plans, internal training material | Not allowed | Allowed |
| Confidential | Contracts, unreleased financial figures, tenders, source code, client documents under a non-disclosure agreement | Not allowed | Only in tools approved for confidential data, by the roles named in the approved list |
| Personal data | Customer records, CVs, e-mails with names, call recordings, HR cases | Not allowed | Only for an approved purpose in an approved tool, with as little data as the task needs |
| Never | Passwords, API keys, payment card data, national identity numbers, health and other special-category data, information under a legal duty of confidentiality | Not allowed | Not allowed, unless a dedicated solution has been assessed and approved for it |
Removing a name does not make text anonymous: under GDPR Recital 26, data is anonymous only when the person can no longer be identified by means reasonably likely to be used. A complaint that mentions a job title, a town and a date may still point to one person.
The European Data Protection Board gives another reason to keep personal data out of prompts. Its ChatGPT taskforce noted that ChatGPT's outputs "are likely to be taken as factually accurate by end users, including information relating to individuals, regardless of their actual accuracy". A wrong sentence about a named customer is a data accuracy problem under GDPR Article 5.
What do business AI accounts change compared with personal accounts?
Business AI accounts change the contract: whether your data trains the vendor's models, who owns inputs and outputs, whether a data processing agreement applies, and whether an administrator can manage users and see logs. Personal accounts, free or paid, fall under consumer terms that the individual accepts and the company cannot control.
| Tool and plan | Training on your data by default? | What else the vendor states | Check before approving |
|---|---|---|---|
| ChatGPT Business, Enterprise, Edu and the OpenAI API | No | OpenAI says the organisation owns its inputs and outputs | European data residency is listed for eligible Enterprise, Edu and API customers, so confirm what applies to your plan |
| Microsoft Copilot with a work account (formerly Microsoft 365 Copilot) | No | Prompts and responses are stored like other Microsoft 365 content and can be managed in Microsoft Purview | Copilot surfaces everything the user has at least view access to. Anthropic models inside Copilot are currently outside the EU Data Boundary |
| Claude Team, Enterprise and the Anthropic API (commercial terms) | No | The customer owns its outputs, and Anthropic's data processing addendum is part of the terms | Conversations rated with thumbs up or down may be used for training. Owners can switch off "Rate chats" in the organisation's privacy settings |
| Gemini in Google Workspace with a work account | No, without the customer's permission | Gemini only retrieves content the user can already access, and content is not reviewed by people | Applies to qualifying Workspace editions, not to personal Google accounts |
| Personal ChatGPT, Claude or Gemini accounts | Depends on the user's own settings | Claude Free, Pro and Max let the user choose; if training is allowed, Anthropic keeps data for up to five years instead of 30 days | The company has no contract, no admin control and no log. Keep work out of these accounts |
A business plan does not remove your duties. Your company is still the controller for any personal data employees enter, the vendor acts as processor and needs a data processing agreement under GDPR Article 28, and the processing belongs in your record of processing activities. For a wider look at plans and vendor terms, see our guide to choosing AI tools for business and checking the data terms and the comparison of ChatGPT, Claude and Gemini.
Datatilsynet's exit report on Microsoft 365 Copilot adds the access point: "Copilot will have access to the same information that the user of the tool has". Loose sharing settings in SharePoint, Teams or Drive become visible the first time someone asks the assistant about salaries.
AI policy template for employees (copy-ready)
The AI policy template below fits a company of roughly 10 to 500 people. Copy everything inside the box, replace the text in square brackets and delete clauses that do not apply. This template is general information, not legal advice. Have your data protection officer, privacy contact or lawyer check it against how your company actually processes data.
Template: AI use policy (copy from here)
[Company name] AI use policy
Version [1.0]. Approved by [name, role] on [date]. Policy owner: [name, role]. Next review: [date, at most 12 months after approval].
1. Purpose and scope
This policy sets the rules for using artificial intelligence (AI) tools in work for [Company name]. It applies to all employees, temporary staff, consultants and contractors who handle company information or use company devices. It covers every AI tool: chat assistants such as ChatGPT, Microsoft Copilot, Claude and Gemini, AI features inside other software, browser extensions, transcription and meeting assistants, and image, video and code generators.
