Phase AI in over about twenty weeks: two weeks to inventory current use and name a lead, a six-week pilot on one low-risk task, two weeks for policy and training, a trustee review with a go or stop decision, then a staged team rollout from week 12 with a full review at week 20. Each phase ends with a decision, not a drift.
Small nonprofits rarely fail at AI because the tools are weak. They fail because three people try three tools, nobody records what worked, and the rules arrive after the first mistake. A phased plan fixes the order of events: learn what is already happening, test one thing properly, write rules from what you learnt, train, get trustees to sign off, then widen use. The timings below suit a team of five to fifteen staff and a pool of volunteers; shorten or lengthen them to your capacity.
The six phases and their decisions, on one page
| Phase | Weeks | Main output | Decision at the end |
|---|---|---|---|
| 0. Look and lead | 1-2 | Tool inventory, named lead, one pilot task | Which task to pilot and who leads |
| 1. Pilot | 3-8 | Time log, quality log, issues list | Continue, change or stop |
| 2. Rules and training | 7-10 | One-page policy, short training, prompt library | Is the team ready to use it |
| 3. Trustee sign-off | 10-11 | Board paper, risk register, minuted decision | Approve rollout and spend |
| 4. Team rollout | 12-18 | Tools in daily use, support routine | Which teams and tasks next |
| 5. Review | 20, then yearly | Results summary, updated risks | Keep, widen or retire each use |
Phases overlap slightly, because policy and training can start while the pilot runs its last two weeks. The overlap is deliberate: the pilot teaches you what the policy needs to say. Treat the weeks as a guide. A busy appeal season may push a phase back, and that is fine so long as each phase still ends with a recorded decision.
Phase 0: find out what is already happening
Before choosing anything, ask. A two-question message to staff and volunteers is enough: which AI tools have you used for work in the last three months, and for what? You will usually discover free chatbots drafting emails, an AI note-taker in a meeting app and image tools for social posts. That inventory is your starting position, and it shows risk (any personal data being pasted) and appetite (what people already find useful).
Then do three things in this phase.
- Name a lead. One person with a few hours a week, not a committee. In a small team this is often the operations manager or a curious fundraiser. The lead runs the pilot, keeps the logs and reports to the chief officer.
- Name a trustee sponsor. One trustee who will ask for updates and take the paper to the board. This avoids a surprise at sign-off.
- Choose the pilot task. One task, one small group, no personal data, easy to check. Good candidates: a monthly newsletter draft, volunteer rota and reminder messages, first drafts of thank-you messages with names left as placeholders, tidying a public web page, or turning meeting notes you have typed up into an action list. Poor candidates: anything with case notes, safeguarding, donor financial data or decisions about individuals.
Write the pilot task on a single line, with the success test: "Newsletter first draft in 30 minutes instead of 90, with no factual errors reaching the edited version." If you cannot say what success looks like, the task is too vague.
Phase 1: the six-week pilot
A pilot is an experiment with a stop rule, not a trial of enthusiasm. Keep the group small, three to five people, and give them the same instructions and tool.
Choose the tool. For a first pilot, use one mainstream assistant. Free tiers work for non-sensitive tasks, but their data settings are weaker, so for anything beyond public text a paid plan with business data terms is safer. Nonprofit offers exist from several vendors. Claude for Nonprofits lists Team seats at $8 per user a month with a two-seat minimum, Google Workspace for Nonprofits includes the Gemini app at no extra cost, and OpenAI also runs a nonprofit offer, whose current price and eligibility rules you should confirm on its own page. Microsoft's nonprofit Copilot add-on is listed at $25.50 a user a month, which is higher than the commercial business price, so compare carefully. Eligibility rules differ and change, so read each vendor's page before budgeting, and see AI tool discounts for charities and nonprofits for what to claim.
Run it in shadow mode where you can. Have people do the task their usual way and also with the assistant, then compare. Piloting AI in shadow mode explains the method, and it removes the risk that an AI draft reaches the public unchecked.
