Skip to content
be-human.ai
← Back to the journal

Human oversight

Responsible AI Policy for Small Companies: A One-Page Outline

A one-page responsible AI policy for a small company: data use, approvals, disclosure and incident reporting, with a four-week plan to adopt it this month.

Be Human typographic cover with the title: Responsible AI Policy for Small Companies: A One-Page Outline
Conceptual illustration created with AI · Be Human

A responsible AI policy for a small company is a short written set of rules that states what data AI tools may use, which actions need a named person to approve them, when customers and staff are told that AI was involved, and how problems get reported. One page is enough to start if every rule names an owner and can be checked. The goal is a clear default for the next decision a person or agent faces, not a legal essay.

Key takeaways

  • A one-page policy covers five parts: scope, data use, approvals, disclosure and incident reporting.
  • Every rule should name an owner and a way to check that it is being followed.
  • Start with the AI tools staff already use, including those on personal accounts, before approving anything new.
  • Write approval rules as a list of specific actions, not as a general feeling about risk.
  • Review the page on a fixed schedule and again after any incident.

Why a small company needs the policy before it needs a large one

Most small companies already use AI tools without a written rule about them. Staff paste customer messages into a chat assistant, a scheduling tool reads a shared inbox, or a marketing draft is generated from a spreadsheet. None of this is unusual, and none of it is automatically wrong. The risk comes from the absence of a shared default, because each person decides on their own what is safe to share, who should check the output and whether a customer needs to know.

A policy closes that gap. It does not need to predict every future tool. It needs to answer a small set of questions the same way each time, so that a new employee, a contractor or an agent working for the company can follow the same rules.

Keeping it to one page is a practical choice. A short document people actually read is more useful than a thorough one stored in a shared folder. The NIST AI Risk Management Framework describes AI risk work as an ongoing cycle of governing, mapping, measuring and managing, and a one-page policy is a sensible first output of the governing step.

The five parts of a one-page policy

A useful page has five short parts. Each can be written in a few sentences, and each should name the person responsible for it.

  • Scope: which tools, agents and people the policy covers, including tools used on personal accounts for company work.
  • Data use: what information may and may not enter AI systems, and how long any vendor may keep it.
  • Approvals: which actions an AI system may take alone, and which need a named person to sign off first.
  • Disclosure: when customers, candidates and staff are told that AI was used.
  • Incident reporting: how a suspected problem is raised, who responds and where it is recorded.

Scope is the part most often skipped, and it matters most. If the policy only mentions the chatbot on the website, it misses the assistant that drafts quotes from the sales inbox, which is usually the tool that carries the most customer detail.

What the data use section should say

Data rules work best when they name categories of information rather than tools. A line that says identity documents, health information, bank details and passwords never go into any AI system is clearer than a general instruction to protect privacy, because a person can apply it at the moment of decision.

Write three lists into this section: data that is never allowed, data that is allowed only under a stated condition, and data that may be used freely, such as published product descriptions. Then add two sentences on vendors. The company checks each vendor's data processing terms before connecting it, and no tool is connected to company records until that check is complete. Brazilian companies also need the section to reflect LGPD obligations, and LGPD and Superintelligence: What Brazilian Companies Should Prepare sets out what to prepare.

If an AI system keeps memory about customers or staff, the section should also say how long that memory is kept and how it is removed when someone asks. Without that line, memory quietly becomes the least controlled dataset in the company.

Illustrative scenario: a small consulting firm lets an assistant summarize client calls. The data section says call summaries may contain project details but not personal health information, and that summaries are deleted after the project closes. When a client later asks for their information to be removed, the operations lead already knows where the summaries live and who can delete them.

Where a person must approve before anything happens

This section decides whether the policy changes behavior. The useful question is not whether AI is risky in general, but which specific actions an AI system may take without a person looking at them first.

A short decision guide covers most small companies:

  • Can the action be reversed within a day, without cost or customer impact? If yes, the system may proceed, and the action is logged.
  • Does the action send something outside the company, such as an email, quote, invoice or message to a customer? If yes, a named person approves the first version and reviews a sample of later ones.
  • Does the action move money, change a price, edit a contract, grant access or delete records? If yes, a named person approves each instance.
  • Does the action affect a decision about a person's health, employment, credit or legal position? If yes, the system only prepares a recommendation, and a human makes the decision.

