Skip to content

How to write an AI policy your staff will actually follow

Short answer

Keep it to one page: a list of approved AI tools, the data that never goes into any of them, and who to ask about a new tool. Microsoft's 2026 Data Security Index found that 32% of data security incidents at surveyed organizations involved generative AI tools, so write the policy before your staff write their own rules.

Key takeaways

  • One page beats a 12-page policy nobody opens.
  • Name the approved tools, and name the data that never goes in.
  • Copilot can see whatever the person can see, so fix permissions before the rollout.
  • Train staff with examples, then review the policy every few months.

Your staff are already using AI. Somebody in accounting has pasted a spreadsheet into a chatbot this month, and nobody told them not to. We've been doing managed IT in New York since 2001, and this is the fastest-moving gap we see.

That's not a character flaw. In a 20-person office with no written rule, everyone guesses, and the person who guesses wrong is usually the one in a hurry. Here's how we'd write a rule that people follow.

Why bother with a policy at all?

Microsoft's 2026 Data Security Index surveyed more than 1,700 security leaders. It found that 32% of data security incidents at those organizations involved generative AI tools, and that 47% of the leaders were adding controls aimed at generative AI. The survey is Microsoft's, so read it as a signal, not a law of nature. The direction is still clear. The same report says that both unsanctioned and sanctioned AI tools must be managed.

Microsoft Data Security Index 2026: generative AI and data security
  • Incidents involving generative AI tools32 %
  • Leaders adding generative AI controls47 %

Source: Microsoft Data Security Index 2026, 1,700+ surveyed security leaders (linked in Sources)

A policy won't stop every mistake. It does give people a clear answer to the question they're asking anyway: what am I allowed to paste into ChatGPT, Copilot or Gemini today? Microsoft's checklist has seven steps and so does ours below.

What goes in a one-page AI policy

Microsoft publishes a seven-step checklist for an employee AI policy. We'd follow its order and keep each step to a few lines.

  1. 01

    Find out what people already use

    Ask. A short, no-blame survey works better than a firewall report, and it tells you which tools your staff actually like.

  2. 02

    Pick the approved tools

    Publish a short list of authorized tools and the way to ask for a new one. Microsoft's checklist calls for exactly this: a list of authorized platforms, restricted tools and a request process.

  3. 03

    Write the acceptable-use rules with examples

    "Drafting a marketing email: fine. Summarizing a client contract in a public chatbot: not fine." Examples do more than principles.

  4. 04

    Say which data never goes in

    Be specific. Passwords, client financials, patient or employee records, and anything covered by a contract or regulation. If you can't list it, people will guess.

  5. 05

    Name an owner

    One person reviews tool requests and answers questions. In a small company that's often the office manager with IT backing them up.

  6. 06

    Train people in plain language

    Twenty minutes with real examples, then a one-page reminder on the intranet. A PDF in an inbox doesn't count.

  7. 07

    Put a review date on it

    AI tools change quickly. Look at the policy again every few months and cut what no longer applies.

A one-page policy you can copy

Here's the skeleton we'd start from. Fill in your own names and tools, and cut anything that doesn't apply.

  1. Approved tools. [Tool names]. Anything else needs approval first.
  2. How to ask. Email [name] with the tool and what you want to use it for. You'll get an answer within [number] working days.
  3. What goes in. General questions, drafts of public material, and internal text with no client details.
  4. What never goes in. Passwords and keys, client financials and contracts, employee and patient records, anything under an NDA.
  5. Check the output. AI can be wrong and confident about it. A person is responsible for anything sent to a client.
  6. If you slip. Tell [name] straight away. Nobody gets in trouble for reporting a mistake in the first hour.

That last line matters more than it looks. If people are afraid of getting in trouble, they won't tell you, and you'll find out from a client.

What we'd start with for data rules

A starting point for data rules (adjust for your contracts and regulators)
Kind of informationApproved AI toolPublic chatbot
Public marketing copy and general questionsFineFine
Internal documents with no client detailsFineAsk first
Client names, contracts, financialsOnly inside tools under your company accountNever
Passwords, keys, SSNs, patient or employee recordsNeverNever
This is our suggested starting point, not legal advice. Your contracts and regulators set the real limits.