Keep two logs. A time log (task, minutes with AI, minutes you estimate without) and a quality log (what the AI got wrong, what you changed). Ten lines a week is plenty. Add an issues list for anything that worried someone.
A worked example (illustrative): say a small environmental nonprofit with six staff and twenty volunteers pilots newsletter drafting and rota messages with four people for six weeks. In week one, drafts need heavy edits and one invented statistic is caught by the editor. By week three, the team has a prompt that supplies the facts and says "use only these facts", and edits shrink. Over six weeks the log shows the newsletter averaging 35 minutes against an estimated 90 and the rota messages at 5 minutes against 20, with eleven quality issues recorded, all caught in review. The stop-or-go rule written in week one was "go if time falls by a third and no error reaches the public": the pilot passes.
The setup, steps and common ways pilots stall are in how to run your first AI pilot project. Expect energy to dip around week four; that is when the time log matters most.
Phase 2: policy and training, written from the pilot
Now the pilot pays off. The issues list tells you which rules matter, and the pilot group are your first trainers. Keep both pieces short.
A one-page policy covering: which tools are approved; what data must never go in (beneficiary details, donor financial data, safeguarding information, staff records); that a person checks every output before it is used; when to tell people AI was involved; who to ask. A fuller guide is in how to write an AI policy for a small charity. Keep trustee-level risk monitoring separate, as in the trustee AI risk register tutorial.
A 60-minute training session, run by the lead with a pilot-group member. Cover three things: a live demo of the pilot task, the do-not-paste list with real examples, and the check routine (facts, names, numbers and tone). Give everyone the prompt library the pilot built: five or six prompts that worked, each with a sentence on what to check afterwards. Record the session and keep the slides for new volunteers.
A sample prompt from the library (illustrative):
Draft a 200-word thank-you message to volunteers after our beach clean.
Use ONLY these facts: [date], [number] volunteers, [amount] collected.
Leave [first name] as a placeholder. Warm, plain English, no
exclamation marks. Do not add statistics or promises.The final line is the pilot's lesson in miniature: the AI fills gaps with plausible numbers unless told not to.
Phase 3: trustee sign-off
A decision on rollout belongs to trustees, and a good paper makes it quick. Keep it to a page and ask for specific decisions. Include: what was piloted and the results from the logs; the risks found and how they are controlled; the policy for approval; the cost, including nonprofit pricing assumptions and staff time; and the proposed rollout, with the first review date.
Wording you can adapt (illustrative):
The board is asked to: (1) note the results of the six-week pilot;
(2) approve the AI use policy attached; (3) approve the use of [tool]
for [tasks] by [teams] from [date], at a cost of [amount] a year;
(4) agree that beneficiary and donor personal data is excluded from AI
tools until a data review is complete; (5) agree a review at the
meeting in [month].Point four is a useful brake. It gives trustees a safe yes: approve the low-risk uses now, hold back the sensitive ones until the evidence exists. Record the decision, and the reasons, in the minutes. If trustees want a risk view, the register in the governance tutorial gives them one without a technical briefing.
Phase 4: roll out in stages, not all at once
Roll out by task and by team, not by switching everyone on. A sensible order for a small nonprofit: communications and admin first (newsletters, social, rotas), then fundraising support (first drafts of letters, donor update drafts with placeholders), then programme support (summaries of public information, easy-read versions of leaflets), and only later anything involving personal data.
- Week 12. Admin and communications teams start, with the pilot group as buddies. Hold a 15-minute check-in weekly for the first month.
- Week 14. Fundraising joins, with the do-not-paste list repeated and a named checker for every donor-facing piece.
- Week 16. Volunteers are brought in for specific tasks, with the one-page policy in their induction. Pair each new volunteer with a pilot-group member for their first two tasks.
- Week 18. Review quality and issues, then decide whether programme teams join.
Support matters more than the tool. Give people one place to ask questions (a shared channel or a named lead with a weekly slot) and a simple way to report a mistake without blame. If staff fear being criticised for an AI slip, they will hide it, and hidden slips are the ones that reach donors.
