An AI use policy tells staff which AI tools they may use, on which accounts, with which data, and who checks the work before it goes out. For a 10- to 50-person nonprofit or agency, it can fit on two pages. The hard part is deciding the rules, not writing them down.
Many small organizations already use AI in some form. Staff may have personal ChatGPT, Claude or Gemini accounts, and may paste member lists, grant drafts or client copy into them. Without written rules, each person decides alone what is safe.
This piece walks through the six sections a staff AI use policy needs: approved tools and plans, data that stays out, human review, disclosure, approving new tools, and rollout. It points to the vendor settings pages that back each rule and to the NIST framework if you want a structure to hang it on. It covers internal staff rules only. It is not legal advice. If you handle health, student or payment data, have counsel review your draft against the rules that apply to you.
The short version: approve business plans, not personal ones, because the major vendors do not train on business-plan data by default. Name the data that never goes into any AI tool. Require a human to check anything that leaves the building. Name one person who approves new tools and connectors. Review the policy every quarter, because the tools change that often.
What an AI use policy covers, and what it does not
A staff AI use policy is your organization’s rulebook for its own people. It is different from a vendor’s usage policy, which sets what the vendor allows its customers to do. We covered that side in our post on AI vendor usage policies.
Keep the scope narrow. The policy should answer the questions staff actually ask: "Can I use this tool?" "Can I paste this file?" "Do I have to say AI helped?" "Who do I ask?" Anything longer tends to go unread.
It also helps to treat the policy as part of a wider rollout. Our explainer on AI implementation covers why deployment decisions matter more than model choice.
Use a framework for structure, not for length
You do not need a framework to write a two-page policy. If your board or funders ask what you based it on, NIST’s AI Risk Management Framework is the common reference.
According to NIST’s AI RMF page, version 1.0 was released on January 26, 2023, and is intended for voluntary use. The framework document describes itself as non-sector-specific, with flexibility for organizations of all sizes. NIST added a Generative AI Profile (NIST AI 600-1) on July 26, 2024. As of November 2026, NIST’s page says AI RMF 1.0 is being revised.
The framework has four functions: Govern, Map, Measure and Manage. A small team can map them to plain policy sections:
| NIST function | What it means for a small team | Policy section |
|---|---|---|
| Govern | Someone owns AI decisions | Who approves tools; review schedule |
| Map | Know where AI is used and on what data | Approved tools; data that stays out |
| Measure | Check whether outputs are right | Human review |
| Manage | Act when something goes wrong | Reporting problems; offboarding |
Section 1: approved tools and plans
Start with a short list of approved tools and the plan for each. The plan matters more than the brand, because training defaults differ between personal and business plans.
- OpenAI. OpenAI’s help article on model improvement says it does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or the API to improve its models by default. On Free, Plus and Pro, content may be used unless the person turns off “Improve the model for everyone.”
- Anthropic. Anthropic’s commercial privacy article says it does not use inputs or outputs from commercial products such as Claude for Work and the API by default. Its consumer article says Free, Pro and Max chats may be used if the person allows it. That covers Ultra too, the renamed Max 20x plan.
- Google. Google’s Generative AI in Google Workspace Privacy Hub says content is not human reviewed or used for model training outside your domain without permission. That applies to qualifying Workspace editions.
- Microsoft. Microsoft’s Microsoft 365 Copilot privacy page says prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs.
A simple rule follows: work happens on organization-managed business accounts. Personal accounts are not approved for work data. Our guide to AI training data settings shows where each setting lives if you allow personal accounts for low-risk tasks.
Section 2: data that never goes into an AI tool
Name the categories. "Sensitive data" is too vague for staff to apply on a busy afternoon. A typical list for a nonprofit or association:
- Donor and member records, including giving history and contact details
- Student, client or program participant information
- Health, disability or counseling information
- Staff HR files, salaries and performance notes
- Passwords, API keys and payment card details
- Anything under a confidentiality agreement with a funder or client
Then say what to do instead: remove names and numbers first, or use an approved business account that admins control.
If your policy allows personal accounts for low-risk work, tell staff about the private modes. OpenAI says temporary chats are not used to improve its models while they remain temporary. Anthropic says Incognito chats are not used to improve Claude, even if Model Improvement is on.
Add one more line about feedback buttons. On personal ChatGPT plans, OpenAI says a thumbs up or down can make the entire conversation available for training, even after opting out. Anthropic says consumer feedback and the related conversation are stored for up to 5 years. The simple rule: no feedback buttons on personal accounts.
