An AI policy for staff in a small business.

Why a small team needs one

Your staff are probably using AI already: to draft replies, summarise long emails, write social posts or fix a spreadsheet formula. Without a policy, each person decides alone which account to use and what to paste into it. That is how a customer's complaint ends up in someone's personal chat history, or a reply with last year's prices goes out under your name.

There is a legal reason too, for businesses in the EU or serving people there. Since February 2025 the AI Act has asked businesses that use AI to look after their staff's AI literacy. Since the 2026 Digital Omnibus, the duty is to take measures that support it, matched to people's knowledge and the context they work in. The Act does not ask for a written policy as such, but a page and a short briefing are the plainest way to show the measures you took.

And the GDPR, plus HIPAA for US health practices, applies to whatever personal data staff put into these tools.

Six parts, one page

A policy a small team will read has six parts:

• Scope: who it covers and what counts as an AI tool.
• Approved tools: which ones, on which accounts.
• What never goes in.
• Checks: who reviews AI work before it reaches a customer or the public.
• Honesty: when you say that AI was involved.
• Problems: who to tell, and how fast.

Add an owner, a date and the date of the next review. Everything else, such as how the tools work or why AI matters, belongs in the briefing, not the policy.

Part 1: approved tools and accounts

Name the tools and the account type, because the account decides what happens to the data. "ChatGPT" is not enough. "ChatGPT Business, signed in with your work email" is. Is it safe to put customer data into ChatGPT? explains the difference plan by plan.

Keep a short register next to the policy, one line per tool:

• The tool and plan, such as Gemini in Google Workspace.
• Who may use it.
• What data it may see: green only, or green and amber.
• The settings you have checked, such as retention and connectors.
• Who owns the account and pays for it.

Count the quiet ones too: AI features inside software you already use, browser extensions that read the page you are on, and meeting note-takers that join calls. A note-taker records everyone in the meeting, so it needs the same approval as any other tool, and the other people on the call should know it is there.

Part 2: what never goes in

Keep this list short enough to remember. It applies to every tool, approved or not, unless the owner has set up a specific tool and agreement for that kind of data:

• Card numbers, bank details, passwords and access codes.
• Passport, national ID and Social Security numbers.
• Health information about customers or staff.
• Information about children.
• HR and disciplinary records, and anything about a colleague's pay or performance.
• Anything a client gave you in confidence, and anything under a non-disclosure agreement.
• Customer personal details, on any personal account.

Then add the one sentence that does most of the work: "Before you paste, take out names, contact details and account numbers, and use [Customer] and [Date] instead."

Part 3: who checks AI work

The rule: a named person reads, checks and owns anything AI helped produce before it reaches a customer, a supplier or the public. AI never sends anything on its own unless the owner has approved that workflow.

Spell out what checking means, because "have a look at it" is not a check:

• Every price, date, name and policy, against your own documents rather than against what the AI says.
• Promises: discounts, refunds, deadlines and anything else that commits the business.
• Tone, for complaints and anything sensitive.
• Facts in public posts and web pages, with sources you can point to.

Raise the bar where mistakes cost more. Quotes, contracts, legal or medical wording and anything about money get a second reader. A Canadian tribunal made Air Canada pay for one wrong answer from its chatbot, holding that a company is responsible for what its website tells customers, and an email from your team is no different. Who is responsible when your chatbot gets it wrong has the case.

Part 4: being honest about AI

Most AI help needs no label: a reply that a person checked and sent is that person's reply. Four cases are different:

• Chatbots and voice agents say they are AI at the start. In the EU, Maine and elsewhere the law expects it, as chatbot disclosure laws explains.
• Realistic AI images, audio or video of real people, places or events are labelled. Under the EU AI Act these are deepfakes, and the business that publishes them must disclose them.
• Nobody writes or invents reviews or testimonials with AI.
• Nobody uses AI to pretend to be a real colleague or customer, in writing or by voice.

In the EU, AI-generated text published to inform the public on matters of public interest also needs a label, unless a person has reviewed it and the business takes editorial responsibility for it. The check in Part 3 covers that.

Part 5: when something goes wrong

People report mistakes when reporting is easy and nobody is blamed. Write down:

• Who to tell: one name and a backup.
• How fast: the same day.
• What to say: what was pasted or sent, when, to whom and from which account.
• What happens next: the owner deletes what can be deleted, decides with whoever handles data protection whether it is a breach, contacts the customer if needed and fixes the cause.

Under the GDPR, a personal data breach must be reported to the regulator within 72 hours where feasible, unless it is unlikely to put people at risk. That clock is why "the same day" matters.

