Short answer: An AI acceptable use policy for a company of 20 to 500 people should fit on two pages. It names the approved tools, sorts data into what may and may not go into them, and makes the user answer for the output. It also sets rules for client disclosure, agents, ownership, incidents, training and review.
An AI acceptable use policy for a company of 20 to 500 people should fit on two pages and answer ten questions: which tools, which data, who checks the output, when clients are told, what agents may do, who owns the output, how to report mistakes, what training comes first, who owns the policy and when it is reviewed. Below is a full template, a one-page staff version and a two-week rollout plan. The template is a starting point to adapt with your lawyer and data protection officer.
This is general information, not legal advice.
Why does a company need an AI policy now?
Because your staff already use AI, with or without rules. In PagerDuty’s Shadow AI Survey of 1,250 office workers in the US, UK, Australia and Japan, 66% had used AI tools they believed their company did not allow. 34% had put customer data into public AI tools, and 31% financial information or confidential documents.
Respondents work at companies with $500 million or more in revenue, and PagerDuty sells AI software. Smaller firms look similar: in Microsoft’s 2024 Work Trend Index, 78% of AI users brought their own tools to work, 80% at small and medium-sized companies.
A policy alone does not fix this. In the same survey, 86% said their company had an AI policy, and 81% believed leadership plays by different AI rules. In the KPMG and University of Melbourne study of more than 48,000 people in 47 countries, 57% of employees said they hide their AI use and present AI output as their own, and only 40% said their workplace had any policy or guidance on generative AI.
In Germany, 57% of companies with 20 or more staff now use AI, according to Bitkom; among them, 66% name data protection as an obstacle and 56% legal uncertainty. Bitkom president Ralf Wintergerst summed it up: “whoever has AI in-house wrestles with data protection and the question of what is allowed” (our translation).
So the job is not to stop AI use but to give people approved tools, clear data rules and a policy that binds the managing director too. All these surveys are self-reported.
Is an AI policy legally required?
No law says “write an AI policy.” But three duties are easier to meet with one.
AI literacy (EU). Since the Digital Omnibus, Article 4 of the AI Act requires companies that use AI to “take measures to support the development of AI literacy” of their staff, without guaranteeing any specific level. The Commission’s AI literacy Q&A says no certificate is needed, an internal record of trainings “and/or other guiding initiatives” is enough, and national authorities enforce from 2 August 2026 (in Germany the Bundesnetzagentur). Some German “KI-Richtlinie Muster” pages still quote the old “sufficient level” wording, which no longer applies.
Instructions under GDPR. Anyone handling personal data under your authority may process it only on your instructions (Articles 29 and 32(4) GDPR). An AI policy is part of those instructions.
Regulators expect it. Germany’s data protection authorities, in their guidance on AI, recommend documented rules on which AI tools may be used for what, with examples of allowed and forbidden uses, and company accounts instead of private ones. Our EU AI Act guide covers the rest of the Act.
What must a short AI policy cover?
Section | The question it answers |
|---|---|
1. Scope | Who, and which tools, including AI inside existing software? |
2. Approved tools | Which tools, which accounts, how to add one? |
3. Data classes | What may go in, where, and what never? |
4. Output and accountability | Who checks, and who answers for mistakes? |
5. Disclosure | When do clients or the public need to be told? |
6. Agents and connected tools | What may AI do on its own? |
7. IP and copyright | What may go in, who owns what comes out? |
8. Incidents | What to report, to whom, how fast? |
9. Training | What comes before access? |
10. Review | Who owns it, when is it checked? |
Most free templates we reviewed cover sections 1 to 4. Few deal with agents, client disclosure or the works council, and almost none include a version staff will read.
Which data can staff put into AI tools?
Classify data, not tools. Green data may go into any approved tool, amber only into tools on company accounts with a data processing agreement, red never unless a specific system is approved for that use. Section 3 of the template defines each class.
The account matters as much as the tool. OpenAI says it does not train on business data by default on its business plans and API. On Anthropic’s consumer plans, users choose whether chats train models, with five-year retention if they agree and 30 days if not. A colleague on a personal account is on consumer terms.
Red data depends on your sector:
Healthcare and life sciences. Patient and other health data is special-category data under Article 9 GDPR. In the US, a vendor that handles protected health information for you must give “satisfactory assurance” that it will safeguard it (45 CFR 164.502(e)), in practice a business associate agreement.