What staff actually paste, and what to say about it

The mistakes we'd expect are boring ones. Someone pastes a client email thread into a chatbot to get a polite reply. Someone uploads a spreadsheet of invoices to get a summary. Someone asks an AI tool to rewrite a contract clause, and the clause names the client.

None of that is malicious. All of it moves client information into a service your company doesn't control. The fix is a plain sentence in the policy: "If it names a client, an employee or a price, it only goes into [approved tool]." Add two or three real examples from your own office and people will remember them.

Where policies usually go wrong

  • Too long. Nobody reads twelve pages, so nobody follows twelve pages.
  • No list of approved tools. "Be careful with AI" isn't a rule. People need a name they can use.
  • No owner. If nobody can say yes or no, people stop asking.
  • Written once, never updated. Microsoft's checklist ends with a review step for a reason.
  • Ban first, answer later. People still use the tool, just off your network and out of sight.

What changes if you roll out Copilot

Copilot can reach anything a user already has access to in Microsoft 365. That sentence is the whole reason we do permissions cleanup first.

Most offices have folders that were shared with "Everyone" years ago and forgotten, like an old HR share or a finance folder from the 2019 audit. A person never opens them. An AI assistant asked a broad question might. So before a rollout we'd look at who can see what in SharePoint and OneDrive, fix the worst of it, and roll out in stages. A rollout shouldn't start with the licenses.

Microsoft's index points the same way. It recommends automated controls that stop unauthorized data exposure and watch for unusual usage, naming Purview data loss prevention and Defender for Cloud Apps. You can read how we handle this on our AI adoption and governance page.

Rolling it out in stages

We wouldn't switch everything on in one go. A staged plan is easier to explain and easier to undo:

  1. 01

    Weeks 1 to 2: policy and clean-up

    Publish the one-page policy, pick the approved tools and look at who can see what in Microsoft 365.

  2. 02

    Weeks 3 to 4: a small pilot

    Give the approved tool to a handful of people who are keen and careful. Ask what works and what confuses them.

  3. 03

    After that: widen it

    Add teams in groups, with a short training session for each. Review the policy once the first few groups have been through.

The timings are our suggestion for a small office, not a rule. A firm with regulated client data may want a longer clean-up.

Questions to ask a vendor before you approve a tool

  • Does our data train your models, and can we turn that off in writing?
  • Where is our data stored, and for how long?
  • Can we use single sign-on and MFA, so only our staff reach it?
  • Can we see who used it and what they asked, if we're ever audited?
  • What happens to our data if we cancel?

A vendor that can't answer these five in a few minutes has told you something. The answers go straight onto your approved-tools list, so the next person who asks gets a yes or no without starting over.

Don't forget who's signing in

An AI tool tied to a stolen login is just another way into your data. Make sure the accounts that reach these tools use strong MFA, and if you're not sure what that means, start with does MFA stop phishing?

Questions people ask

Do we need a policy if we don't use AI officially?

Yes, because your staff probably use it unofficially. A short policy that names approved tools and off-limits data is the cheapest control you can add.

Should we just ban public AI tools?

You can, but people tend to work around bans. We'd rather give them an approved tool that's good enough that they don't need the workaround.

Who should own the AI policy?

One named person who can say yes or no to a new tool. In a small business that's usually the office manager or owner, with your IT provider advising.

How do we know the policy is working?

Look for three things: staff asking about new tools instead of just using them, a tool list that matches what people actually use, and nothing from the "never goes in" list turning up in an incident. If none of those are happening, shorten the policy and add examples.

Sources

  1. Microsoft Security blog: New Microsoft Data Security Index report explores secure AI adoption (2026)
  2. Microsoft: How to create an AI policy employees can follow
NS

Written by

Neteam SolutionsManaged IT and cybersecurity, NYC since 2001

Neteam Solutions Inc. runs IT departments and security for businesses in New York City, Nassau County and northeast New Jersey from offices in Astoria and Williston Park.

Last updated October 28, 2026