Editorial note: this article was drafted with AI assistance and reviewed, fact-checked and edited by Steven May Jr. before publication. Read our editorial process.
AI Policy for Nonprofits: A Starter Template
A definition-first, section-by-section template nonprofits can adapt for approved tools, review, and incidents.

An AI policy for nonprofits is a short internal document that tells staff what they may enter into an AI tool and which outputs require review before use. It names who owns approvals and what to do if something goes wrong. It is not a legal filing or a compliance certificate, and it should not read like a mission statement about innovation. A working policy answers an ordinary workday question: can I use this to draft a volunteer email, and who checks it before it goes out?
What should the purpose, scope, and approved-use sections cover?
The purpose and scope sections of an AI policy for nonprofits set the boundary before any rule gets written. Write a purpose statement tied to your mission: AI may assist with approved drafting, summarizing, and administrative tasks, while people remain responsible for decisions and final outputs. Specify who the policy covers, including staff, contractors, and volunteers where relevant, and include tools embedded in software you already use. A writing feature inside an email platform is still an AI feature, even when nobody purchased a separate assistant.
Distinguish experimenting with public information from processing confidential records, and separate internal drafts from material sent to donors or program participants. Those differences should determine the review requirement for each task. Avoid treating every use as equally risky or equally safe, since brainstorming a public event title and summarizing a client's case history call for very different levels of oversight.
A starter approved-use rule could read: team members may use approved tools to draft nonconfidential content, summarize approved public documents, and build internal outlines. They may not submit restricted information or delegate consequential decisions without an explicitly reviewed process. Customize restricted information to your organization; it commonly includes donor payment details, personal case notes, staff records, credentials, or information covered by an agreement.
Name examples of consequential decisions, such as determining program eligibility, making hiring decisions, or approving payments, rather than assuming staff already share your definition. State plainly that a tool's suggestion is not authority to bypass the normal approval process. Give people a way to request a new use instead of leaving them with a flat prohibition.
How should approved tools, data rules, and human review work together?
Approved tools, data rules, and human review work together by tying each tool to a named owner and a review step before anything reaches an outside audience. Maintain a short table with the tool name, approved account or plan, owner, approved tasks, and information restrictions, and keep it beside the policy itself. State that unapproved connectors and public sharing are not permitted for confidential materials, and that vendor terms and settings must be reviewed before sensitive information is processed. A paid account is not automatically approved; the actual handling terms and configuration matter more than the price.
For a first version, consider limiting use to public or nonconfidential information while leadership reviews more sensitive workflows. Include instructions for minimizing data: staff should not upload a whole folder when one approved paragraph answers the question. Give people a named contact who can resolve uncertainty before information leaves your systems.
Assign a person to every outward-facing output, with responsibility for checking facts, names, dates, tone, permissions, and whether the message meets its purpose. A draft is not approved simply because it sounds polished. For grant applications, financial reports, and communications about individual people, the reviewer should compare claims against authoritative records. An assistant can omit an important qualification or state something plausible with no real source behind it.
State who can publish, send, or authorize actions, and make clear that staff remain responsible for work submitted under their names. Protect reviewers from unrealistic workloads; if a review takes longer than doing the task manually, revisit the workflow rather than the reviewer's effort.
What do transparency, incident reporting, and rollout look like in practice?
Transparency, incident reporting, and a measured rollout turn the written policy into something staff can actually follow. Decide when the organization will disclose AI assistance and how automated interactions will be described, and make sure a website assistant does not pretend to be a human staff member. For content drafting, follow the relevant funder, platform, contractual, or professional requirements that actually apply to you, rather than inventing one universal disclosure rule. Record which obligations matter for your organization and who checks them when a new channel is added.
Prohibit invented testimonials, case studies, citations, and program outcomes, and require permission plus accurate context before using any real story. AI cannot turn a fictional example into donor proof or a beneficiary's words; a draft's fluency does not make a quote authentic. When a team needs an illustration, label it hypothetical and keep it separate from reports of actual work.
Name the policy owner and the reporting channel, and ask staff to report accidental uploads, inappropriate outputs, unauthorized tool use, and unexpected automated actions promptly. Specify the first response: pause the affected workflow, preserve the facts needed to investigate, and notify the owner. Avoid promises about deletion or legal outcomes that staff cannot verify, since the right response may need vendor assistance or qualified advice depending on the information involved.
Review the policy on a scheduled basis and whenever a material change occurs, since new models, tools, connectors, or program uses can alter the risks. Version the policy and record approvals so people know which rules currently apply, and involve the people who actually perform the tasks rather than writing the document in isolation.
Rollout examples make the rules concrete for staff who are new to the policy.
- Safe starting example: draft a newsletter outline from approved public program descriptions.
- Approval-needed example: summarize internal survey responses that may identify people.
- Prohibited-until-reviewed example: upload donor payment data to an unapproved assistant.
- Review example: compare an AI-prepared board summary against original program totals.
- Incident example: report a mistakenly shared document through the named channel immediately.
Use these as discussion prompts rather than universal classifications, since your records and obligations may change the right answer. Give staff a chance to ask questions and practice with fictional data before the policy goes live. Show them exactly where the approved tool list lives and how to request access.
After a short pilot, ask staff where they hesitated and where the rules were unclear. Were they unsure which documents counted as confidential? Did the approved account differ from the one on their phone? Did the review checklist actually catch factual errors? Update those specifics rather than adding vague language just to make the policy look more thorough.
Track examples of permitted use, issues raised, and changes made, which gives leadership an honest picture of adoption without collecting unnecessary personal detail. Begin with a narrow set of uses and expand after testing, rather than granting blanket permission and trying to recover control later. If no one can explain who owns an automated process, it is not ready for routine use.
Editorial note: this article was drafted with AI assistance and reviewed, fact-checked and edited by Steven May Jr. before publication. Read our editorial process.