2. Approved tools and accounts
- 2.1 Use only the tools on the approved list at [location, for example an intranet page], and only through the company account you have been given ([single sign-on with your work e-mail]).
- 2.2 Do not use personal accounts, free or paid, for work tasks. Do not install AI browser extensions or plug-ins that are not on the list.
- 2.3 The approved list shows, for each tool: plan, the highest data class it is approved for, the roles that may use it and the owner. Example: [Microsoft Copilot Chat, work account, internal data, all staff, IT manager].
- 2.4 New AI features that appear in software we already use count as a new tool until [policy owner] has checked them.
3. What you may enter: data classes
- 3.1 Public information may be used in all approved tools.
- 3.2 Internal information may be used only in approved company accounts.
- 3.3 Confidential information (contracts, unreleased figures, tenders, source code, client material under a confidentiality agreement) may be used only in tools approved for confidential data and only by the roles listed for that tool.
- 3.4 Personal data (any information about an identifiable person) may be used only for the purposes listed in [the approved list or record of processing], in approved tools, and limited to what the task needs. Remove names and identifying details where the task works without them.
- 3.5 Never enter passwords, access keys, payment card data, national identity numbers, health information or other special categories of personal data, or information covered by a legal duty of confidentiality, unless [policy owner] has approved a specific solution for it.
- 3.6 Uploading files, sharing screens and connecting an AI tool to e-mail, calendars, drives or other systems follow the same rules as pasting text.
- 3.7 If you are unsure which class applies, treat the information as the stricter class or ask [contact].
4. Human review of output
- 4.1 You are responsible for anything you send, publish or decide with help from AI, exactly as if you had written it yourself.
- 4.2 Check facts, figures, names, quotes, sources and links, and any legal, financial or technical statement, before you use the output. AI tools can produce convincing text that is wrong.
- 4.3 AI may support but must not make decisions about individuals, such as hiring, assessments, pay, credit or complaints. A person makes the decision and can explain it.
- 4.4 AI-generated code is reviewed and tested like any other code before it reaches production.
5. Disclosure
- 5.1 Tell customers and other external people when they communicate directly with an AI system, such as a chatbot or an automatic reply, unless that is obvious.
- 5.2 Label images, audio and video made or altered with AI that show real people, places or events, or could be taken as real.
- 5.3 Tell the person who reviews your work when a deliverable relies heavily on AI.
- 5.4 Follow customer contracts that limit or prohibit AI use. [Account managers] keep the list of such customers.
6. Approval of new tools and new uses
- 6.1 Request a new tool, or a new use of an approved tool (for example using it with customer data), through [form or e-mail address] to [policy owner].
- 6.2 Before approval, [policy owner] checks: the data processing agreement, whether the vendor trains on our data, where data is stored and processed, retention and deletion, admin controls and log access, and security documentation. [Privacy contact] decides whether a DPIA is needed and updates the record of processing activities.
- 6.3 You get an answer within [10 working days]. Approved tools are added to the list with the data class they are approved for.
7. Logging
- 7.1 Company AI tools may store your prompts, uploaded files and the responses. Administrators can access these records for [security, troubleshooting, handling legal requests and responding to incidents].
- 7.2 The records are not used to monitor or assess individual employees. Access to one person's records follows [our routine for access to employees' e-mail and files].
- 7.3 Records are kept for [period] and then deleted, unless an incident or legal obligation requires longer storage.
8. Training
- 8.1 Everyone completes [basic AI training, length] before getting access to company AI tools, and a refresher when this policy changes in substance.
- 8.2 People who use AI with confidential or personal data complete [role-specific training].
- 8.3 [Policy owner] keeps a record of who has completed which training.
9. Reporting incidents
- 9.1 Report to [contact and channel] at once, and no later than the same working day, if you entered information you should not have, if AI output with errors reached a customer or was published, if a tool shows other people's data or behaves unexpectedly, or if you suspect an account has been compromised.
- 9.2 Reporting your own mistake quickly is expected and is not in itself a reason for sanctions.
