AI policy for a small business: from rules to everyday decisions
An AI policy should help someone make a decision before they paste information into a tool, connect it to company files or send its output to a customer. For a small business, the practical starting point is simple: approve a particular tool and work account for a defined task, specify the information it may use, and name the person who checks the result.
If you already have an AI acceptable use policy, start there. The next step is to turn its principles into instructions your team can follow during a busy working day. A document saying “use AI responsibly” leaves too much for each person to interpret.
This guide shows how to make those instructions concrete. It is an operational starting point to use alongside your existing policies and any legal or specialist advice your business needs, rather than a complete legal policy template.
Start with the work people are trying to finish
Ask your team which AI tools they use, which accounts they use them through and what they are trying to achieve. Include personal subscriptions, browser extensions and features built into other software. Keep the conversation practical and open: a colleague may have found a useful way to tackle work that currently takes too long.
The NCSC's September 2026 article on shadow AI recommends understanding staff needs and providing suitable approved alternatives. Its shadow IT guidance also explains why slow or ineffective approval processes can encourage workarounds. The useful response is a clear route to approval, with someone responsible for answering.
For each proposed use, record the decision below. Replace roles with actual names before sharing it with staff. Approval of one task should not be treated as permission for every possible use of the same AI product.
An approval record your team can actually use
The following is an illustrative record for a limited trial, not a completed policy for your business.
Decision to record | Example for an appointment-email trial |
Task and boundary | Draft a reusable appointment-confirmation template. A person checks and sends every message. |
Exact tool and account | Enter the product, subscription and business-controlled account that your business has reviewed and approved. Do not substitute a personal account. |
Permitted information | Public service wording, a manager-approved template and fictional details for the trial. No real customer information is needed. |
Information and actions outside approval | No customer email threads, confidential files, passwords or access keys. No connected mailbox, automatic sending or changes to business records. |
Human reviewer | Name the office manager who checks facts, missing details, tone and commitments before the template is used. |
Questions and incidents | Name the policy owner for new uses or uncertainty. Separately name the IT/security contact for a suspected exposure or unsafe action. One person may hold both roles. |
Review trigger | Set a review date. Recheck sooner if the task, information, product terms, account, settings or connected features change. |
Clear approval lets staff repeat that task without asking again while its agreed boundaries remain unchanged. Information or actions outside those boundaries still go back to the owner.
Keep this record somewhere staff already look for working instructions. The wider policy can cover responsibilities and expectations; the record makes one approved use clear enough to follow.
Give staff three clear routes
Within the agreed tool, account, task and information boundaries? Prepare the draft and complete the agreed human checks before using it.
New task, different information or uncertainty? Ask the policy owner before entering the information or enabling the feature. While waiting, use the existing approved process.
Possible data exposure or an unexpected action? Stop that activity and report promptly through your incident route. Record which tool/account was involved and what happened, without spreading the information further. Do not assume that deleting a chat reverses the exposure; let the responsible person decide the next steps.
These routes should be easy to find. A policy is much less useful if a colleague must guess who owns it or waits indefinitely for an answer.
Worked example: a customer email, from prompt to checked template
Consider a fictional small service business that wants a consistent appointment-confirmation template. Its initial trial uses no real customer details and no connection to its mailbox. The aim is to agree useful wording and a repeatable checking step before considering any wider use.
1. Give the tool a narrow job
The office manager uses the approved work account and supplies the public service description plus this instruction:
“Draft a short, friendly appointment-confirmation template in UK English. The appointment is a service review. Keep placeholders for the customer's name, date, time and location. Ask them to reply if they need to rearrange. Do not invent prices, cancellation terms, arrival promises or contact details. Return a draft for a person to check; do not send anything.”
2. Inspect the output, rather than trusting the instruction
An illustrative draft might read:
“Hello [name], we're looking forward to your service review on [date] at [time], at [location]. Please reply if you need to rearrange. There is no charge for changing your appointment. Kind regards, [team name].”
The sentence about changing an appointment is a problem. No cancellation or rearrangement terms were supplied. Even a clear prompt does not remove the need to check the output.
3. Make the human correction visible
Check | Decision before the template is used |
Is every statement supported? | Remove “There is no charge for changing your appointment.” Do not invent a replacement policy. |
Are details missing? | Retain the placeholders. The sender fills them from the approved appointment record, outside this limited AI trial. |
Is the message appropriate? | Check the wording matches the actual service and the business's usual tone. |
Who sends it? | A person checks the completed message and recipient in the normal email process. The AI tool has no sending permission in this trial. |
The approved template is then: “Hello [name], we're looking forward to your service review on [date] at [time], at [location]. Please reply if you need to rearrange. Kind regards, [team name].”
The useful result is a consistent starting point with a known check, rather than each colleague making up their own rules. Whether it saves worthwhile time depends on the task and the checking it needs. Test that in your own process; this example is not a customer result or a claim of measured savings.
Where the same approval stops
Now imagine a colleague wants to paste a genuine customer complaint into the same tool to draft a reply. That is a different use, with different information and potentially different consequences. The appointment-template approval does not cover it.
Pause and ask the policy owner to review the proposed use and data. Simply removing the customer's name may leave them identifiable from the remaining details. Do not assume that a paid account, a familiar brand or a “not used for training” statement makes every input appropriate.
Before approving a tool or expanding a use, the responsible person should check how the relevant account and features handle information, including retention, access, sharing and any connected services. Training use is only one part of that assessment. Microsoft’s enterprise-data-protection documentation, for example, distinguishes core service handling from web queries and agents; the feature being used matters.
Connecting a mailbox, searching company folders, sending messages or changing records also expands the scope. Require a separate decision about access and actions, rather than allowing a drafting trial to quietly become an automated workflow.
Put the policy into practice, then review what happens
Choose one useful, bounded task and walk through it with the people who will use it. Show both a good result and an output that needs correcting. Make sure they can find the approved record, ask a question and report a mistake without worrying that raising it will make matters worse.
Check whether the workflow is helping: are people getting a useful result, how much correction is needed, and are they staying within the agreed boundaries? Update the instructions when you find ambiguity. Review again when the tool, settings, access or purpose changes.
The ICO’s own internal AI use policy provides an example of an organisation combining approved tools, human review and reporting responsibilities. It is useful background, but its internal rules are not a ready-made policy for every small business.
If you are still choosing which repeatable task to start with, our guide to choosing a first business automation can help frame the process. Keep that choice separate from permission to use particular information or carry out actions.
A practical next step for your business
Choose one task your team already wants help with. Write down the tool and account, the permitted information, the human check and the person who owns the decision. If any of those is unclear, that is the place to start the conversation.
Use our free AI Acceptable Use Policy template as a starting point. Adapt it to your business, name your approved tools and accounts, and set clear rules for the information staff may use. The template is not a guarantee of compliance; review it alongside your existing policies and any specialist advice your business needs.
Cache4’s AI Solutions start with how your business works and where AI could genuinely help. Book a free discovery call to discuss your team's everyday tasks and an appropriate next step.
Sources and further reading
NCSC: The hidden risks of shadow AI, published 7 September 2026.
ICO: Internal AI Use Policy, the ICO's own organisational policy.
Sources checked 19 October 2026.
Comments