Many organizations have an AI policy. Fewer have one that staff could describe. The usual version is several pages of legal language, written to protect the company, that nobody consults when deciding whether to paste a customer email into an assistant. A policy only works if a person can apply it in the moment.
Why do long policies fail?
People use AI tools in the middle of other work. If finding the answer to “am I allowed to do this?” takes longer than the task, they will guess. And if the official answer is always no, they will stop asking and use a personal account instead, where you have no visibility at all.
The aim is a page that makes the safe choice the convenient one.
The four questions a policy must answer
- Which tools may I use?
- What information may I put into them?
- When does a person have to check the output?
- When do I have to say AI was involved?
1. Approved tools
List them by name, with the account type that is approved, since a company account and a personal account of the same product often have different data terms. Say how to request a new tool and how long a decision takes. A slow or silent approval process is where unapproved use begins.
2. Data classes
Three classes are enough, provided each has examples from your own work.
| Class | Examples | Rule |
|---|---|---|
| Public | Published web copy, job adverts, general research | Any approved tool |
| Internal | Meeting notes, plans, code, procedures | Approved company accounts only |
| Restricted | Customer personal data, health or financial records, credentials, contract-protected material | Only in systems specifically cleared for it |
If someone is unsure which class applies, the rule is to treat it as the higher one and ask.
3. Human review
- Anything sent to a customer, regulator or the public is read and owned by a named person.
- Facts, figures and citations produced by AI are checked against a source.
- AI-written code goes through the same review as any other code.
- Decisions about a person, such as hiring, credit or claims, are not made by AI output alone.
The principle to state plainly: the person who uses the output is responsible for it. “The AI wrote it” is not an explanation.
4. Disclosure
Say when staff must tell someone that AI was used. Common cases are customers talking to an automated assistant, recorded or transcribed meetings, and content where a client contract requires it. Keep this section short and specific to your obligations.
What else belongs on the page?
- An owner, by role, who answers questions.
- One place to ask, such as a channel or an email address.
- What to do if something went wrong, written so that people report it early.
- A review date, since tools and rules change within months.
How do you get people to follow it?
Publishing the page is the smallest part. Walk each team through it using their own tasks: a recruiter with a CV, an engineer with a stack trace, an account manager with a client email. Ask them which class the data is and which tool they would use. Ten minutes of that does more than an annual acknowledgement form.
Then watch what people ask for. Requests for a tool you have not approved are information about what the business needs, not a compliance problem.
This is an outline for an internal working policy, not legal advice. Have your own counsel check it against the laws and contracts that apply to you.