- 9.3 [Privacy contact] assesses whether the incident is a personal data breach that must be reported to the data protection authority within 72 hours.
10. Roles
- 10.1 [Policy owner] keeps this policy and the approved list up to date and handles requests. [Privacy contact or data protection officer] advises on personal data, DPIAs and breaches.
- 10.2 Managers make sure their teams know the policy and have completed training. Everyone follows it and asks when unsure.
11. Review and acknowledgement
- 11.1 This policy is reviewed at least every [12 months], and also when a tool is approved for confidential or personal data, when a vendor changes its terms, when the law changes or after a serious incident.
- 11.2 Breaches are handled under our ordinary employment routines.
- 11.3 I have read and understood this policy. Name: [ ] Date: [ ]
End of template.
How do you adapt the template to your company?
Adapt the template by making three decisions before you edit any wording: which tools you approve, how far personal data may go, and who owns the policy.
- Tools. Start with what you already pay for. The Microsoft 365 or Google Workspace assistant with work accounts may cover most needs if sharing permissions are tidy. Add a dedicated assistant only for a documented need.
- Personal data. A strict first version that keeps personal data out of all AI tools is easy to explain and easy to loosen later. Open specific uses (for example summarising customer calls) one by one, each with its own check and, where needed, a DPIA.
- Owner. Name one person, often the operations lead or IT manager, with the privacy contact as adviser.
If you employ people in Norway, check clause 7 against the Norwegian e-mail access regulation. Datatilsynet considered it likely that a log of Copilot interactions "could fall under the prohibition against monitoring employees' use of electronic equipment", with possible exemptions for managing the network and handling security breaches. Purpose, access and retention of AI logs therefore need to be written down, and control measures must be discussed with employee representatives under section 9-2 of the Working Environment Act.
Watch out
A policy that bans all AI use rarely stops it. People move to personal accounts on their phones, where the company has no contract, no log and no way to help after a mistake. Approve one safe option before you forbid the unsafe ones.
How do you roll out an AI policy in four weeks?
Roll out an AI policy in four weeks by finding out what people use today, approving tools, filling in the template, getting it approved and training everyone before the effective date.
| When | Step | Owner | Output |
|---|---|---|---|
| Week 1 | Ask staff, anonymously, which AI tools and accounts they use and for what | Policy owner | A list of tools and real use cases |
| Week 1 | Check vendor terms and data processing agreements for the candidate tools | Policy owner with IT | A shortlist with the data class each tool can handle |
| Week 2 | Fill in the template, decide on DPIAs and update the record of processing | Policy owner with privacy contact | Draft policy and approved list |
| Week 2 | Discuss logging and the draft with employee representatives | Management | Comments and agreed changes |
| Week 3 | Approve the policy, set up accounts with single sign-on, switch off features you have not approved | Management and IT | Signed policy, working accounts |
| Weeks 3–4 | Train each team with exercises on their own tasks and on the data classes | Policy owner or training provider | Training record |
| Week 4 | Publish the policy and the approved list, collect acknowledgements | Policy owner | Policy in force |
| Day 30 and day 90 | Review incident reports, tool requests and questions, and adjust | Policy owner | Version 1.1 |
Training carries most of the weight. Datatilsynet's report states that it is "the responsibility of the organisation deploying M365 Copilot that users of the solution have sufficient knowledge, awareness and training in its use". It works best when people practise on their own tasks, such as removing identifying details from a real e-mail or checking a summary against its source. Our guide to prompt engineering covers the prompting part.
Example
A fictional engineering consultancy in Bergen with 35 employees finds four chat tools in use in week 1, all on personal accounts. It approves Copilot Chat in its existing Microsoft 365 tenant for internal data and buys 6 business seats of one dedicated assistant for the bid team, approved for confidential tenders but not personal data. With personal data kept out, the privacy contact finds no DPIA is needed. Three 75-minute sessions of about 12 people cover all staff in week 3. The 30-day review shows two incident reports (a tender pasted into a personal account, a meeting summary with a wrong figure sent to a client) and five tool requests, two approved. Version 1.1 adds a rule that meeting summaries are checked against the recording before they leave the company.