Phase 5: review at week 20 and every year
At week 20, repeat the measures from the pilot: time saved, quality issues, incidents, and a short staff question set on how it feels. Add cost: licence fees plus the lead's time. Report to trustees in one page: what is in use, what it saved, what went wrong and what is next. Decide for each use whether to keep, widen or retire it.
A realistic result at this point, for the example nonprofit, might be: three tasks in daily use across eight people, about six hours a week saved across the team, two near misses caught in review and one prompt rewritten. Those numbers are illustrative, but the shape is typical of small-team results: modest and real, not transformational. Report modest results honestly. Trustees and funders trust a charity that shows its working.
Stop-or-go criteria, written in week one
The most useful single page in the pilot is the one that says, before anything is tried, what result means "go". Agree it with the sponsor trustee and the chief officer, and keep it with the logs. A filled-in version (illustrative):
| Measure | Go if | Change if | Stop if |
|---|---|---|---|
| Time on the pilot task | Falls by a third or more | Falls by under a third | Rises |
| Errors reaching the public | None | One, caught and fixed within a day | More than one, or any involving personal data |
| Staff view (5 questions, 1-5) | Average 3.5 or above | Between 2.5 and 3.5 | Below 2.5 |
| Rule breaches (data pasted) | Zero | One, with a fix agreed | Repeated after a reminder |
"Change" is the outcome people forget to plan for. It means keep the task but alter the tool, the prompt or the checking step, and run two more weeks. Having it on the page stops a middling result being argued up to a success or down to a failure.
A second walk-through: a youth club with part-time staff
Say a youth club has three part-time staff and a rota of volunteers, and almost no spare hours. The plan compresses without changing order. Phase 0 is one afternoon: the manager asks who is using what, and learns that a volunteer uses a chatbot to write funding emails. The pilot task is session reminders to parents and volunteers, which contain no sensitive data when written with placeholders. The pilot runs for four weeks with two people. The one-page policy is written in an evening from two issues found (names pasted into the chatbot once, and a wrong date in a draft). Training is a 30-minute chat at the end of a session. The trustee paper is a half page, and rollout adds newsletters and social posts only. The whole plan takes about twelve weeks and costs one free-tier assistant, so the main investment is attention, not money.
The contrast with the first example is useful. Neither organisation did anything exotic; both ran the same order of events at the pace their capacity allowed. A very small team should treat the plan as a menu of checkpoints, not a project: the decision points are what must survive, and the weeks can flex.
What to tell volunteers and supporters
Volunteers and supporters will ask whether AI is writing to them. Decide your line during phase 2 and keep it honest. A short note for the volunteer handbook might read (illustrative): "We use AI tools to help draft routine messages and first versions of newsletters. A member of staff reads and edits everything before it is sent. We never put your personal details into an AI tool without your agreement." If your communications include AI-generated images or a synthetic voice, say so wherever a reader could reasonably be misled. Check your privacy notice too, because adding AI processing of personal data may need updated wording, and a qualified adviser can confirm what applies.
Funders are also an audience. If an application used AI for a first draft, be ready to say how a person verified every claim and number. Some funders publish their expectations; read them before you apply. Using AI for grant writing without funders rejecting it covers that conversation.
Four ways phased plans go wrong
No stop rule. The pilot has no success test, so it cannot fail or succeed, and it drifts. Write the stop-or-go rule in week one.
The lead has no time. A pilot run "in spare moments" stalls. Give the lead a protected two to three hours a week, and say so to the team.
Policy before evidence. A policy written in week one is generic and ignored. A policy written from six weeks of real issues is specific and followed.
Skipping trustees. The team rolls out quietly, then a funder asks who approved it. Sign-off takes one meeting and protects everyone.
There is also the opposite failure: a plan so careful that nothing happens for a year. If six weeks seems too much, run a two-week test on one task and still follow the same order of events. The sequence matters more than the length.
A quick sum for the budget line
Say a nonprofit with eight people on a $8-per-user plan pays about $64 a month, or $768 a year, before any discount differences or taxes. If six hours a week are saved across the team and staff time is valued at $20 an hour, that is $120 a week, or around $6,000 a year of staff time, which you do not cash but can redirect to services. If the pilot saves one hour a week only, the licence still roughly pays for itself on those assumptions. These figures are illustrative, not a promise: use your own seat count, your own price from the vendor's page on the day, and the time log from your pilot.
If you would rather map the phases with someone who has done it, that is what my AI implementation consultation covers; you can reach me through the contact page.
Questions small nonprofits ask about phasing AI in
How long should a nonprofit AI pilot run?
Six weeks is a good default for a small team: long enough to see real tasks repeat, short enough to stop if it is not working. Decide the stop-or-go criteria in week one, not at the end, so the decision is not driven by enthusiasm or sunk effort.
Do trustees need to approve AI use?
For a pilot on non-sensitive tasks, trustee awareness is usually enough. Before a team-wide rollout, or any use involving beneficiary or donor data, get a trustee decision recorded in the minutes. Your governing document and your regulator's guidance may set further expectations, so check them.
What should we pilot first?
Pick one task that is frequent, low-risk and easy to check: drafting newsletters, rota messages, first drafts of meeting follow-ups or tidying public-facing text. Avoid anything involving personal data, safeguarding or funding decisions in the first pilot.
Are there discounts for nonprofit AI plans?
Several vendors offer nonprofit pricing, and Google Workspace for Nonprofits includes the Gemini app at no extra cost, but eligibility rules and prices change. Check each vendor's nonprofit page and your eligibility before budgeting, and confirm prices on the day you buy.
Further reads
- OpenAI for Nonprofits: Who Qualifies and What to Check First — Check eligibility for OpenAI's nonprofit offer first.
- AI Consultant for Nonprofits: What to Hire For and What to Pay — What outside help costs for a nonprofit.
- Charity AI Risks: Data Protection, Deepfakes, and Donor Trust — Data and donor-trust risks to cover in the policy.
- Donor Thank-You Letters With AI That Still Feel Personal — A first low-risk pilot task: thank-you letters.
- Do Small Charities Need Outside Help to Adopt AI? — Whether a small charity needs outside help.
- How to Survey Your Staff Before an AI Rollout (With Questions) — Question wording for the phase 0 staff survey.
- Why Your AI Pilot Stalled, and How to Get It Live — Read in week four, when pilot energy dips.
- ChatGPT Prompts for Nonprofit Organisations: 10 for Admin and Donors — Ready-made prompts to test in your own pilot.
- Fireflies for Nonprofits: Can It Write Board Minutes Safely? — Set up Fireflies for board meetings the safe way: consent notice, a minutes template, a review step before trustees approve, and where the transcript lives.
- Grammarly Pricing for Nonprofits: Is a Team Plan Worth It? — Grammarly Pro is $12 a seat on annual billing, and a TechSoup offer takes 30% off for a year. Seat maths and data checks for a small charity.
- HeyGen for Nonprofits: Donor Update Videos and Disclosure Rules — How a small nonprofit can make donor update videos in HeyGen: avatar consent, ready-to-copy disclosure wording, translation, and when to film a real face.
- AI Bookkeeping for Nonprofits: Restricted and Unrestricted Funds — A sample ledger walked from bank export to AI fund-coding suggestions, accountant review and an audit trail for a small nonprofit.
- Google AppSheet for Nonprofits: Volunteer Rota App Without Code — A build-or-buy check for an AppSheet volunteer rota: the steps, the yearly cost for 42 volunteers, and the limits on reminders and outside users.
- How to Manage a Community Centre: Where AI Helps and Where It Cannot — Four tests that sort community-centre jobs into AI drafts and person-only decisions, with worked examples for bookings, rotas, newsletters and grants.
- Introduction to AI Skills for Nonprofits: What Staff Learn First — A manager's order of teaching for a small charity: donor-data rules, prompting basics, a review habit, and a first-week practice plan for five staff.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor nonprofit pages (OpenAI for Nonprofits, Claude for Nonprofits, Google Workspace for Nonprofits, Microsoft nonprofit offers) as checked in September and October 2026; eligibility and prices vary and should be confirmed before purchase.