Consumer health and DTC brands. Customer order data is amber; AI-written health claims need their own check, covered in AI-written health claims under EU law.
Who answers for AI output, and when do you tell clients?
The person who uses it. KPMG found that 66% of employees rely on AI output without checking it, and 56% have made mistakes at work because of AI. Courts see the result: Damien Charlotin’s database listed 2,145 decisions involving AI-invented content, mostly fake citations, as of 4 October 2026. A useful test comes from the Ghostty open-source project’s AI policy: if you cannot explain the work without the AI tool, it is not ready.
Disclosure is narrower than many templates suggest. The AI Act requires telling people they are talking to a chatbot, and labeling deepfakes and AI-generated public-interest text that no person has reviewed and taken responsibility for (Article 50). Our rule: disclose when the law, the contract or the client asks, and before client confidential information goes somewhere the engagement did not foresee.
What should the policy say about AI agents?
More than most templates do, because agents act: they send emails, update the CRM and move files. In Bitkom’s survey, 11% of German companies using or considering AI already run agents. In ISACA’s 2026 poll of more than 3,400 IT governance and security professionals, 56% did not know how long it would take to halt an AI system during a security incident.
Section 6 of the template sets three rules: no AI tool gets connected to company systems without approval, including through a “Sign in with Microsoft” screen; every agent has a named owner; and a person approves anything that leaves the company or cannot be undone. Our AI agent security checklist has the permission ladder and 15 checks before go-live.
Who owns what AI produces?
Not always you. The US Copyright Office concluded in January 2025 that AI output is protected only where a human author determined sufficient expressive elements, “but not the mere provision of prompts.” German law protects only “personal intellectual creations” (§ 2(2) UrhG). So keep a person shaping anything you need to own, such as logos, and keep content you have no rights to out of prompts.
The AI acceptable use policy template
Copy it, fill in the brackets and cut what does not apply. It is a starting point to adapt with your lawyer and data protection officer, not a finished legal document.
AI Acceptable Use Policy, [Company]. Version [1.0] · Owner: [name, role] · Approved by: [managing director] · In force from: [date] · Next review: [date, at most six months away]
1. Scope. This policy applies to everyone who works for [Company], including management, freelancers and temps, on any device. “AI tool” includes AI features inside software we already use, such as [Microsoft 365 Copilot / our CRM assistant]. We want you to use AI where it helps; this policy sets the limits.
2. Approved tools. For company work, use only the tools on the approved list at [link], signed in with your company account. Personal or free accounts may be used for green data only. To request a new tool, write to [owner]. We decide within [five working days] after checking the data processing terms, training use, data location and cost.
3. Data.
Green, any approved tool: public information; your own text without personal, client or confidential details.
Amber, company accounts only: internal documents; business contact details; contracts; source code; non-public figures not listed as red.
Red, never, unless [owner] has approved a specific system for that use in writing: patient and health data; other special-category data; [privileged or confidential client material]; HR, applicant and performance data; passwords, keys and tokens; payment and bank data; [unannounced results, deal information].
Unsure? Treat it as red and ask. Remove names when the task works without them.
4. Output and accountability. You answer for anything you produce with AI, as if you had written it yourself. Before AI output reaches a client, the public or a decision about a person, check facts, figures, quotes and citations against the source, and make sure you can explain it without the tool. [Sector check: health claims against [list] / citations in [database].] AI does not make final decisions on hiring, pay, promotion, termination, credit or patient care, and we do not use it to read the emotions of staff or applicants.
5. Disclosure. We tell people when they talk to a chatbot or voice agent we run, and label AI-made images, audio and video that could pass as real. We disclose AI use when a client’s contract requires it, when a client asks, and before client confidential information goes into a tool the engagement did not cover. If asked whether AI was used, answer truthfully.
6. Agents and connected tools. Do not connect any AI tool, plugin or agent to company email, files, the CRM or other systems without approval from [IT owner]. Every agent has a named owner, its own account and the narrowest access its job needs. Agents may read and draft; a person approves before an agent sends anything outside, deletes, shares files externally, pays or changes access. [Owner] can stop any agent within [15 minutes].
7. IP and copyright. Do not put in third-party material we have no right to use, such as paid reports or licensed images, beyond short quotations, and do not ask a tool to copy a named creator’s work. Ask [owner] before using AI-generated elements in logos, brand assets or core product material.