Does the EU AI Act require an AI policy?
The EU AI Act does not require a written AI policy for ordinary use of chat assistants, but Article 4 has required providers and deployers of AI systems to take measures on AI literacy for their staff since 2 February 2025. A policy plus a training record is the simplest way to show that you have done so.
The Digital Omnibus on AI (Regulation (EU) 2026/1744), published on 24 July 2026 and in force since 27 July 2026, rewrote Article 4. Providers and deployers must now take measures to support their staff's AI literacy, without having to guarantee any specific level for any individual. The Article 50 transparency duties, such as telling people they are dealing with an AI system and labelling deepfakes, apply from 2 August 2026, and clause 5 of the template covers them. Generative AI systems on the market before that date get until 2 December 2026 for the machine-readable marking in Article 50(2). Our guide to setting up an AI chatbot covers disclosure for customer-facing bots in more detail.
Norway has not yet incorporated the AI Act into the EEA Agreement, so it does not apply in Norway yet. A draft Norwegian AI law (KI-loven) was on consultation from 30 June 2025. On 4 August 2026 the government announced a revised draft for a new consultation in autumn 2026, aiming to present the bill to the Storting in spring 2027. Existing rules, including the GDPR, already apply to AI use in Norway.
When is a policy like this not enough?
A two-page policy is not enough when AI systems make or prepare decisions about people, process health or other special-category data at scale, or are connected to customer systems and act on their own. Those cases need a DPIA for the specific processing, a legal review and technical controls such as access limits and logging built into the system.
The same applies when AI moves into your workflows, such as an assistant that answers from your own documents. Our overview of AI automation for business explains where a person should stay in the loop, and our AI implementation service builds such systems with access control and logging from the start.
Frequently asked questions
Do we need an AI policy if only a few people use ChatGPT?
Yes, if those people handle any customer, employee or confidential information. A few people on personal accounts is exactly the situation where data ends up under terms the company never accepted. A one-page version with the data classes and an approved account is a good start.
Can employees use ChatGPT with customer data?
Only in a business account covered by a data processing agreement, for a purpose recorded in your record of processing, and with as little personal data as the task needs. Personal ChatGPT accounts should not be used with customer data. Check whether a DPIA is required before you open this use.
Is it enough to remove names before pasting text into an AI tool?
Often not. Data is anonymous only when the person cannot be identified by means reasonably likely to be used, and details such as job title, place and date can identify someone. Treat pseudonymised text as personal data unless you are confident nobody could link it back.
When the policy is approved, the next job is making it stick in daily work. Our AI training for teams includes drafting or updating the company's AI guidelines and trains each role on its own tasks, including what may and may not be entered into public AI tools.
Sources
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Articles 4, 50 and 113
- EUR-Lex: Regulation (EU) 2026/1744, Digital Omnibus on AI (Official Journal, 24 July 2026)
- European Commission: AI literacy, questions and answers (updated July 2026)
- EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 4, 5, 9, 28, 33 and 35 and Recital 26
- Datatilsynet and NTNU: Copilot through the lens of data protection, sandbox exit report (November 2024)
- European Data Protection Board: Report of the work undertaken by the ChatGPT Taskforce (May 2024)
- Regjeringen: Lov om kunstig intelligens i Norge sendes nå på høring (30 June 2025)
- Regjeringen: Tung vil sende KI-loven med endringer på høring (4 August 2026)
- Lovdata: Regulation on employer access to e-mail (FOR-2018-07-02-1108), section 2
- Lovdata: Working Environment Act, chapter 9 (control measures)
- OpenAI: Business data privacy, security and compliance
- OpenAI: Enterprise privacy
- Microsoft Learn: Data, privacy and security for Microsoft Copilot (updated October 2026)
- Anthropic: Commercial Terms of Service (effective 17 June 2025)
- Anthropic Privacy Center: Is my data used for model training?
- Anthropic: Updates to our consumer terms (28 August 2025)
- Google Workspace: Generative AI in Google Workspace Privacy Hub
- Future of Privacy Forum: What changes under the AI Omnibus (2026)
