Book a fit chat

Writing an AI policy your team will actually follow

By Sam Kim · · 6 min read

Writing an AI policy your team will actually follow

What should a small business AI policy actually say?

Four things, on one page. Which tools are approved and on which accounts. Which information never goes into an AI tool. When a person has to check the output before it is used or sent. And who to ask when something is not covered. A policy that tries to cover every scenario gets skimmed once and ignored, while a short one that answers the questions staff actually have gets used, because it saves them guessing.

Why do most AI policies get ignored?

Because they are written to cover the business rather than to help the person at the desk. A twelve-page document with definitions, principles and a risk framework gets read once at induction. Six weeks later a staff member wants to tidy up a client email with AI, has a specific question, and the policy doesn't answer it in under a minute. So they guess.

Three other patterns come up often:

  • The policy names no tools. "Use approved AI tools responsibly" leaves the reader to decide what approved means.
  • It bans things without saying why. People follow a rule they understand and work around one they don't.
  • Nobody owns it. The tools change every few months, the page doesn't, and staff learn that the document is out of date.

What goes on the page?

Four short sections. Write each one in the words your team uses, not the words a template uses.

Approved tools, and which accounts

Name the tools. Say whether staff use the firm's business account or may use a personal one (usually: business account only, for anything to do with work). If a tool is approved for some tasks and not others, say which.

The account matters more than the brand. A business plan from a provider can come with different terms on data retention and training from a free personal login to the same product. Check the current terms for each tool you approve, because they change, and keep a note of what you found and when.

Information that never goes in

This is the section people will actually use, so make it concrete. A typical list:

  • Client names together with anything about their matter, account or health.
  • Tax file numbers, identity documents, bank details, passwords.
  • Anything a client agreement says stays inside the firm.
  • Staff personal information, including performance and pay.

Then give the alternative. "Remove the name and identifying details first" is something a person can do. "Do not enter sensitive information" leaves them to work out what counts. The client confidentiality checklist goes deeper on de-identifying material and the questions to ask a vendor.

When a person checks the output

AI writes a wrong answer in the same confident tone as a right one. Match the review to the stakes:

  • Anything that reaches a client, a regulator or a supplier: a person reads and approves it before it goes.
  • Anything with a price, a date, a figure, a name or a commitment in it: checked against the source, every time.
  • Internal summaries and first drafts: the person using them is responsible for what they rely on.

Say plainly that the person who sends something owns it, whoever drafted it. Where AI gets it wrong has more on setting a review rule that doesn't eat the time saved.

Who to ask

One name and one way to reach them. Also say what to do when someone has already pasted something they shouldn't have: tell that person straight away, with no blame attached. If owning up gets someone in trouble, the next mistake stays hidden, and the hidden one is the one that hurts you.

How do you write it so it sticks?

Start from the work, not the template. Before you write anything, spend twenty minutes asking the people who already use AI three questions:

  1. Which tools do you use, and on which account?
  2. What do you use them for, in an ordinary week?
  3. What have you been unsure about?

The answers tell you which tools to approve, which tasks to name, and which questions the policy has to settle. They also tell you how far use has spread already, which is usually further than the owner thinks.

Then write the page in plain language, with examples from your own work. "Don't paste a client's file note into a free chatbot" is clearer than any definition of personal information. Keep it to one page. If something needs more detail, link to it rather than adding it.

How do you roll it out?

Walk the team through it in a short meeting instead of emailing a PDF. Ten minutes is enough: what's approved, what never goes in, what gets checked, who to ask. Then take questions, because the questions show you what the page missed.

After that:

  • Put the page somewhere people already go, not in a policies folder nobody opens.
  • Set up the approved business accounts before the meeting, so the right way is ready when the policy goes live.
  • Add a line to new-starter induction.
  • Review it every quarter, or whenever someone wants a new tool. Date each version.

If you are adopting AI more deliberately, the policy and the workflow belong together. A rule that says "check every client email" works best when the workflow puts the draft in front of a person before it can be sent. The pattern in getting your team to use a new workflow applies here too: make the approved way the easy way.

What won't a policy do?

It won't make a bad tool choice safe, and it won't catch mistakes on its own. A policy sets the rules. The review step, the account settings and the person who owns the page do the work. It also won't tell you where AI is worth using in the first place. That comes from looking at where your team's time goes, which is a different exercise. What AI won't do covers that line from the other side.

What to do next

If you'd rather start from a draft, the free AI policy builder writes the one-page version from a few answers.

  • Ask the people already using AI the three questions above, this week.
  • Pick the tools you'll approve and set up business accounts for them. Check each one's current data terms and note what you found.
  • Write the four sections on one page, using examples from your own work.
  • Name the owner, date the page, and walk the team through it.
  • Book the first review for three months from now.

If you'd like a second pair of eyes on which tools fit and where the risks sit, the AI Tools Assessment starts with exactly that inventory.

General information, not legal, tax or financial advice · Drafted with AI tools to my written rules · Terms

Common questions

Do we need an AI policy if only a couple of people use it?

Yes, and that is the cheapest time to write one. If a couple of people are using AI, they are already making decisions about which tool, which account and what goes in, each on their own. A one-page policy turns those private choices into one shared rule before the habits spread to everyone else.

Should we just ban AI until we have worked it out?

A ban mostly moves the use onto personal phones and free accounts, where you can see even less. A short policy with one approved tool and a clear list of what never goes in usually gives you more control than a ban, and it tells staff where the line is rather than leaving them to guess.

Who should own the AI policy in a small firm?

One named person, usually whoever already looks after systems or risk, not a committee. Their job is to answer questions, approve new tools, and update the page when something changes. If nobody owns it, the policy describes last year's tools within a few months.

Next step

Start with the job you'd pay most to stop doing.

It's the core question on the intake, and rough answers are fine. Or book a free 15-minute fit chat first. If the honest answer is "not yet", you'll hear that.

WhatsApp0450 253 297