01Set staff rules

AI Policy Examples for Small Businesses, by Firm Type

Six fictional AI policy examples for small firms, with a comparison table, a sample tools register entry and a real decision for each firm.

Written by
AI Ground Rules
Last reviewed

These six fictional firms all use the same data rules. What changes by industry is which information falls into each group, and where the hidden risk sits. Compare the table with your own firm, then read the example closest to yours.

All firm names and details are invented. Square brackets mark details you would replace. To write your own policy, start from the AI policy template.

Compare the examples

Every firm sorts information into three groups. “Approved tool” means the exact product, plan and work account listed in the firm’s approved tools register, for the tasks and data types listed there. Paying for a tool does not make it approved.

Firm type Fine in an approved tool Ask first Never, unless approved in writing for one tool and task Extra rule to watch
Accounting or bookkeeping Public tax guidance, own notes with no client, staff or confidential details, invented examples Client books, statements, emails Payroll, tax records, bank and payment-card details, ID numbers New AI features in bookkeeping apps
Marketing agency Public brand material, own drafts with no client, staff or confidential details Client briefs, unreleased campaigns, customer lists Client staff HR or health details Client contracts may limit AI or set ownership
Online shop Public product details Order details, customer lists, supplier prices Customer bank or card details Customer-facing chatbots need notice
Recruiting Public job info, own outreach drafts CVs, interview notes, client hiring plans Salary history, ID numbers, health details AI screening or ranking needs separate approval
Legal-adjacent services Public law and guidance Client contracts, case files Staff disputes with health or HR detail, ID numbers Engagement terms may refuse AI
MSP or IT firm Public vendor docs, own scripts with no client, staff or confidential details Support tickets, client configurations Client HR or payroll seen in their systems AI agents may not act on client systems alone

Two rules apply to all six. Passwords, access codes, private keys, sign-in links and MFA codes never go into any AI tool, with no exceptions. When groups overlap, the stricter rule wins.

A completed tools register entry

Each firm’s approved tool is an entry like this one, from the fictional accounting firm below. The register is what makes a tool approved, not the invoice.

Field Entry
Tool and plan ExampleAI Business (fictional), business plan
Account Firm work account only (@harbourledger.example). Personal accounts on the same product are not approved.
Approved tasks Drafting client emails, summarising public tax guidance, checking spreadsheet formulas
Approved data Public information, own notes with no client, staff or confidential details, invented examples. Client books and statements for the tasks above, unless the client’s engagement letter says no.
Not approved Payroll, tax records, bank and payment-card details, ID numbers, passwords
Limits Use the minimum needed. A person checks every figure against the source.
Approved by Priya Shah, data decision owner. Backup: Tom Reyes.
Approved on / review by 3 March 2026 / 3 March 2027

To list what AI is already in use, including connected apps and AI agents, see the AI inventory template.

Example 1: accounting or bookkeeping firm

Fictional: Harbour Ledger, 18 staff, bookkeeping and tax returns for local businesses.

  • Data decision owner. The practice manager.

Use client books and statements only in [approved tool], with your work account, for the tasks in our register. Payroll, tax records, bank and payment-card details and ID numbers never go into AI unless [data decision owner] has approved one named tool for one named task, in writing. Check every figure, date and tax point against the source before it reaches a client. “The AI said so” is not a reason we accept.

  • Who checks. The client manager signs off any figure or advice that came from an AI draft.
  • Client disclosure. Explained in the engagement letter, with the client’s agreement where needed.
  • Decision: refused. A bookkeeper asked to paste a client’s payroll export into the approved tool to find errors. The data decision owner said no: payroll is not approved for that tool, and a spreadsheet check does the job.
  • Watch out. A new AI feature in the bookkeeping app counts as a new tool. Approve it before staff switch it on.

Example 2: small marketing agency

Fictional: Brightside Creative, 12 staff, social media and campaigns for small brands.

  • Data decision owner. The managing director.

Use [approved tool] to draft ideas, outlines and copy from public or approved material. A client brief, customer list or unreleased campaign goes in only if the client’s contract allows it and the register approves that tool for it. A person edits and checks every piece before it is published or sent.

  • Who checks. The account lead reads all client-facing copy and checks facts, claims and names.
  • Client disclosure. Tell clients in the proposal how AI is used, and ask before using AI images in their brand work.
  • Decision: allowed. An account manager asked to draft a launch post from a client’s published product page. The data decision owner approved it: the material is public and the client’s contract does not restrict AI.
  • Watch out. Client contracts may limit AI use or say who owns the work. Check the contract before using AI on a new account. Also check that a reused draft does not still carry another client’s name.

Example 3: online shop

Fictional: Fern and Flint, 9 staff, sells homewares online.

  • Data decision owner. The owner.

This shop treats two uses very differently.

Drafted product copy. Staff use [approved tool] to draft product descriptions, ads and email ideas from public product details. A person reviews every piece before it goes live.

Customer-facing chatbots. A bot that talks to customers is a separate tool and needs its own approval. Before switching one on, check the notice rules where you sell, and make sure the bot says it is automated. The European Commission says that under the EU AI Act, people using chatbots “should be made aware that they are interacting with a machine” (European Commission).

