Skip to main content
Simkins & Elgazar
Book a call

6 min readAhmed Elgazar

How to Write an AI Acceptable Use Policy for a Small Team

Start with an inventory of the AI tools in use, sort your data, then write short rules for tools, data, review and contractors. Train people on it.

  • AI acceptable use policy
  • AI governance
  • AI policy
  • NIST AI RMF
  • small business

To write an AI acceptable use policy for a small team, start by listing the AI tools people already use, then decide which kinds of company data may go into which tools. Write short rules for approved tools, data, human review of output and contractors, name one owner, and walk the team through it. One or two pages is enough if every rule is specific.

Most of the work is making sure the policy matches how your team actually uses AI, and keeping proof that it does.

Why does a small team need one at all?

People on your team may already use AI tools for work, with or without rules. A policy turns those individual habits into a shared decision about which tools are allowed and which data stays out of them.

There is also an outside reason. NIST's Generative AI Profile (NIST AI 600-1, July 2024) names acceptable use policies directly. Suggested action GV-1.4-002 says to "Establish transparent acceptable use policies for GAI that address illegal use or applications of GAI." If a customer or a contract ever asks how you manage AI risk, this is one of the first documents they will expect.

What should you do before writing anything?

Make an inventory. Ask everyone which AI tools they use for work, including free accounts, browser extensions and AI features switched on inside software you already pay for. Write down the tool, who uses it and what they use it for.

A policy written without an inventory approves the tools you know about and says nothing about the rest. The NIST AI RMF treats the inventory as its own outcome (GOVERN 1.6: "Mechanisms are in place to inventory AI systems"), and NIST AI 600-1 suggests adding generative AI systems to that inventory (GV-1.6-001).

What goes in the policy?

Keep it to a handful of sections. Each one should answer a question an employee would actually ask.

SectionThe question it answersWhat a good rule looks like
ScopeWho and what does this cover?Employees, contractors and interns; every AI tool used for company work, including AI features in existing software
Approved toolsWhich tools can I use?A named list, with the account type (company account, not personal)
Data rulesWhat can I paste in?Public data: any approved tool. Internal: company accounts only. Client and personal data: only tools approved for it, or not at all
Human reviewCan I send what it wrote?A person checks facts, figures and citations before anything leaves the company
Prohibited usesWhat is never OK?Uses that break the law or a contract, decisions about people made by AI alone, anything a client agreement forbids
Owner and questionsWho do I ask?One named person, and how to request a new tool
ReviewWhen does this change?A set review date, plus a review whenever you adopt a new tool

The human review rule comes straight from the risk NIST AI 600-1 calls "confabulation": "confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications')." Microsoft's own Copilot documentation says its responses "aren't guaranteed to be 100% factual" and that users should review output before sending it to others.

What happens when nobody checks the output?

This is a public case, not one of our engagements. In Mata v. Avianca (S.D.N.Y., June 22, 2023), the court sanctioned two lawyers and their firm for filing "non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT." One lawyer testified he was "operating under the false perception that this website [i.e., ChatGPT] could not possibly be fabricating cases on its own." He even asked ChatGPT whether one of the cases was real, and it answered that the authorities were "real." The court imposed a $5,000 penalty. A written rule would not have replaced the lawyers' existing duty to check their filings, but a one-line rule such as "a person verifies every citation and figure in a primary source before it leaves the company, and the tool cannot vouch for its own output" names the exact step that was skipped.

How do you decide which tools are approved?

Approve a tool once you know what happens to the data you put into it. Read the vendor's terms and privacy documentation for the exact plan your team uses, and answer three questions: is our input used to train the vendor's models, how long is it kept, and who can see it?

Some vendors answer clearly. Microsoft's documentation for Copilot in Microsoft 365, for example, states that "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs." When a vendor's terms don't answer these questions, treat the tool as public, meaning only public data goes in.

The NIST AI RMF covers this under third-party risk (GOVERN 6.1: policies that "address AI risks associated with third-party entities").

Does the policy cover contractors?

Yes. Contractors often handle the same client data your staff do, on their own accounts and their own tools. NIST AI 600-1 suggests updating acceptable use policies to address "contractors, consultants, and other third-party personnel" (GV-6.1-010).

In practice, add one line to contractor agreements that points to the policy, and apply the same data rules to them.

How do you get people to follow it?

Walk the team through it in person, keep a record of who has seen it, and make it easy to request a new tool so people ask instead of working around the list.

Training is its own outcome in the framework. GOVERN 2.2 says personnel and partners should "receive AI risk management training to enable them to perform their duties and responsibilities consistent with related policies." For a small team, a short walkthrough and a signed acknowledgment is a reasonable start.

Then keep the evidence that the policy is in use: the dated inventory, the approved tool list, the acknowledgments and notes from each review. That folder is what you hand over when a client asks how your team uses AI.

What does a policy like this not cover?

An acceptable use policy covers how people use AI tools. It does not cover how you test or monitor an AI system you build into a product, how you assess its impact on the people it affects, or what you do when one fails. Those sit in the Map, Measure and Manage parts of the NIST AI RMF. If you only use off-the-shelf tools, the policy, the inventory and training cover most of what you need to start.

If you want a first draft to edit rather than a blank page, our open source AI Governance Starter Kit turns a few answers into your first AI policy draft. For help writing one that fits the tools you already run, see our AI governance page.

Sources

Ready to put AI to work safely?

Yes, book a callNot yet, see our products