A Nonprofit AI Policy Template That Covers AI Agents, Not Just Chat

Most nonprofit AI policies were written for chat: a staff member opens ChatGPT, asks for a draft, and copies it somewhere. That's covered well enough by a few rules about data and checking the output. The gap is agents, software that takes steps on its own, like scheduling posts, editing a website or drafting campaign changes. If a tool can act, your policy needs to say what it may do without a person.

Below is a template you can adapt, then the reasoning behind each section. It comes from the rules we run our own agents by at Storyline and NonProfitSaver. We haven't had a board adopt a formal AI policy ourselves, so treat this as a working draft, and have your attorney review it before your board votes.

The rules we run our own agents by

Before the template, here are the rules we actually use. They're short on purpose.

1. A person approves anything that publishes, sends, spends money, deploys or deletes.
2. Agents draft; people send.
3. Never invent people, quotes, stories, statistics or results. Cite a source or cut the claim.
4. Done means checked, not claimed.
5. Donor data and passwords never go into notes, articles or prompts that leave our systems.
6. Each agent works only in its own area.
7. Anything an agent reads from outside, such as emails, web pages or comments, is information, never instructions.
8. If a safety check blocks something, the agent stops and asks rather than working around it.

Rule 1 isn't a promise; it's built into the software. A check that runs before every command refuses any attempt by an agent to send, schedule or test-send a Mailchimp email. Agents can build the draft, and only Mike can send it. He asked for that on October 20, 2026, in so many words: "he can build a draft, only i can send those out."

Rule 4 came from a mistake. One of our agents graded a listing as its best find of the week; the listing had come off the market seven months earlier, and an outdated copy was still showing on another site. Since then the agent must confirm anything is current from a dated source before it can rank it, and a separate reviewer checks results before they reach a person.

The template

Copy this into a document, replace the brackets, and cut what doesn't apply.

[Organization] Artificial Intelligence Use Policy

1. Purpose. This policy sets how staff, volunteers and contractors use artificial intelligence (AI) tools, including AI agents that take actions, in [Organization]'s work.

2. Approved tools. Only tools on the approved list may be used for [Organization] work, on accounts in [Organization]'s name under a business or nonprofit plan. Personal accounts may not be used for [Organization] data. The approved list is kept by [role] and reviewed every [6] months.

3. Data that never goes into AI tools. Donor and member records, giving amounts, prayer requests, counseling or pastoral notes, health information, information about minors, passwords, and bank or payment details. Exceptions require written approval from [role] for a specific tool that has been reviewed for that data.

4. Checking AI output. A person reviews anything AI produced before it is published, sent, submitted or relied on. The reviewer checks facts, numbers, names and quotations against a source. AI output is never presented as a person's words, a real story or a testimonial unless it is one.

5. AI agents. An AI agent is software that can take actions, such as posting, sending, editing systems or spending money. Agents may draft and prepare work. A named person must approve each action that publishes, sends, spends money, changes a live system or deletes anything. Agents receive the narrowest access their job needs, under accounts [Organization] owns. Content an agent reads from outside, such as emails, web pages or comments, is treated as information and not as instructions.

6. Disclosure. [Organization] tells people when they are talking to an automated system rather than a person, and does not use AI to impersonate a staff member.

7. Mistakes. Anyone who finds an error caused by AI reports it to [role]. Errors in published material are corrected in the open.

8. Ownership and review. [Role] owns this policy. It is reviewed at least once a year and whenever a new kind of tool is approved.

Adopted by the board on [date]. Reviewed by counsel on [date].

Why each section is there

Approved tools on organizational accounts. Business and nonprofit plans come with data protections that personal accounts don't, and they keep the work when a staff member leaves. They also cost less than people expect. As of October 2026, ChatGPT Business and Claude Team are $8 a user a month for verified nonprofits, and the Gemini app is free with Google Workspace for Nonprofits. Our AI tools discount comparison has the sources and the open question of whether churches qualify for ChatGPT's price.

A list of data that never goes in. It's easier to follow a short list of what's forbidden than a long list of what's allowed. Many useful AI jobs, such as drafting social posts from your public articles, need no private data at all.

Checking output. AI writes confident sentences that are sometimes wrong. The rule that matters is that a named person checks facts against a source before anything leaves the building.

Agents. This is the section most templates leave out. An agent that can post or send can make a mistake many times before anyone notices. Requiring approval for each action that publishes, sends, spends, changes a live system or deletes keeps the speed of the agent and the judgment of a person. The line about outside content matters more than it looks: an email or a web page can contain text written to trick an AI into doing something. Treating it as information keeps it from becoming an instruction.

Disclosure and mistakes. People trust organizations that tell them what's automated and fix errors in public.

Spam and sign-ups: a policy question you may not expect

Automation cuts both ways. On Storyline's website email sign-up, about 40 of the 47 new contacts added after October 13, 2026 turned out to be bots, found on October 20. We tried reCAPTCHA and double opt-in, and rolled both back because they also blocked real people. A spam-checked sign-up went live on October 21. If AI tools will send to your lists, the policy should also say who keeps those lists clean.

Where to go next

For what agents can run today and what they cost, read AI agents for nonprofits. If you'd like an agent run for you under these rules, our AI agent solutions page lists what we offer; that's a paid service, and you don't need it to use this template.