Most policy documents in this area fail for the same reason: they are written to protect the organization legally, not so that an employee knows what to do on Tuesday at ten o'clock. The result is a document that gets signed and not read. The direct consequence is that employees keep using AI tools at their own personal discretion, with no connection to what the document says — and in effect the policy exists only on paper.
Seven Sections, and No More
What may and may not be entered — as a list
Not "sensitive information" but a concrete list: customer names, account numbers, medical information, source code, legal documents. An employee should be able to check against the list in two seconds. Build this list together with the professionals on the ground, not only with legal counsel — they know exactly which data passes through day-to-day work.
Which tools are approved
A named list, and who approves additions. "Approved tools only" without a list is an empty instruction.
Organizational accounts only
The most important distinction, and the one employees are least aware of. Explain why, not just the rule.
Responsibility for the output
Whoever uses the output is responsible for it. AI makes mistakes with complete confidence, and "the system said so" is not an explanation to a customer.
When AI use must be disclosed
To customers, in official documents, and in external correspondence. ⚠️ In Europe this is already a legal requirement for branded chatbots.
Agents and automated actions
Which actions an agent may perform on its own, and which require human approval. Payments, deletion and sending outside the organization are a reasonable starting point.
Who to contact
A name and an address. An employee who has found an interesting tool needs to know whom to ask — otherwise they will simply use it.
What Gets a Policy Read
- Two pages, not twenty.
- Examples from the organization's day-to-day work, not generic wording.
- An explanation of "why", not just "what" — an instruction that is understood gets applied; an instruction that is imposed gets bypassed.
- A review date. The field moves fast, and a two-year-old policy does more harm than having none.
A policy that is not accompanied by an approved alternative is a list of prohibitions. It will drive usage underground and create an illusion of control — which is worse than the starting point. The simplest fix is to pair every prohibition with a concrete approved alternative — not only "do not use unapproved tools", but also "here is the approved tool that meets the same need".
What This Guide Does Not Cover
- This is a recommended structure, not binding wording. A policy document meant to bind employees requires review by legal counsel and human resources.
- Organizations in regulated sectors will need additional sections according to their sector regulator.
Want to Go Through It Together?
A guide is a starting point. An assessment of your environment will give you a precise answer.