Use AI to draft product copy from public product details. Do not paste customer names, emails, addresses or order details into any AI tool unless the register approves that tool for it. If you want help with a customer issue, use invented details, such as “a customer wants a refund for a cracked vase”. No customer-facing bot goes live without [data decision owner]’s approval, and every bot must say it is automated.

  • Who checks. The owner or shop manager reads product copy and promotions before they go live.
  • Customer disclosure. None needed for reviewed product copy. Any chatbot tells customers up front that it is automated.
  • Decision: held. A staff member wanted to switch on the shop platform’s AI chat assistant. The owner put it on hold until she has checked what order data it can read and how it tells customers it is a bot.
  • Watch out. Removing a customer’s name alone does not make a message safe. An order number, address or unusual product can still identify them. Use invented details instead.

Example 4: recruiting firm

Fictional: Northgate Talent, 25 staff, places finance and admin staff.

  • Data decision owner. The operations director.

Use AI to draft job ads, outreach messages and interview questions. Upload CVs or candidate notes only to a tool the register approves for candidate data. AI may help summarise, but a recruiter decides who is shortlisted, rejected or put forward. Using AI to screen, score or rank candidates needs separate approval from [data decision owner], after checking hiring, discrimination, audit and notice rules where the role is based.

  • Who checks. The recruiter who owns the role reads any AI summary against the original CV. They check for anything left out and for wording that could reflect bias, before sending it to a client.
  • Client disclosure. Tell hiring clients if AI is used to summarise or screen. Tell candidates where your privacy notice or local rules require it.
  • Decision: refused. A consultant asked to use a CV-ranking feature on a role based in New York City. The data decision owner said no. The city bars employers and agencies from using an automated employment decision tool unless it has had a bias audit within one year, the audit information is public, and candidates have been given notice (NYC DCWP). The firm had none of these.
  • Watch out. AI note-takers on interview calls connect to email and calendars and need approval like any tool. Recording follows the firm’s recording procedure, including notice and any consent the law requires.

Fictional: Calder Advisory, 40 staff, HR advice, contract reviews and compliance help for small employers.

  • Data decision owner. The head of operations.

Client documents may go only into [approved tool], for the tasks in our register, and only if the client’s engagement terms allow it. A client restriction beats any internal approval. Removing names does not make a document safe: dates, figures and context can still identify people. To test an approach, use an invented example. Check every legal point, date and citation against the original. AI does not give advice to clients. Our advisers do.

  • Who checks. A senior adviser reviews anything that states a legal position or quotes a law or case.
  • Client disclosure. Set out in the engagement terms. Ask permission before using AI on a client’s documents if the terms do not already cover it.
  • Decision: held. An adviser wanted to summarise a client’s staff grievance file. The head of operations held it: the file includes health details, and the client’s terms are silent on AI. Before any use, it needs written approval from the data decision owner for one named tool and task, and the client’s permission.
  • Watch out. Some clients will say no. Record each client’s answer where staff can see it.

Example 6: an MSP or IT firm, for itself and its clients

Fictional: Keystone IT, 30 staff, runs Microsoft 365 for about 60 small businesses.

  • Data decision owner. The technical director.

Never enter passwords, keys, sign-in links, MFA codes or access codes into any AI tool. Client network details and support tickets go only into a tool the register approves for them. Use AI to draft scripts and documents, then test scripts in a safe setting. An AI agent may suggest a change, but an engineer approves it before anything runs on a client system.

  • Who checks. A second engineer reviews any AI-written script before it runs on a client system.
  • Client disclosure. Each client gets a short note on which AI tools are switched on and what they can reach.
  • Decision: allowed. An engineer asked to use the approved tool to draft a PowerShell script from Microsoft’s public documentation. The technical director approved it: no client data goes in, and a second engineer tests the script first.
  • Watch out. Microsoft states that Copilot “only surfaces organizational data to which individual users have at least view permissions” (Microsoft Learn). So if a payroll folder is open to everyone, Copilot can bring it into an ordinary question. Check sharing before you switch it on, using the Microsoft 365 Copilot checklist. For agents that act, see the AI agent human approval checklist.

If something goes in by mistake

Stop the task, and pause or disconnect the app if you can straight away. Keep a record: do not delete the chat unless told to. Then tell the data decision owner (or the backup) through [channel]: which tool and account, when, and what kind of information. Do not copy the sensitive content into the report. The AI acceptable use policy has the full steps.

What the six have in common

Each names one data decision owner, keeps a register of approved tools, tasks and data, and has a person check output before it leaves the firm. For the staff rules, see the AI acceptable use policy. For what may go into which tool, see client data and AI tools.

For a full policy to adapt, see the AI policy template.

How these examples were made

The firms are invented. Each one applies the same three data groups used across this site’s guides. Vendor and legal statements were checked against the linked official pages on the review date above. These examples are not legal advice. Your contracts, professional rules and local law may ask for more. A written policy also does not set up the tool. Someone still needs to configure accounts, sharing and permissions to match it.

04Next step

The kit includes the complete editable policy, an approved tools register, a data use guide and fictional worked examples.