AI Use Policy Template for Small Business: The Clauses That Matter
The clauses a small company needs in an AI use policy: data, approvals, logging and agent ownership, with the questions to answer for each.

An AI use policy for a small company is a short written document that sets out which AI tools and agents may be used, what data they can touch, which actions need a person to approve them, what gets logged, and who owns each agent. A workable template covers four blocks: data, approvals, logging and ownership. The template is the easy part, because each clause only protects the business once someone has written the specific answer for that business.
Key takeaways
- Build the policy around four blocks: data, approvals, logging and ownership.
- Name the data groups each agent may read and send, and the restricted data it may never touch without a named approval.
- Put a person in front of any action that speaks for the company, moves money or changes records.
- Give every agent a business owner, a technical owner and a named person who can pause it at once.
- Review the policy when a tool, permission, incident or staff role changes, not only once a year.
Start with scope before you write any clause
Before drafting, list the AI uses that already exist in the company, including the ones nobody approved. Staff often use general chat assistants, browser extensions and meeting note tools with company material, and a policy that ignores them will not protect data that is already leaving the building. Cover three things: the tools in use, the people allowed to use them, and the data classes those tools can reach.
Write one purpose sentence at the top. It should state what AI is meant to do for the business and what it will not do without a person, such as making a hiring or credit decision on its own. Keep the document to a few pages. A policy that tries to predict every future tool becomes unread within weeks, while a short one gets followed.
Data clauses: what agents may see, keep and send
Start with three data groups in plain language. Public material, such as published catalog pages, may go into any approved tool. Internal material, such as procedures, project notes and supplier terms, may go only into tools covered by a written data agreement. Restricted material, such as personal data about customers and staff, payment details, health information and contracts under negotiation, should not reach an agent unless a named owner has approved that specific flow.
Then add three clauses. Agents read only the sources listed in their setup record, and adding a source requires approval. Agents keep personal data in memory only as long as the stated purpose requires, and the policy says how someone asks for deletion. No agent sends company data to an outside destination without a person reviewing that destination first. Ask each vendor where data is processed and which subcontractors can see it, because the answer may sit in a separate subprocessor list rather than the main contract. LGPD and AI Agents: What Brazilian Companies Should Check covers the questions to ask when personal data is involved in Brazil.
Approval clauses: which actions need a person
Approval clauses decide what an agent may do alone. The simplest workable rule is that an agent may read, summarize, classify and draft freely, while any action that is hard to reverse or speaks for the company needs a named approver. Write the list down so that staff do not have to guess in the moment.
- Ask a person before any message goes to a customer, supplier or candidate in the company's name.
- Ask a person before any payment, refund, purchase or price change.
- Ask a person before an agent edits customer, pricing, payroll or contract records.
- Ask a person before an agent gains a new tool, channel or permission.
- Ask a person before any deletion, unless it is a routine retention job already approved in writing.
- Allow the agent to act alone on internal drafts and read-only lookups that a person reviews anyway.
Two rules keep this list honest. The approver must have authority over the affected process, and must see the proposed action in full rather than a summary that hides the detail. High-stakes actions may need a second approver, as described in Two-Person Approval for AI Agents. Keep the list short, because approval requests that arrive all day teach people to click yes without reading.
Illustrative scenario: a small office lets an agent draft replies to supplier invoices. Under the policy it can categorize invoices and write drafts, but any reply that confirms a payment date goes to the finance lead. In the first week the drafts start proposing dates that do not match the purchase orders, and the approval step stops them before a supplier relies on a wrong date.
Logging clauses: what you record and who reads it
A log is what lets you explain an agent's action after the fact. Write a clause that lists the events recorded for each agent: the request received, the sources consulted, the tools called, the approval given or refused, the output sent, and any error or blocked step. The business owner should be able to read these records without help from the builder, since the person who answers for an agent is not always the person who configured it.
Logs contain personal data too, so the same clause should say who may read them, how long they are kept, and when they are deleted. Prefer references and short summaries over full copies of sensitive documents. Blocked actions deserve their own review rhythm, because a guardrail that stops an agent is information about what the agent tried to do. Blocked AI Agent Actions: How to Review What Your Agent Tried to Do describes a simple way to run that review. Engineering teams often describe this kind of record with concepts such as traces and events, which the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) explains in general terms.
Ownership clauses: who owns each agent
Every agent needs two named owners. The business owner decides what the agent is for, what success looks like and whether it keeps running. The technical owner maintains its tools, credentials and updates. The policy should also name who may pause an agent immediately, and that person should not need permission from the builder to do it. Pausing should leave logs and queued work intact so that nothing is lost during the stop.
Agent ownership is separate from two other questions. The first is who owns the work an agent produces, which depends on your contracts, your vendors' terms and the law where you operate, so ask counsel to confirm the wording. The second is who answers when an agent makes a mistake. The policy should name a company role for that, rather than point to a vendor's terms of service, because your customers deal with your company, not with the tool. Staff changes need a clause too: when someone leaves or changes role, their access to agent tools, memory and approvals moves with the responsibility. Offboarding Staff From AI Agent Access gives a practical order of steps.
Review, incidents and exit: keeping the policy alive
A policy that never changes is usually a policy nobody follows. Set review triggers in the document itself, so that a review happens when something changes rather than when someone remembers. Useful triggers include:
- A new tool, data source or channel is added to any agent.
- An agent is granted a permission it did not have before.
- A logged incident, a pattern of blocked actions or a customer complaint involves an agent.
- A person with approval authority leaves or changes role.
- A vendor changes its terms, subprocessors or data location.
Finish the policy with an exit clause that says how an agent is retired, what happens to its memory and logs, and how its credentials are revoked. Keep it short and test it once before you need it.
In a managed engagement, Be Human starts by writing the data groups, approval list and log fields with the business owner, then configures each agent's memory, tools and permissions to match those answers. Monitoring, recurring evaluations and permission reviews check that the running agent still follows the policy, and incident response uses the pause and review steps agreed in advance. When the records show a clause that no longer fits how the work happens, the change goes through the same review, so the policy keeps matching the system it governs.
When you are ready to decide where a first agent would help most, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Is a free AI use policy template enough for a small company? A template gives you the headings and a sensible order for the clauses, which saves time. It does not know which data your business holds, which actions your staff take with money or customers, or who can pause an agent. Those answers have to be written into the template before it protects anything.
How long should an AI use policy be? Aim for a few pages that a new employee can read in one sitting. Put the data groups and the approval list on the first page, because those are the parts people consult during daily work. Move detailed procedures into separate documents that the policy references.
Should we ban staff from using public AI tools? A blanket ban often pushes use out of sight, which makes it harder to see what data is leaving the company. A better approach names the approved tools, states which data groups may go into them, and gives people a clear route to ask about a tool they want to use. Review the rules whenever a new tool appears in your work.
Who should approve the policy? The business owner or a senior operations leader should sign it, because that person answers for the business outcomes. Technical owners should review the logging and permission clauses so the rules match what the systems can actually enforce. Share the signed version with everyone who uses or manages an agent, and set the date of the next review when you sign.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


