Shadow AI policy for small businesses: five practical rules

Shadow AI is what happens when staff start using AI tools before the business has decided what is approved. In a small team, that can happen quickly and quietly.

Someone rewrites an email. Someone pastes meeting notes into a summariser. Someone asks a chatbot to improve a proposal. Someone tests a browser extension because it saves time.

The answer is not to panic. The answer is to give people rules they can remember.

Keep the policy short enough to use

A small business AI policy should fit on one page. If staff need legal training to understand it, it will not work in the moments that matter.

Start with five plain rules:

1. Use approved tools for business work
2. Do not paste sensitive customer, staff, financial or confidential data into unapproved tools
3. Check AI output before it leaves the business
4. Ask before connecting AI tools to company accounts or shared drives
5. Report mistakes, leaks or unusual outputs quickly

That is enough to change behaviour while the business builds a more mature approach.

A worked example: allowed, prohibited and awaiting approval

This is a fictional policy example for a small team, not a description of a customer or an endorsement of a particular AI supplier.

Allowed: a writing assistant in the business's approved account may help draft public website copy using already-public information. A named staff member checks facts, tone and promises before publishing. Approval applies to this workflow, not every feature the tool offers.

Prohibited: staff must not paste payroll, customer lists, private emails, contracts, passwords or unpublished financial information into unapproved tools. Removing a name alone does not necessarily make a document suitable to share.

Awaiting approval: a meeting bot or browser extension wants access to company email, calendars or shared files. Do not connect it just because the standalone writing tool is approved. The owner must review the actual supplier, account type, data handling, requested permissions and removal process first.

The approval record can be short: tool and account, named owner, allowed task/data, forbidden data, connected permissions, decision and review date. Staff ask the owner before changing that scope; silence is not approval. If data is shared by mistake, report the facts promptly without forwarding the sensitive content to more people.

The NCSC's guidance on shadow AI highlights the need to understand unapproved use and give staff an appropriate route for useful tools. The example above is a practical starting point to adapt, not a compliance certificate.

Define sensitive data in normal language

People need examples. "Confidential data" can sound abstract. Say what it means for the business.

For many small businesses, off-limits data should include customer lists, invoices, contracts, passwords, employee information, unpublished financials, private emails, supplier agreements and anything covered by a non-disclosure agreement.

This is also where AI readiness is not buying Copilot matters. Buying a tool does not answer the data handling question by itself.

Make output checking explicit

AI output should be treated as a draft, not a decision. Staff should check facts, tone, calculations, customer promises and anything that could create legal, financial or reputational risk.

This rule is especially important for customer-facing work. A polished answer can still be wrong. A confident summary can still miss context.

Control plugins and connected apps

The riskiest AI usage is often not the prompt box. It is the tool that asks to connect to email, files, calendars, CRM data or browser activity.

Small businesses should have a simple approval step before any AI tool is connected to company systems. That approval does not need to be bureaucratic. It does need to ask what data the tool can see and who owns the decision.

Give people a safe way to report mistakes

If someone pastes the wrong data into a tool, the business needs to know quickly. A blame-heavy policy makes people hide problems. A practical policy gives them a route to say what happened so the business can respond.

For a small team, that might be as simple as reporting to the business owner, operations lead or IT contact.

Review it every quarter

AI tools change quickly. The policy should be reviewed regularly, especially when new tools are adopted or suppliers start using AI in their own service delivery.

The aim is not perfect governance. The aim is a clear baseline that makes good decisions easier.

For the surrounding account, email and recovery checks, see what a small business AI and security review should cover.