Company AI policy and acceptable use (approved tools, shadow AI, confidential data)

Your staff are already using ChatGPT and similar tools, whether you have allowed it or not. This guide helps you get from "nobody knows" to a two-to-four-page AI policy. The policy names the approved tools, puts data into classes that say what may go where, sets clear duties for users, and is backed by an actual company tool so people have no reason to use private accounts.

Why it matters

• Shadow AI is already common. Bitkom surveyed 604 German companies with 20 or more employees in mid-2025. In 8% of them, staff widely use private AI tools at work (2024: 4%). Another 17% reported individual cases, and a further 17% suspect it without knowing for sure. That makes roughly 4 in 10 companies. Only 26% give employees official access to generative AI, and only 23% have AI rules (2024: 15%) [1].
• Data leakage and loss of control. The BSI lists missing confidentiality of entered data, dependency on the provider, hallucinated output and AI-specific attacks among the main risks of generative AI models. It recommends a risk analysis for each concrete use case before deployment [2][3].
• GDPR accountability stays with you. According to the German data protection authorities (DSK), an organisation that uses an LLM for its own purposes and allows staff to use it is generally the controller. It must therefore settle the legal basis, the processor agreement and internal rules [4]. Personal data typed into a free private account is processed without an AVV (Art. 28 GDPR) and often with training use enabled.
• AI Act Art. 4. Since 27 Jul 2026, deployers must take measures to support the AI literacy of their staff. The law does not require a guaranteed level. The Commission says it will not prescribe a fixed "sufficient level" [5]. A written policy plus a briefing is the cheapest documented measure you can show (see GUIDE-006).
• Trade secrets. Under the GeschGehG, information only counts as a protected trade secret if you have taken "appropriate confidentiality measures". A documented rule on what may not go into external AI tools is one of those measures.

When you need it

• Now, if you have no rule at all. Shadow AI does not wait for your AI strategy.
• Before you roll out any company AI tool (M365 Copilot, a ChatGPT/Claude/Le Chat business plan, an internal RAG assistant). The policy and the tool should go live together.
• When a new class of tool arrives: AI meeting recorders, browser extensions, coding assistants, AI agents with access to email or ERP.
• At least once a year, and whenever the legal dates change (AI Act Art. 50 transparency since 2 Aug 2026, Annex III high-risk duties from 2 Dec 2027; see GUIDE-014).

Who is involved

Step by step

1. Find out what is actually used (1–2 weeks, low effort). Run a short anonymous survey: which tools, for which tasks, with which data. Add a look at proxy/firewall logs or SaaS discovery for AI domains. Report the results only in aggregate, and talk to the works council first if you analyse logs. You are looking for demand, not culprits. Typical findings: email drafting, translation, summarising, Excel formulas, code snippets, image creation.

2. Provide an approved alternative before you prohibit (2–6 weeks). A ban without an alternative pushes use onto private phones. Choose at least one general-purpose assistant on a business or enterprise plan with: - an AVV (DPA) and EU data residency or standard contractual clauses, - contractual exclusion of training on your inputs, - SSO and company-managed accounts (no private accounts), - admin controls for retention, connectors and sharing. Use GUIDE-002 for the build-vs-buy choice and GUIDE-010 for vendor checks.

3. Classify data in plain language (1 day workshop). Most SMBs already have, or need, three or four classes. Map each class to the tools it may go into:

4. Write the policy (2–4 pages, not 20). Recommended structure: 1. Purpose and scope (all staff, temps, contractors; company and private devices used for work). 2. Approved tools list. Keep it as a separate, frequently updated page, with an owner. 3. Prohibited: private accounts for work data, unapproved browser extensions and AI meeting bots, uploading data above the permitted class, fully automated decisions about people. 4. Data rules (the table above) plus "anonymise first" guidance (see DOC-018). 5. User duties: check output before using it (facts, figures, legal statements, code), take responsibility as the author, no AI-generated content passed off as verified expertise. 6. Transparency: label AI-generated content where Art. 50 AI Act requires it (e.g. chatbots facing customers, deepfakes, AI text published on matters of public interest without human editorial review). Also label where your customers or contracts require it. 7. Copyright and confidentiality: no uploading of third-party material you have no rights to, and respect NDAs. 8. Requesting new tools: a simple form (purpose, data class, vendor, cost) with a 2-week decision target. 9. Incidents: whom to tell if confidential data went into the wrong tool. No blame for fast reporting. 10. Training and contact person. Consequences of breaches. Review date.

5. Clarify co-determination and data protection (parallel to step 4). In 2024 the Hamburg labour court (ArbG Hamburg, 16 Jan 2024, 24 BVGa 1/24, interim proceedings) found no co-determination right when staff merely were allowed to use ChatGPT via their own private accounts and the employer received no usage data [6]. Do not generalise this. Once you provide company accounts whose logs or admin reports could show who did what, the BAG's "objective suitability" test applies. §87(1) no. 6 BetrVG is then normally triggered, as with M365 Copilot. Rules on conduct can also touch §87(1) no. 1. Plan a framework works agreement on AI instead of negotiating tool by tool (GUIDE-004). The vbw (Bavarian employers' association) has published a model works agreement on AI use [7]. Have the DPO record the processing in the record of processing activities and decide whether a DPIA is needed.

