Raphael Friedmann

AI policy template

My recommendation from the IT side: one page that answers which tool, which data, who checks the result and whom to ask. Fill the highlighted fields, print it or take the Markdown into your own document. Why each line is there →

Download Markdown

AI policy of [Company]

Version [1.0] of [date]Next review: [date, at most 6 months later]Owner: [name, role]

1Purpose

We use AI tools to save time. These rules protect our data and the data of our customers. They apply to everyone who works for [Company]. Company data goes only into approved tools on company devices, never into AI apps on private devices or accounts.

2Approved tools

Tool and planUse forHighest data class
[Chat assistant, business plan with data processing agreement, no training on inputs][Writing, research, code][Yellow]
[Self-hosted model on our own servers][Internal and confidential documents][Red, with approval]

Every other tool is not approved for work, including free personal accounts, browser extensions and transcription apps. New tools are requested from [contact].

3Data classes

  • Green, any approved tool: public website text, published documents, questions without names, customers or figures.
  • Yellow, only tools with a contract: internal documents, meeting notes, non-public code, customer names in normal business context.
  • Red, only after a data protection check and written approval from [role]: health and HR data, unannounced financials, data under NDA.
  • Never, in any tool: passwords, API keys and access tokens.

Logs and config files are fine for troubleshooting once secrets are removed and personal data (names, mail addresses, IP addresses) is anonymised; they then count as Yellow. If in doubt, the higher class applies.

4Responsibility

AI output is a draft. Whoever sends, publishes, merges or signs it has checked it and owns it like their own work. Decisions about people (hiring, performance, pay) are made by a person. AI that infers the emotions of employees is prohibited (AI Act Art. 5).

5Labelling

Realistic AI images, audio and video (deepfakes) are always labelled, also internally. AI text that informs the public on matters of public interest is labelled unless a person reviewed it and holds editorial responsibility. Customer chatbots say that they are an AI. [Communications] decides edge cases.

6Agents and automation

Every agent runs with the minimum rights its task needs: its own account or API key, only the scopes and endpoints it uses, read-only where reading is enough. No AI agent gets write access to production systems, mail sending or payments without a named owner and an approval step. Content an agent reads from outside (mails, web pages, tickets) is untrusted. Built-in safety checks (auto mode, permission prompts) stay on; modes that skip every check run only in a sandbox.

Changes to systems go through code: the AI writes the change (e.g. an [Ansible playbook]), a person reviews and approves it, then the agent rolls it out. The larger the blast radius, the smaller the steps: [test system], [one production host], [all].

7Training

Before getting access to an approved tool, everyone completes the short training on this policy. Attendance is recorded (EU AI Act, Art. 4).

8Questions and requests

[Name or shared mailbox], answer within [two working days].

Before it goes live

  1. Fill every highlighted field and delete tool rows you do not need.
  2. Use the labels of your existing information classification for Green, Yellow and Red.
  3. Check every tool with your data protection officer: processing agreement (GDPR Art. 28), transfer basis for providers outside the EU (EU-US Data Privacy Framework or standard contractual clauses), entry in the record of processing, a DPIA where the risk is high.
  4. Involve the works council if there is one: in Germany, technical systems that are objectively suitable to monitor performance or behaviour need its consent (§ 87(1) no. 6 BetrVG).
  5. Run the short training and keep the attendance list.

Template as of November 2026, not legal advice.