Part 6: the briefing, and AI literacy

Give the page to every new starter and talk it through in a short session: which account to use, the never list with real examples from your trade, how to strip a message before pasting, what checking means and who to tell. Repeat it when a tool or a plan changes.

Keep a simple record of who was briefed and when. If your business is in the EU, that record and the page show the measures you took under the AI Act's literacy duty. The amended Article 4 says the measures should fit people's technical knowledge, experience, education and training and the context the AI is used in, and it does not require any individual to reach a set level. The EU AI Act guide covers the rest of the Act.

Test it on three real situations

Before you hand the page out, check that it answers the questions people will actually have. At a ten-person cleaning company, for example:

• The office manager gets an angry email about a missed clean and wants help with a calm reply. The policy says: use the work account, replace the customer's name and address with [Customer] and [Address], and read the draft against the booking record before sending.
• A supervisor wants a new commercial contract summarised before signing. The policy says: contracts are confidential, so either the owner approves the work account for it or the supervisor reads it the old way.
• A new starter installs a free browser extension that rewrites emails. The policy says: ask the owner first, because the extension can read every page and email it touches.

If the page leaves any of these unanswered, add a line now. Those are the questions people would otherwise answer on their own.

The policy, ready to copy

Replace the brackets and cut anything that does not apply to you:

[Business name] AI use policy. Owner: [name]. Last reviewed: [date]. Next review: [date].

• Scope. This policy covers everyone who works for [Business name], including part-timers and contractors, and any tool that writes, summarises, translates, transcribes or creates images, including AI features inside other software.
• Approved tools. For work, use only [ChatGPT Business, Gemini in Google Workspace or Copilot with your work account], signed in with your work account. Do not use personal AI accounts for work. Ask [owner] before you try a new tool or browser extension.
• What never goes in. Card or bank details, passwords, ID numbers, health information, information about children, HR records, anything a client gave us in confidence, and customer details on any personal account. Before you paste, replace names, contact details and account numbers with [Customer], [Date] and so on.
• Checks. You are responsible for anything AI helped you write. Check every price, date, name, policy and promise against our own documents before it goes to a customer or online. Quotes, contracts and anything about money, health or legal matters also need a second reader: [name].
• Honesty. Our chatbot says it is an AI. We never publish AI-written reviews or testimonials, never use AI to pretend to be a real person, and we label realistic AI images of real people or places.
• Problems. If you paste something you should not have, or AI work goes out with a mistake, tell [name] the same day. Nobody is blamed for reporting.
• Training. Everyone gets a short briefing on this policy when they start and whenever our tools change.

Extra lines for regulated trades

Add the line for your trade under "What never goes in":

• Clinics covered by HIPAA: "Patient information goes only into tools covered by a business associate agreement: [list]. None of our other AI tools may see it." See HIPAA and AI chatbots, and what patients ask a practice on our dental clinics page.
• Law firms: "Client and prospective client information goes only into [approved tool], and only for the matter it belongs to." The ABA's Formal Opinion 512 applies the duty of confidentiality to generative AI tools.
• Accountants and bookkeepers: "Client financial records stay in [practice software]. Use AI only on text with the client's details removed."
• Recruitment agencies: "No AI tool may rank, filter or reject candidates without [owner]'s approval." In the EU, AI that filters applications or evaluates candidates is high-risk, with extra duties from 2 December 2027.

Rolling it out without drama

• Ask the team which AI tools they use now, and promise there will be no blame. You need the honest list.
• Pick the work plan and set it up: members, retention and connectors.
• Fill in the template and the register of tools.
• Brief everyone in one short session, with examples from your own work.
• Have everyone confirm they have read it, by email or with a signature.
• Put the never list where people work: on the wall, in the team chat, at the top of the shared drive.

If the honest list shows people pasting the same procedures or price lists into chat tools again and again, the need is real. An AI assistant that answers from your own documents, with agreed access and a person checking what matters, may serve them better than another rule.

Mistakes that make a policy useless

• Banning AI outright. People use it on their phones anyway, out of sight.
• Twenty pages nobody reads. One page that people remember beats a thorough one nobody opens.
• Naming tools without naming accounts.
• No owner, so nobody approves new tools or answers questions.
• No examples from your own work, so the rules stay abstract.
• Forgetting AI features inside existing software, browser extensions and meeting note-takers.
• Writing it once. Vendors change plans, names and settings, so look at it again whenever they do.

A ban does not stop staff using AI. It only stops you seeing which accounts they use and what goes into them.