01Set staff rules
AI Policy Template for Small Businesses
A free, copyable AI policy template for small firms: approved tools, data rules, review, disclosure and incident steps, with a completed example.
- Topic
- Set staff rules
- Written by
- AI Ground Rules
- Last reviewed
- Sources
- 1 checked source
This AI policy template is the full set of rules your staff follow when they use AI at work. It answers five questions: which AI tools and accounts staff may use, what information may go into them, who checks the results, when to tell a client that AI was involved, and what to do when something goes wrong. It also names one person who owns the decisions, a backup, and a review date. Copy the block below, replace the square-bracket fields, and adjust it to your work.
The decisions to make first
A policy records decisions. Make them before you write anything down.
| Decision | Question to answer | Who usually decides |
|---|---|---|
| Approved tools | Which product, plan and work account may staff use, for which tasks and data types? | Data decision owner, with IT |
| Information | What is fine, what needs approval, and what never goes in? | Data decision owner, checked against client agreements |
| Review | Who checks AI-assisted work before it reaches a client or the public? | Team leads or a named reviewer |
| Disclosure | When do clients or readers need to know AI was used? | Data decision owner, checked against client agreements |
| Incidents | Where do staff report a problem, and who responds? | Data decision owner and a backup |
If you cannot answer a row yet, leave that use unapproved until you can. Paying for a tool does not make it approved. An approval covers the listed tool, plan, work account, tasks and data types, and nothing more.
The AI policy template
Replace every square-bracket field. The disclosure rule is an illustrative example: change it to match your client agreements. This sample is a starting point, not legal advice.
[Business name] AI policy
Policy owner (data decision owner): [name, role]. Backup: [name, role]. Version: [1.0]. Effective date: [date]. Next review date: [date].
1. Purpose and scope. [Business name] allows AI to support approved work when people protect information, check results and stay responsible for decisions. This policy applies to employees and contractors using AI for our work, including AI features inside other software.
2. Approved tools. Use AI for our work only in an approved tool: the exact product, plan and work account listed in our approved tools register at [link], for the tasks and data types listed there. Paying for a tool does not make it approved. A personal account on the same product is not approved. Connected AI apps need approval like any tool. That includes note-takers, mailbox or file summarisers, browser extensions and apps given access to our accounts. Ask the data decision owner before trying a new tool or a new use.
Each register entry records:
Tool and plan Work account Approved tasks Data groups allowed Conditions Approved by and date [Product, plan] [Firm login type] [Tasks] [Group 1, or Group 1 and named Group 2 items] [Settings, limits] [Name, date] 3. What information may go in. Use the minimum information needed. When groups overlap, the stricter rule wins.
- Fine in an approved tool: public information we may use, such as published pages and public guidance; your own notes and drafts with no client, staff or confidential details; invented examples.
- Ask first: client documents, emails and files; internal documents; customer lists and order details; spreadsheets and PDFs you have not checked fully. These need approval for this tool, account, data type and task. Also check client contracts and engagement letters. A client restriction beats any internal approval.
- Never, unless the data decision owner has approved one named tool for one named task, in writing: payroll, salaries, HR records, health details, ID numbers (such as SSN, NI or passport numbers), tax records, bank and payment-card details. The default answer is no.
- Absolute never, with no approval or exception: passwords, access codes, private keys, sign-in links, MFA codes.
An approved tool never overrides these limits or a client restriction. Removing names does not make a document safe: context, figures, file names and logos can still identify a person or client. Use invented details instead. If you are unsure, stop and ask the data decision owner.
4. Check the work. Treat AI output as a draft. You own what you send. Check facts, figures, names, dates and citations against reliable originals. A person checks AI-assisted work before it goes to a client or is published. [Reviewer or role] approves client-facing and public work. If the reviewer cannot verify a material claim, hold the work. Do not use AI alone to decide hiring, dismissal, eligibility, legal rights, financial commitments or another serious matter.
5. Disclosure. We follow each client’s agreement first. Where an agreement requires permission before we use AI on that client’s work, we get it in writing before we start. When AI drafted a substantial part of a deliverable, we add this line: “Parts of this document were drafted with AI assistance and checked by [name].” The data decision owner approves any other wording. Before recording or transcribing a meeting with an AI tool, follow our recording procedure at [link], including notice and any participant consent the law requires.