6. Launch with a briefing, not an email (1–2 hours per group). Run a 45–60 minute session per team: the approved tools, the data-class table with real examples from your company, and 3–4 good prompts for their work. Have staff confirm they have read the policy. This also documents your Art. 4 measures (GUIDE-006). Name AI champions per department (GUIDE-008).

7. Enforce proportionately and review. Technical measures: block known high-risk tool categories (e.g. unvetted AI browser extensions, transcription bots), and use DLP rules for clearly marked confidential files where your licences allow. Organisational measures: quarterly review of the approved-tools list, a yearly policy review, and a review after every incident. Track the numbers: active users of the approved tool, open tool requests, incidents.

Realistic effort for a 50–250 employee company: 8–15 person-days internally over 6–10 weeks, plus licence costs for the approved tool. Add more time for works-council negotiations.

Checklist

• [ ] Anonymous survey of current AI use done; results shared in aggregate
• [ ] At least one approved AI tool with AVV, no-training clause, SSO, company accounts
• [ ] Data classes defined and mapped to tools, with company-specific examples
• [ ] Policy (2–4 pages) approved by management, DPO, IT security, HR
• [ ] Works council informed early (§90); works agreement concluded where §87(1) no. 6 applies
• [ ] Record of processing activities updated; DPIA decision documented
• [ ] Tool request process with owner and decision deadline
• [ ] Incident reporting path known to everyone, with a no-blame rule for quick reports
• [ ] Briefing held, acknowledgement collected, Art. 4 measures documented
• [ ] Approved-tools list reviewed quarterly; policy reviewed yearly

Templates & tools

• IHK Schwaben – Muster-KI-Richtlinie: free model AI policy (purpose, approved programmes, data protection, responsibility) [8].
• vbw – Muster-Betriebsvereinbarung zur Nutzung von KI (July 2025) [7].
• DSK – Orientierungshilfe "Künstliche Intelligenz und Datenschutz" (May 2024): checklist-style criteria for selecting and using LLM-based tools [4].
• BSI – "Generative KI-Modelle: Chancen und Risiken für Industrie und Behörden" (updated Jan 2025, DE/EN): risk catalogue and countermeasures, useful as the basis for the risk section of your policy [2][3].
• EU Commission – AI literacy Q&A and living repository of practices [5].
• The Mittelstand-Digital Zentren and your local IHK offer free introductory workshops on AI governance.

Common pitfalls

• Ban without an alternative. This leads to more shadow AI, not less. Fix: approved tool first, then rules.
• Policy written by lawyers for lawyers. Twenty pages nobody reads. Fix: 2–4 pages, a data-class table, and real examples ("an offer to customer X: only in the approved tool").
• Approved list never updated. Staff go around a list that is out of date. Fix: one owner, a quarterly review, a 2-week decision target for tool requests.
• Relying on the Hamburg decision. It covered private accounts with no employer access to data. Company tools with logging are a different case.
• Forgetting embedded AI. AI features switched on inside existing software (Office suite, CRM, DMS, video calls) count too. Include them in the inventory and the approved list.
• Punishing honest reporters. If reporting a mistake leads to sanctions, you will stop hearing about incidents.
• Treating the policy as the whole of AI literacy. Signing a document does not build skills. Combine it with training (GUIDE-006).

Sources

1. Bitkom: "Beschäftigte nutzen vermehrt Schatten-KI", press release, 21 Oct 2025 (survey of 604 companies ≥20 employees, July–Aug 2025). https://www.bitkom.org/Presse/Presseinformation/Beschaeftigte-nutzen-vermehrt-Schatten-KI
2. BSI: "Generative KI-Modelle – Chancen und Risiken für Industrie und Behörden". https://www.bsi.bund.de/SharedDocs/Downloads/DE/BSI/KI/Generative_KI-Modelle.html
3. BSI: "Aktualisierte Fassung veröffentlicht: Chancen und Risiken generativer KI-Modelle", 21 Jan 2025. https://www.bsi.bund.de/DE/Service-Navi/Presse/Alle-Meldungen-News/Meldungen/Chancen-Risiken-generative-KI_250121.html
4. Datenschutzkonferenz (DSK): Orientierungshilfe "Künstliche Intelligenz und Datenschutz", press release 6 May 2024. https://datenschutzkonferenz-online.de/media/pm/2024_05_06_DSK_PM_OH_KI_und_Datenschutz.pdf
5. European Commission: AI Literacy – Questions & Answers. https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers
6. ArbG Hamburg, decision of 16 Jan 2024, 24 BVGa 1/24, summarised by CMS: "Kein Mitbestimmungsrecht des Betriebsrats bei ChatGPT & Co." https://cms.law/de/deu/legal-updates/kein-mitbestimmungsrecht-des-betriebsrats-bei-chatgpt-co
7. vbw: Muster-Betriebsvereinbarung zur Nutzung von KI (July 2025). https://www.bds-bayern.de/wp-content/uploads/2026/05/Muster-Betriebsvereinbarung-zur-Nutzung-von-KI-Juli-2025-vbw.pdf
8. IHK Schwaben: KI-Richtlinie (Muster). https://www.ihk.de/schwaben/produktmarken/digitalisierung/kuenstliche-intelligenz/ki-richtlinie-6644000

Accessed 2026-10.