Section 3: human review before anything leaves the building
AI tools produce confident text that can be wrong. Our post on why AI hallucinates explains the causes.
Write the review rule by output type:
- Published content (web pages, newsletters, social posts): a named editor checks facts, names, numbers and links.
- Grant reports and board materials: the person who signs it checks every figure against the source.
- Code and site changes: the usual code review applies, with no exception for AI-written code.
- Internal notes and drafts: the author is responsible; no second review needed.
Add a line on AI tools that read web pages or email for you. Text hidden in that content can steer the tool. Our explainer on prompt injection covers what to watch for.
Section 4: disclosure for client and public work
Decide when staff say AI helped. This is a choice about trust with your readers, members and clients. Common internal rules:
- Tell clients when AI tools touch their content or data, and on which plan.
- Follow the client’s own AI rules when they are stricter than yours.
- Do not present AI-generated images as photographs of real events.
- Keep a note of which tool and plan produced client deliverables.
Agencies should put the client rule in the contract or statement of work too, so it does not depend on memory.
Section 5: who approves new tools and connectors
Name one person or a small group. New AI features arrive often. Some turn on by default, and many connect to mail and files.
Connectors deserve the most care. Microsoft’s privacy page says Copilot only surfaces organizational data that the user can at least view. That means an AI assistant inherits your sharing mistakes. Fix overshared folders before you connect AI tools to them.
Give staff a short request form: the tool, the plan, the data it would touch, what it connects to, and who will own it. Anthropic’s Claude for Nonprofits announcement describes a skill where staff ask "Can I use Claude for this?" and get an answer based on the organization’s own AI policy. Our post on the Claude for Nonprofits plugin covers what it installs. A tool like that only helps if the AI use policy exists and is current.
Section 6: roll out, train and review your AI use policy
A policy nobody has read does not change behavior. Roll it out in three steps:
- Train. Walk staff through the approved list and the data list with real examples from your own work.
- Make it easy to ask. Name the contact, and keep a running list of answered questions.
- Review quarterly. Check vendor plan changes, new default-on features and incidents.
Anthropic’s announcement points to a Claude Academy tutorial on data handling and drafting an AI use policy, and it scheduled an admin webinar on the same topic for November 10, 2026.
Close the loop at departure. When someone leaves, remove them from every AI workspace. Our employee offboarding checklist covers ChatGPT and Claude workspaces alongside Microsoft 365 and Google Workspace.
A one-page AI use policy outline
- Purpose and scope: applies to all staff, contractors and volunteers using AI for work.
- Approved tools: tool, plan and account type for each.
- Not approved: personal accounts for any work data.
- Data that stays out: your named categories.
- Private modes and feedback: when to use them, and no feedback buttons on personal accounts.
- Human review: who checks what, by output type.
- Disclosure: when and how staff say AI helped.
- New tools and connectors: the request form and the approver.
- Problems: how to report a mistake or a data slip, and to whom.
- Review: the owner and the quarterly date.
Frequently Asked Questions
What is an AI use policy?
A short set of internal rules telling staff which AI tools they may use, on which accounts, with which data, and who reviews the output.
How long should an AI use policy be?
For a small nonprofit or agency, one to two pages. Put examples and answered questions in a separate living document.
Is this legal advice?
No. This covers internal staff rules. If you handle regulated data such as health or student records, have counsel review your draft.
Do ChatGPT Business and Claude Team train on our data?
Not by default, according to OpenAI and Anthropic. OpenAI covers ChatGPT Business, Enterprise, Edu and the API. Anthropic covers its commercial products, including Claude for Work and the API.
Should staff be allowed to use personal AI accounts?
Most small organizations should say no for any work data. If you allow them for low-risk tasks, require the training opt-out and private modes.
What is the NIST AI Risk Management Framework?
A voluntary framework NIST released on January 26, 2023, built around four functions: Govern, Map, Measure and Manage. NIST says it is being revised as of November 2026.
Do temporary or Incognito chats keep data out of training?
OpenAI says temporary chats are not used to improve its models while they remain temporary. Anthropic says Incognito chats are not used to improve Claude.
Who should own the AI use policy?
One named person, often the IT lead or operations director, with a backup. They approve new tools and run the quarterly review.
How often should we update the policy?
Every quarter, and whenever a vendor changes plan terms or turns on a new AI feature by default.