6. Actions. Do not let an AI tool send messages, publish, change records, delete files, make purchases or connect new systems unless that exact workflow has been approved in writing. Actions that need human approval must wait for it.
7. If something goes wrong. If information may have gone into the wrong tool, an AI tool did something unexpected, or an error could affect someone:
- Stop the task, if it is safe to do so. Pause or disconnect the app if you can straight away. Keep a record: do not delete the chat unless the data decision owner says so.
- Tell the data decision owner, or [backup], straight away through [channel]. Say which tool and account, what time, and what kind of information was involved.
- Do not copy the sensitive content into the report.
- Do not wait to be certain. Reporting early is expected.
8. Confirmation. Staff confirm they have read this policy when they join, and again after each new version.
Owner incident checklist
When a report comes in, the data decision owner (or backup) works through these steps.
- Contain. Stop the task. Where you can, pause the workflow, suspend the account or disconnect the connected app. If a password, access code or sign-in link was entered, change or revoke it now.
- Record. Write down the tool, account, time, what information was involved, who was involved and what has been done so far. Describe the information; do not copy it into the record.
- Escalate. Tell IT or your IT provider, and your adviser [name] if client or personal information was involved. Ask whether you must tell the client, a regulator or the people affected, and whether a deadline applies.
- Fix affected work. Find any AI-assisted work that went out and correct or withdraw it. Tell the client where needed.
- Restart. Resume only once the cause is fixed. Update the register or this policy if a rule was unclear, and tell staff what changed.
First-response script. Use this when someone reports a problem:
Thanks for telling me straight away. Please stop that task now. Tell me which tool and account you used, roughly what time, and what kind of information was involved. Do not paste the information itself. I will take it from here and get back to you by [time].
Name the owners
A policy works when a real person can answer questions and act. In a small firm one person often holds several roles, but each role needs a backup.
- Data decision owner. Approves tools, tasks and data types, checks client restrictions, handles exceptions and leads the response to incidents. Usually also owns the policy and its review.
- Tool and access owner. Sets up work accounts and settings, and can suspend access.
- Work reviewer. Checks AI-assisted work before release and can stop it.
Before you share the policy, send a test message to the reporting channel and confirm someone sees it.
Share the policy with staff
- Publish the policy beside the approved tools register, in a place everyone can find.
- Send a short announcement. Say what the policy covers, where the register is, and who to ask.
- Run a 20-minute practice session. Ask a colleague to find the approved work account, decide whether an invented client brief can be used, spot a fact to check in an AI draft, and find the reporting route. If they have to guess, rewrite the unclear rule.
- Ask staff to confirm they have read it. Record the version and date.
- Set the next review date. Quarterly is a reasonable default. Review sooner when a tool, connected app, client obligation or agent permission changes.
Acceptable use
For the staff-facing list of what they may do, what needs a question first and what is never allowed, see the AI acceptable use policy.
Examples by firm type
For worked versions of this policy for accounting, agency, professional services and other firms, see AI policy examples by firm type.
What a policy does not do
A written rule does not configure a tool or enforce an approval. Someone still needs to set up accounts, permissions and settings to match the policy. The policy also does not replace legal, privacy or professional advice for regulated work or client contracts. The NIST AI Risk Management Framework is voluntary guidance that gives useful context. Following it does not certify your firm.
A completed example
Here are the key fields filled in for a fictional 12-person design studio.
| Field | Completed |
|---|---|
| Policy owner (data decision owner) | Operations lead |
| Backup | Managing director |
| Version and dates | 1.0. Effective 1 December 2026. Next review 1 March 2027. |
| Register row | One AI writing tool, business plan, work logins only. Task: drafting marketing copy from the studio’s published work. Data: group 1 only. Condition: no client files. Approved by the operations lead, 20 November 2026. |
| Reviewer | Content lead checks every public draft. |
| Disclosure | The standard line above, unless a client agreement says otherwise. |
| Reporting channel | The internal help channel, with the managing director’s phone as backup. |
02Evidence