# An AI policy for your company: one page people can follow
> Why a ban on AI tools only moves use to private accounts, and what a one-page usage policy must say: tools, data classes, review duty, agents, AI literacy. With a skeleton to copy.
Source: https://rfriedmann.de/blog/an-ai-policy-for-your-company/
Published: 2026-11-21 · Track: learn · Level: Ops
At some point every IT lead gets the request: "we need an AI policy". My advice from
the IT side: keep it to one page. A policy is read once, at rollout, and then looked
up when someone is unsure, so it has to answer four questions at a glance: which tool,
which data, who checks the result, and whom to ask. This post suggests that page, with
the reasoning behind each line and the legal hooks as they stand in November 2026.
## Why a ban does not work
The tempting first draft is one sentence: "AI tools are not permitted." It is short,
it feels safe, and it does not stop anyone. It moves the use to private phones and
personal accounts, where you see nothing.
In Microsoft and LinkedIn's [2024 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part),
a survey of 31 000 knowledge workers in 31 markets, 78% of the people who used AI at
work brought their own tools. That is a vendor's survey, so treat the exact figure with
care, but the pattern is what matters: people use what helps them, approved or not.
A ban removes the conversation, not the use. The better order is the reverse: approve
one tool that is good enough that the private account has no advantage, then write the
rules for that tool.
## What the page has to say
Seven points. Each one should fit in two or three lines on the final page.
**1. Approved tools and tiers.** Name the products, and name the tier. The same chatbot
can come as a free consumer account and as a business plan with a data processing
agreement and training on your inputs switched off. Only the second belongs on the
list, and for a provider outside the EU only with a transfer basis under the GDPR: a
company certified under the EU-US Data Privacy Framework, or standard contractual
clauses. If you run a [self-hosted model](/blog/self-hosting-an-llm-for-your-team/) for
data that must not leave the house, list it too, with what it may be used for.
Everything not on the list is not approved, and the policy says how to request an
addition.
**2. Data classes.** This is the part people will actually remember, so keep it to
three colours with examples they recognise from their own work.
Which data may go into which tool (example classification)
[diagram omitted — see the page for the chart]Illustrative. Your classes should reuse the labels of your existing information classification, so nobody has to learn a second scheme.
Passwords, API keys and access tokens get their own line: they never go into a prompt,
in any tool, for any reason. Someone will paste a config file to ask why a service does
not start. The policy has to name that case, and allow it: logs and config files are some
of the most useful things to hand a model, once secrets are removed and personal data
such as names, mail addresses and IP addresses is anonymised. A ban on logs would push
exactly the troubleshooting that saves the most time back to private accounts.
**3. Review duty.** The model's output is a draft. The person who sends, merges or
signs it is responsible for it, exactly as if they had written it themselves. Models
[produce confident errors](/blog/why-models-make-things-up/), and some tasks are
[better done without AI at all](/blog/when-not-to-use-ai/). One sentence covers it:
"You check it, you own it."
Decisions about people need a line of their own. The GDPR (Art. 22) restricts decisions
based solely on automated processing that have legal or similarly significant effects
on a person. AI used for hiring or performance evaluation is high-risk under the AI Act,
with duties for the deploying company from 2 December 2027. And AI that infers the
emotions of employees at work has been prohibited since 2 February 2025 (Art. 5(1)(f)),
apart from medical or safety reasons. The policy says it plainly: a person decides about
people, and emotion recognition is out.
**4. Labelling.** Most AI-assisted work needs no label. Two cases do. Since 2 August 2026,
[Article 50 of the AI Act](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)
requires deployers to disclose deepfakes (image, audio or video that looks real), and
that includes internal use: the Commission's guidelines on Article 50 name a corporate
newsletter as an example. It also requires a label on AI-generated text published to
inform the public on matters of public interest, unless a human reviewed it and someone
holds editorial responsibility; texts that stay inside the organisation do not count as
published. And a chatbot
in front of customers has to tell people they are talking to an AI unless that is
obvious, a duty on whoever builds it. Your policy names who decides these cases,
usually communications.
**5. Agents and automation.** A chatbot answers; an agent acts. The policy draws the
line: no agent gets write access to production systems, mail sending or payments
without a named owner and an approval step. Below that line, every agent gets the
minimum rights its task needs: its own account or API key, scoped to the endpoints it
actually calls, read-only where reading is enough, and never the admin token of the
person who set it up. An agent cannot misuse a permission it does not have, and its own
key shows in the logs what it did and can be revoked alone. Anything an agent reads from outside, a
mail, a web page, a ticket, can carry instructions to it, which is
[prompt injection](/blog/prompt-injection/).