8. Incidents. Report to [contact or channel] within [24 hours] if confidential, client or personal data went into an unapproved tool; wrong AI output reached a client or the public; an agent did something unexpected; or a tool seems to have been manipulated by text it read. Report your own mistakes too; fast reporting is what we want. Some data breaches must reach the authority within 72 hours.
9. Training. Before getting access to approved tools, everyone completes [45 minutes] of training for their role: how the tools work, where they fail, this policy and examples from our own work. [HR] records who attended and when.
10. Review and contacts. [Owner] reviews this policy by [date], and whenever we add a major tool or agent or the law changes. Questions: [contact]. Breaches are handled under [policy], in proportion to the harm. [Germany: works council informed on [date] / agreement [reference].]
The one-page staff version
Put this on the intranet and on top of the policy.
AI at [Company]: the short version
Use AI. We want you to, with the approved tools at [link] and your company account.
Public information and your own text: fine in any approved tool.
Internal documents, contacts, contracts, code: company accounts only.
Patient, health, HR and client-confidential data, passwords and payment data: never, unless a specific system is approved for it.
You own what you send. Check facts, figures and sources, and be able to explain it without the tool.
AI never makes final decisions about people.
Do not connect AI tools to email, files or the CRM without asking [IT owner].
Customers are told when they talk to a bot. Clients are told when they ask or their contract requires it.
Something went wrong? Tell [contact] within 24 hours. Reporting your own mistake is the right call.
Questions or a new tool request: [contact]. These rules apply to everyone, including management.
How do you roll it out in two weeks?
Days 1 to 2: find out what is in use. Run a short anonymous survey (which tools, for what) and check expense claims for AI subscriptions. Announce an amnesty for whatever people report.
Day 3: choose the approved tools. Usually one general assistant on business accounts plus AI features in software you already have. Check processing terms, training settings and data location.
Days 4 to 5: fill in the template. Data protection officer and lawyer review it. In Germany, inform the works council now.
Days 6 to 8: train by role. 45 minutes per team, built on their own tasks. Record attendance.
Day 9: publish. Policy, staff page, tool request route and incident channel go live together.
Day 10: close the gap. Retire unapproved tools that now have an approved alternative, and set the review date.
If a works council agreement is needed, the timeline depends on those talks. For the rest of year one, see our AI implementation roadmap.
Do you need the works council in Germany?
Often for the tools, if not for the policy itself. Under § 87(1) no. 6 BetrVG, the works council co-determines the introduction and use of technical systems designed to monitor employees’ behavior or performance. In a 2024 decision involving a medical technology maker, the Hamburg labor court held that ChatGPT guidelines concerned how work is done, not workplace order under § 87(1) no. 1, and that private accounts gave the employer no usage data, so no. 6 did not apply. That was one interim ruling at first instance.
Company-provided tools are different: business accounts, Copilot and agents usually log who did what, which can bring no. 6 into play. The employer must also inform the works council in good time when planning work processes that use AI (§ 90(1) no. 3), and the council may bring in an expert to assess AI (§ 80(3)).
A works council agreement may be needed; GDPR expressly allows such agreements to set rules on employee data (Article 88 and Recital 155). Involve the council early and let your employment lawyer judge your case.
Where to start
This week, run the survey and pick the approved tools. Next week, fill in the template, train each team and publish. In our view, a policy that names the tools people may use does more than one that only lists what they may not.
Policy out, and ready to put AI into one real workflow? Book a fit check. Projects start at €10,000 (from $12,000 in the US). Next, read AI agent security: what to check before an agent touches your email, files or CRM and the EU AI Act in 2026.
Frequently asked questions
Is an AI policy legally required?
No law says to write an AI policy. It does make three duties easier to meet: the AI literacy measures under Article 4 of the AI Act, the instructions staff need under GDPR when they handle personal data, and the documented rules on AI tools that German data protection authorities recommend.
Which data can staff put into AI tools?
Classify data, not tools. Green data, such as public information, may go into any approved tool. Amber data, such as internal documents and contracts, goes only into company accounts with a data processing agreement. Red data, such as patient, HR or payment data and passwords, never goes in unless a specific system is approved for it.
What should an AI policy say about AI agents?
