What an enterprise AI policy has to contain to be worth writing
A short AI policy is enough if it answers the questions that come up in practice, in language an employee can apply without calling legal. Ten sections do that.
What the law requires, and what staff already do
The legal floor is Article 4 of the AI Act, applicable since 2 February 2025:
Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff ... This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
Regulation (EU) 2024/1689, Article 4(1), as replaced by Regulation (EU) 2026/1744
Microsoft and LinkedIn's 2024 Work Trend Index, published on 8 May 2024 from 31,000 knowledge workers across 31 markets, found that “75% of knowledge workers use AI at work today” and that “78% of AI users are bringing their own AI tools to work (BYOAI)”. The policy therefore governs a habit that staff already have.
The ten sections a policy needs
- 1. Which tools are approved. Name them, in a list short enough for staff to recall.
- 2. What data may be entered. Use classes, not adjectives: public, internal, confidential, restricted, with one line each on what is allowed where.
- 3. Who may use what. Different roles have different permissions, and the system enforces them, not the document.
- 4. Which outputs need a human check. Be specific about what triggers review: anything sent to a client, anything that becomes a commitment, anything that touches a regulated process.
- 5. What the AI may decide alone. Usually the answer is nothing that affects a person's rights, employment, credit or care. Write it down anyway.
- 6. How usage is monitored. The organisation should be able to state which AI systems are in production. The inventory you need to classify your systems under the AI Act asks the same question.
- 7. What happens when the model is wrong. A reporting route with no blame attached, because the alternative is silence.
- 8. What happens when the provider changes. Plan for model retirement, price changes and terms changes as operational events.
- 9. Who owns AI governance. One named role, not “IT and legal”.
- 10. What happens to company data. For each approved system: processing location, retention, whether inputs train anything, subprocessors, who may access, how deletion works, and the security controls.
Why a ban is the weakest option
A prohibition competes with a tool that saves an employee time every day, and it loses quietly: the use continues and the visibility ends. A policy that offers a route faster than the unapproved one keeps both the productivity and the audit trail. Its shortest form: use this system for this class of work, with this class of data, and ask before you go outside it.
With Bastion
What Bastion changes
Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.
An internal system makes sections 1, 2, 3, 6 and 10 easier to write, because they can describe one named tool with one processing location and one set of permissions instead of a shifting list of external services.
Sections 4, 5, 7, 8 and 9 stay the organisation's work.
Questions this article answers
- What should a company AI usage policy contain?
- Ten things: approved tools, permitted data classes, who may use what, which outputs need human review, what AI may decide alone, how usage is monitored, what happens when it is wrong, what happens when the provider changes, who owns governance, and what happens to company data.
- Is an AI policy a legal requirement?
- The policy itself is not named in the AI Act, but Article 4 requires providers and deployers to take measures that support their staff's AI literacy, and it has applied since 2 February 2025. A policy is how that duty becomes something an employee can follow.
- Should the policy ban AI tools that are not approved?
- A prohibition competes with a tool that saves time every day and loses quietly. The stronger version names one system for defined work and defined data classes, and asks people to check before going outside it.
On the record
- 1
EUR-Lex
Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligenceread 2026-09-22
- 2
- 3
Microsoft WorkLab
AI at Work Is Here. Now Comes the Hard Part. 2024 Work Trend Index Annual Reportread 2026-09-22
- 4
Read next
- Shadow AI: the tool arrives before the policy doesShadow AI is what happens when the demand for a tool arrives before the approved version of it. The measured scale, why a ban produces the worst version of the outcome, and what an approved path has to contain.
- AI literacy is already an obligation, and it is not a training courseWhat the Article says, what the Regulation's own recital says it is for, why the duty is role-based rather than uniform, and a five-module programme a small company can run without a consultant.