Modern agents bring their own guard: an auto mode or permission prompts, where a second
check looks at each action before it runs and stops the risky ones, a mass delete, a
force push, data sent somewhere it should not go. Keep it on. It is a model judging a
model, though, so it catches most mistakes, not all of them. It is one more layer on top
of minimum rights and the approval step, not a replacement for them. The opposite switch,
the mode that skips every check, belongs in a sandbox with nothing real to break. How much autonomy is reasonable is an
operations question I answered [separately](/blog/letting-an-agent-loose/); the policy
only needs the rule that it is decided per case, not per enthusiasm.
For changes to infrastructure, this is the flow I like to use:
The AI writes the change as code
An Ansible playbook, not a command it runs itself. The change exists as a diff before anything touches a system, and ansible-playbook --check --diff shows what it would do.
A person approves it
The review happens on the playbook and the dry-run output, where it is cheap: one diff, read once, instead of watching a live session.
The agent rolls it out
Only the approved playbook, with the rights to run exactly that and nothing else.
Step size follows the blast radius
A change to one service on one host goes out in one run. A change that touches every server goes to a test system first, then one production host, then the rest.
The human gate sits where it costs the least time and stops the most damage, and the
agent still does the writing and the repetitive rollout. The template below puts this
into one line.
**6. Whom to ask.** One named contact or one shared mailbox, for "may I put this in?"
and "can we get tool X?". If the answer takes weeks, people stop asking and go back to
their private accounts.
**7. Review cycle.** Tools, prices and laws change every few months. The page carries a
version date and a fixed review date, every six months or when a listed tool or the law
changes, whichever comes first.
## The legal hook: AI literacy
You do not need a legal reason to write this page, but there is one.
[Article 4 of the AI Act](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4)
requires providers and deployers of AI systems to "take measures to support the
development of AI literacy" of their staff, taking into account their knowledge and
the context of use. Any company whose staff use an approved AI tool for work is a
deployer.
The duty has applied since 2 February 2025. Regulation (EU) 2026/1744, the "Digital
Omnibus on AI", amended it in July 2026: the text now says explicitly that it "does not
require providers or deployers to guarantee any specific level of AI literacy of any
individual". According to the Commission's
[AI literacy Q&A](https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers),
national market surveillance authorities supervise it from 2 August 2026, no certificate
is required, and organisations can keep an internal record of trainings to show what
they did.
In practice: the one-page policy, a short training on it, and a list of who attended.
That is a documented measure. The data protection side, GDPR for whatever goes into a
prompt, is a separate law with its own checklist, covered in
[AI and the law](/blog/ai-and-the-law-gdpr-ai-act/).
In Germany there is a third party to involve: the works council, where one exists.
Under § 87(1) no. 6 BetrVG it has a say on technical systems intended to monitor
performance or behaviour, and the Federal Labour Court reads "intended" as objectively
suitable, whatever the employer plans to do with it. A company-run AI tool that logs
who asked what is typically such a system. The Hamburg labour court decided in January 2024
(24 BVGa 1/24, summary proceedings) that allowing ChatGPT through employees' own
browser accounts did not need consent, because the employer had no access to that data.
Once the company provides the tool, plan the works council in.
## A one-page skeleton to copy
| Section | What to write |
|---|---|
| Purpose | One sentence: we use AI tools to save time; these rules keep our data and our customers' data safe. |
| Approved tools | Product, tier, what it is for, highest data class allowed. Everything else needs a request. |
| Data classes | Green, yellow, red with three examples each. Passwords, keys and tokens: never, in any tool. Logs and configs with secrets removed and personal data anonymised: yellow. |
| Your responsibility | AI output is a draft. Whoever sends, merges or signs it owns it and has checked it. A person decides about people; no emotion recognition of staff. |
| Labelling | Deepfakes are always labelled, also internally. Published public-interest texts are labelled unless reviewed and edited by a person. Customer chatbots say they are AI. Communications decides edge cases. |
| Agents and automation | Minimum rights: own account or API key per agent, only the scopes it uses. No write access to production, mail sending or payments without a named owner and approval. Treat external content an agent reads as untrusted. Changes as code, approved by a person, rolled out in steps sized by blast radius. Built-in safety modes stay on; skip-all modes only in a sandbox. |
| Training | Everyone with access to an approved tool completes the short training; attendance is recorded (AI Act Art. 4). |
| Questions and requests | Contact or mailbox, expected answer time. |
| Version | Date of this version, date of the next review (at most six months later). |
That table is the policy. Anything longer belongs in an appendix that the people who
need it can find.
The same structure written out as a policy with fields to fill, printable on one A4
page or as Markdown: [AI policy template](/ai-policy/).
## Bottom line
A ban does not stop AI use, it hides it. A one-page policy that approves a decent tool
and then says which data goes where, who checks the output, what agents may do, and whom
to ask, is something people can follow. It also gives you the documented measure that
Article 4 of the AI Act asks for. Write it short, date it, and review it twice a year,
because the tools and the rules will both have moved by then.