Three rules: no AI tool gets connected to company systems without approval, every agent has a named owner, and a person approves anything that leaves the company or cannot be undone. Each agent also gets its own account and the narrowest access its job needs.
How long does it take to roll out an AI policy?
About two weeks. Survey what is in use, choose the approved tools and fill in the template with your data protection officer and lawyer. Then train each team by role, publish, and retire unapproved tools. If a works council agreement is needed, the timeline depends on those talks.
Do you need the works council for AI in Germany?
Often for the tools, if not for the policy itself. Company tools such as Copilot and agents usually log who did what, which can bring co-determination under § 87(1) no. 6 BetrVG into play. Involve the council early and let your employment lawyer judge your case.
Related reading
Sources
PagerDuty. “PagerDuty Report Finds Two-Thirds (66%) of Office Professionals Have Used Unauthorized AI Tools at Work.” Press release, June 11, 2026. Survey by Wakefield Research of 1,250 office professionals. https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/
Microsoft and LinkedIn. “2024 Work Trend Index Annual Report: AI at work is here. Now comes the hard part.” May 8, 2024. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
KPMG. “Global study reveals trust of AI remains a critical challenge reflecting tension between benefits and risks.” Press release, April 28, 2025, on the University of Melbourne and KPMG study “Trust, attitudes and use of artificial intelligence: A global study 2025.” https://kpmg.com/xx/en/media/press-releases/2025/04/trust-of-ai-remains-a-critical-challenge.html
Bitkom. “Erstmals nutzt die Mehrheit der Unternehmen KI.” Press release, September 14, 2026. Telephone survey of 603 companies with 20 or more employees in Germany. https://www.bitkom.org/Presse/Presseinformation/Erstmals-nutzt-Mehrheit-Unternehmen-KI
ISACA. “AI Use Accelerates, While Governance and ROI Lag, Says New ISACA Research.” Press release, May 5, 2026. https://www.isaca.org/about-us/newsroom/press-releases/2026/ai-use-accelerates-while-governance-and-roi-lag-says-new-isaca-research
Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1(5), replacing Article 4 of the AI Act. Official Journal, July 24, 2026. https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 5 and 50. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
European Commission. “AI Literacy: Questions & Answers.” Last updated July 27, 2026. https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers
Bundesnetzagentur. “Bundesnetzagentur übernimmt zentrale Rolle bei der Umsetzung der KI-Verordnung.” Press release, July 29, 2026. https://www.bundesnetzagentur.de/SharedDocs/Pressemitteilungen/DE/2026/20260729_KI_VO.html
Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 9, 29, 32(4), 33, 88 and Recital 155. https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679
Datenschutzkonferenz (German federal and state data protection authorities). “Orientierungshilfe Künstliche Intelligenz und Datenschutz,” version 1.0, May 6, 2024. https://www.datenschutzkonferenz-online.de/media/oh/20240506_DSK_Orientierungshilfe_KI_und_Datenschutz.pdf
OpenAI. “Enterprise privacy at OpenAI.” Updated January 8, 2026. https://openai.com/enterprise-privacy/
Anthropic. “Updates to Consumer Terms and Privacy Policy.” August 28, 2025. https://www.anthropic.com/news/updates-to-our-consumer-terms
US Code of Federal Regulations. 45 CFR 164.502(e), disclosures to business associates. https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.502
Damien Charlotin. “AI Hallucination Cases Database.” Last updated October 4, 2026. https://www.damiencharlotin.com/hallucinations/
Ghostty project. “AI Usage Policy.” GitHub, accessed October 4, 2026. https://github.com/ghostty-org/ghostty/blob/main/AI_POLICY.md
Betriebsverfassungsgesetz (German Works Constitution Act), §§ 80, 87 and 90. https://www.gesetze-im-internet.de/betrvg/__87.html
Arbeitsgericht Hamburg, 24th Chamber. Order of January 16, 2024, 24 BVGa 1/24. https://www.landesrecht-hamburg.de/bsha/document/NJRE001564562
US Copyright Office. “Copyright Office Releases Part 2 of Artificial Intelligence Report.” NewsNet Issue 1060, January 29, 2025. https://www.copyright.gov/newsnet/2025/1060.html
Urheberrechtsgesetz (German Copyright Act), § 2. https://www.gesetze-im-internet.de/urhg/__2.html