Writing the answers as a list lets staff tell the difference between the assistant drafted this and the assistant sent this. The reasoning behind separating routine actions from high-risk ones is developed in Human Approval Gates for Superintelligent Agents: Where People Sign.

How to disclose AI use without confusing customers

Disclosure rules should match what a reasonable customer would expect to know. The policy can set three defaults. Customer-facing text drafted with AI is reviewed by a person before it is sent, unless the approval rules say otherwise. Any automated chat or voice system identifies itself as automated at the start of the conversation and whenever a customer asks. Staff are told which internal tools use AI and what kinds of information those tools can see.

The OECD AI Principles name transparency and accountability as core values, and a disclosure section is the most direct way for a small company to act on them in daily work. Keep the wording plain. A sentence such as this message was drafted with an automated assistant and checked by our team is enough for most written contexts. Avoid claims the company cannot verify, such as saying a system is unbiased or never makes mistakes.

How to respond to an AI incident in the first day

An incident is any event where an AI system did something it should not have, exposed information, or produced output that caused harm or a complaint. Define the term broadly enough that staff report near misses as well as failures, because near misses are where the policy learns.

A workable sequence for the first day:

  • Stop the system or the specific action, using the pause available in the tool or by removing its connection. Do not delete logs or queued work.
  • Write down what happened, when, which system was involved and who noticed it.
  • Assign one owner who decides whether customers, staff or a regulator need to be told.
  • Check whether personal or confidential information left the company, and record which vendor was involved.
  • Correct the affected messages or records, and log each correction.
  • Agree on the fix and name the person who approves a restart.
  • Update the policy page with the rule that failed, the change made and its date.

The pause matters because people under pressure often delete things to make a problem disappear, and then the review cannot happen. The technical side of stopping an agent without losing its records is covered in AI Agent Kill Switch: Halt Actions Without Losing Logs or Queued Work.

Illustrative scenario: a small accounting office lets an assistant draft payment reminders from a shared spreadsheet. One morning a column change causes reminders to go to the full client list instead of the overdue accounts. The owner pauses the assistant from its settings, exports the send log, confirms that no bank details were included, sends a short correction to the affected clients, and asks the operations lead to add a rule that any bulk send needs a second approver. The incident goes onto the policy page the same week. This scenario is invented for illustration and does not describe a real company.

A four-week plan to adopt the policy this month

The policy does not need a long project. A small team can adopt it in four weeks by working through one part each week.

  • Week one: list every AI tool in use, including personal accounts used for work, and name an owner for each. Write the scope paragraph.
  • Week two: write the data categories and the approval guide, then check the terms of each vendor that touches company data.
  • Week three: agree the disclosure wording and the incident steps with the people who would carry them out, and walk through one realistic incident as an exercise.
  • Week four: publish the one-page policy, ask each person who uses AI tools to confirm they have read it, and set the date of the first review.

Be Human approaches this as an operating matter rather than a document. In a first engagement, the team defines one useful role for an agent, the memory and tools that role needs, the approval boundaries for each action and an observable success criterion, so the policy describes a specific system rather than AI in general. In managed operations, the same rules are checked through monitoring, recurring evaluations, permission reviews and incident response, and the written policy is updated when the agent's behavior or the company's needs change. The company keeps control of its identity, memory, channels and permissions throughout.

When the policy is in place, the next useful question is which task is worth handing to an agent first, and the free Be Human AI diagnostic is built to help you find it.

Frequently asked questions

Does a company with only a few people really need a written AI policy? Yes, although it can be very short. Staff already make decisions about AI tools every day, and a written default replaces many small, inconsistent choices with one shared rule. A page that names data categories, approvals, disclosure and incident steps is enough to start.

Who should own the policy in a small company? One named person, usually the owner or an operations lead, should hold the policy and its review date. Other people can own specific rules, such as the person who approves outbound customer messages. Ownership should be written on the page itself so that nobody has to guess who decides.

Does the policy need a lawyer to review it? For most small companies, a lawyer or data protection specialist should review the data and disclosure sections, especially when personal data, health information or a regulated sector is involved. The policy is an operating document, but legal obligations differ by country and sector, so the review step should be stated explicitly on the page.

How often should the policy be reviewed? Review it on a fixed schedule, and also after any incident or whenever a new tool or agent is connected to company systems. Each review checks whether the approval list still matches what the systems actually do. Keep a short dated change log at the bottom of the page so that every edit can be traced.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →