Teach three skills first, in this order: what never goes into an AI chat (donor and beneficiary data), how to ask for a draft with enough context to be useful, and a review habit for everything that leaves the building. Tool tricks come later. A team of five can cover all three in one week of short sessions.
The order matters because the costly mistakes in a charity are about data and promises, not clumsy prose. A fundraiser who pastes a donor list into a free chatbot has done more harm than any weak prompt could. Training catalogues tend to be organised around what the tool can do, which leaves the manager's real questions open: what may we paste, who checks, and what do we stop doing?
The order of skills, and why each sits where it does
Skills one to three make up the first week. The rest follow across the first month.
| Order | Skill | Time to teach | Why it sits here |
|---|---|---|---|
| 1 | Data rules: what never goes into a chat | 20 minutes | One careless paste cannot be undone, so this comes before any practice |
| 2 | The five-part prompt: task, audience, facts, format, check | 30 minutes | Produces useful drafts and builds the habit of supplying facts |
| 3 | The review habit: four checks and a second reader | 30 minutes | Fluent text hides errors; checking is what makes use safe |
| 4 | Choosing tasks: repeated, low-risk, easy to check | 15 minutes | Stops people starting with the hardest and riskiest job |
| 5 | Sharing prompts in a library | 10 minutes | One person's fix then helps all five |
| 6 | Long documents, spreadsheets, meeting notes, grant criteria | Weeks 3 to 6 | Each adds data risk, so they wait until the habits are in place |
Skills one to three need 80 minutes of teaching in total. The rest of the first week is practice, because nobody learns to check a draft by being told how.
Why data rules come before prompting tips
The one-line rule for staff: if a sentence would identify a donor, beneficiary, volunteer or colleague, it does not go into a chat on a personal account, and on a work account it goes in only if the task truly needs it. A traffic-light list turns that rule into something people can apply in seconds.
| Colour | What it covers | Examples in a charity | Rule |
|---|---|---|---|
| Green | Public or generic | Published annual report text; website copy; a job advert; an event description | Use freely on any account |
| Amber | Sensitive, but you can remove the identifiers | A case study with names and places removed; donor groups as counts ("120 donors gave between $50 and $100"); a letter draft with [first name] in place of names | Use on a work account, after removing identifiers |
| Red | Identifies someone, or is confidential | Donor names with giving history; beneficiary case notes; safeguarding concerns; volunteers' contact details; bank details; passwords; unpublished grant budgets | Never in a chat. Ask the manager if a task seems to need it |
Most red items can be turned amber with a minute of work, and that skill is worth practising on day one. Compare two versions of one request. The unsafe one reads: "Write a thank-you to [donor's full name], [street address], who gave $500 in memory of her husband." The safe version reads: "Write a thank-you to a donor who gave $500 in memory of a family member. Greet them with [first name]. Leave out any address." The real name and address are merged in later by your own mailing system, which is where they already live.
The minute-long clean-up has five steps. Replace names with [first name] or a code such as D014. Delete addresses, emails and phone numbers. Round or group amounts and dates when exact figures are not needed. Remove anything unusual that could identify someone on its own, such as a rare job or a named village. Keep the key that links codes to people in your donor system, never in the chat. The tutorial on donor data that must never be pasted in goes through further examples, and the same rules apply to every AI assistant, not only ChatGPT.
Beneficiary details show why combinations matter. A caseworker's note reads: "Client, 34, referred by a health professional after an eviction notice, two children, lives above a shop on [street name]." No name appears, but age, referral route, street and circumstances together point to one person. The amber version is: "A parent in their thirties facing eviction, referred by a health service, with two children." The rule of thumb is that if someone who knows the community could name the person from the details, the details are still red.
Prompting basics a fundraiser can learn in half an hour
People learn prompting fastest from a bad prompt and its fix. The bad one: "Write a thank-you email to donors." It returns something like "Dear Supporter, thank you for your generous gift. Your support changes lives." It is not wrong, just empty, and every charity's email now reads that way. The fix is a five-part pattern: the task, the audience, the facts to use, the format and limits, and a request to list anything that needs checking.
Task: Draft a thank-you email to people who gave to our autumn food-bank appeal.
Audience: first-time donors. Warm, plain, no jargon. We are a small local charity.
Facts to use (only these): the appeal raised $4,310 from 96 donors; at $2 a meal
(our own figure) that covers about 2,150 meals; the next collection day is
Saturday 21 November; replies go to [name], donor care.
Format: subject line plus body, under 150 words, greeting "Dear [first name]".
Limits: do not claim anything that is not in the facts. No exclamation marks.
Avoid phrases like "changes lives".
At the end, list every fact in the draft that I should double-check.A good response (illustrative) opens "Thank you for your gift to our autumn appeal. Together, 96 of you raised $4,310, enough for about 2,150 meals," adds the collection date, and closes with a line listing two facts to check: the $2-a-meal figure and the date. Even so, a first draft might include "your gift will feed families all winter". That sentence is a promise the facts do not support, because 2,150 meals is a number of meals and not a number of families or weeks, so it comes out. The two habits this prompt builds matter more than any trick: paste in the facts you want used, and make the tool say what it is unsure of. This tutorial of ten prompts for admin and donors gives more patterns to copy.
The review habit: four checks and a second reader
Review is the skill that separates a useful tool from a liability, and it is quick once it is routine. Four checks, in this order:
- Facts. Every number, date, name and place traced back to a source you can open.
- Promises. Claims about impact, where the money goes, or what happens next. If you could not defend it to a funder, remove it.
- People. Is anyone identifiable? Is the language dignified and free of stereotypes? Would the person described be comfortable reading it?
- Voice. Does it sound like your charity? Cut anything you would never say aloud.
Anything that leaves the building, such as a funder, donor or press item, also gets a second reader who did not write it. Take an AI-drafted appeal paragraph as the test case. The draft: "Last year, 87% of the families we supported moved into stable housing within six months. Every penny you give goes directly to the people we help." Check one finds the 87% has no source: the model invented a plausible figure, so it goes, or it is replaced with a number from your own records. Check two catches "every penny", a claim that fails if any of your gifts pay for admin or fundraising costs. The corrected paragraph says what you can prove: "Last year we supported 42 families; 31 of them were in stable housing six months later." Those counts would come from your own case records, with the identifiers stripped before any AI touches them.
Some grant-makers now ask applicants to disclose AI use in their applications, so keep a note of which applications and reports were AI-assisted and read each funder's guidance before you submit. A short log is enough: date, document, what AI did, who reviewed it.
A first-week plan for a team of five
This plan assumes five staff: a director, a fundraising officer, a programme coordinator, a finance and admin officer, and a volunteer coordinator. Each session is short, so it fits around a normal week.
| Day | Time | Who | What happens | Output |
|---|---|---|---|---|
| Monday | 45 min | All five | Traffic-light rules and one live demo of the five-part prompt; each person lists five items from their own job as green, amber or red | A one-page rules card on the shared drive |
| Tuesday | 30 min | Each alone | Practice task one from the list below, using green data only | A draft and a list of facts to check |
| Wednesday | 30 min | Pairs | Swap drafts and run the four checks on each other's work | Notes on what was caught |
| Thursday | 30 min | Each alone | Practice task two with amber data after removing identifiers; save the best prompt | A prompt added to a shared library |
| Friday | 20 min | All five | Show one prompt and one error caught each; decide keep or stop for each task; estimate time saved | A decision list and training-log entries |
That is 155 minutes a person, or about 12.9 hours of staff time across the team. It is an investment, not a saving, and the last section does the sum on when it pays back.
Monday sets the tone, so script it. A 45-minute run sheet: five minutes on what you are doing and why (everyone tries it and nobody is being tested); ten on the traffic-light list, with two examples from your own work; fifteen on a live demo where you run the vague prompt and then the five-part prompt on a green task, and then show an error in the output on purpose; ten for each person to sort five items from their own job into colours while you read out the borderline ones; five for questions and for agreeing where the rules card will live. Three sentences are worth saying aloud. "Checking is the skill, not knowing the answer." "Tell me about a mistake the day it happens and nobody gets blamed for it." "If you are unsure which colour something is, it is red until we have talked."
The Thursday output, a shared prompt library, needs no software. A single document with one row per prompt is enough, and it should record who checks the result and when the prompt was last tested, because tools change behaviour from month to month and an old prompt may no longer do what its row claims.
| Task | Prompt saved as | Data colour | Checked by | Last tested |
|---|---|---|---|---|
| Appeal thank-you email | Five-part prompt with appeal facts | Amber | A second reader | 12 November |
| Volunteer rota reminder | Shift list in, short message out | Green | Volunteer coordinator | 12 November |
| Funder guidance summary | Ten bullets, page number per deadline | Green | Director | 11 November |
One practice task per role, with the check that goes with it
People learn faster on their own work. These five starter tasks use the plan above, and each names the check that matters most for that role. All are illustrations; adapt them to your own job descriptions.
Director: a funder briefing. Paste the public guidance for a grant (green data) and ask: "Summarise this guidance in ten bullet points under three headings: what they fund, what they require, deadlines. Give the page number for each deadline." The check is every deadline against the original, because a wrong date is the costliest error a summary can contain.
Fundraising officer: a thank-you email. Use the five-part prompt above with one appeal's real figures and a de-identified donor group. The check is promises: nothing beyond what the facts support.
Programme coordinator: a volunteer briefing. Turn rough session notes into a one-page briefing: "Rewrite these notes as a briefing for new volunteers: what to bring, timings, who to ask for help. Keep my wording on safeguarding exactly as written." The check is that no beneficiary details slipped into the notes and that the safeguarding wording is the policy's own. AI does not write your safeguarding procedures; it only tidies the page around them.
Finance and admin officer: an overdue-invoice reminder. Draft a polite chase for venue-hire invoices using placeholders for names and amounts: "Write a courteous reminder email for an invoice that is 14 days overdue. Use [name], [invoice number] and [amount]. Offer a phone call if there is a problem." The check is amounts and dates, and the rule that totals are added up in the spreadsheet, not by the assistant.
Volunteer coordinator: a rota reminder and a recruitment post. Draft a Thursday reminder message from a list of shifts with first names only, and a short recruitment post for Saturday sorters. The check is dates and times against the rota, and that the post does not promise anything about training, expenses or references that the charity has not agreed.
Choosing the account the team practises on, and what it costs
Run the week on work accounts, one login per person. Both OpenAI's and Anthropic's terms bar sharing credentials, so a single shared login is a breach as well as a visibility problem. On personal consumer plans, chats can be used to improve the provider's models unless training is switched off in the privacy settings, while business plans from both vendors do not train on your content by default. That difference is the main reason to put the team on a business-grade plan.
As of October 2026, both vendors have nonprofit offers. OpenAI for Nonprofits lists ChatGPT Business Standard seats at $8 per user a month on annual billing ($10 monthly), with the Premium seat excluded and eligibility checked through Goodstack. Claude for Nonprofits lists a Team plan at $8 per user a month with a two-seat minimum, also through Goodstack. For five staff that is $40 a month on either, and eligibility rules vary, so check yours before you plan around the price. Other suites have their own nonprofit programmes, and it is worth looking in the admin panel of any suite you already use to see what assistant it includes. The OpenAI eligibility tutorial and the one on charity tool discounts cover the sign-up details.
If the budget is zero, the free plans are enough for week one provided the team uses green and de-identified amber data only and switches the training option off. The skills you are teaching do not depend on which assistant is open.
A short paper trail, and the AI Act's literacy duty
If your charity operates in or serves people in the EU, Article 4 of the EU AI Act asks providers and deployers of AI systems to take measures to support the AI literacy of their staff and other people using AI on their behalf, taking account of their knowledge, experience and training and the context of use. After the 2026 amendment, the text says this does not require guaranteeing any specific level of literacy. In other words, the standard is proportionate effort, and a documented week like this one is a sensible form of it. The wording covers "other persons" using AI on your behalf, which for a charity may include volunteers. Whether and how the duty applies to you is a question for your adviser.
Wherever you operate, the log is useful. For each person, record the date, what was covered, which tool and account they use, and the version of the rules card they were given. Five minutes per session keeps it current. The tutorial on training record logs lists what to capture, and the small charity AI policy tutorial shows where the rules card grows into a policy.
How to tell the week worked, and what to teach next
At the end of the week, ask each person five questions. Can you name three things you will not paste? Can you show a prompt that supplies the facts and sets limits? Can you show a fact you checked and where you found it? Can you say when you would ask the manager before using AI? Can you show one task you now do with it, and how long it takes? If any answer is shaky, repeat that part with that person rather than moving on.
Suppose the scorecard for the week reads as follows. The team completed ten practice tasks. Seven drafts were used after editing and three were dropped because the tool was no better than writing from scratch. Reviews caught nine factual problems: three invented statistics, four wrong dates or numbers, and two unsupported promises. No red-category data was pasted. Self-reported time saved was 40 minutes for the director, 25 for fundraising, 30 for programmes, 20 for finance and 35 for volunteers, a total of 150 minutes.
The sum: 775 minutes of staff time invested (155 each for five people) against about 150 minutes saved a week. If those savings repeat, the week pays back in a little over five weeks, and the habits it builds are what stop the next donor-data mistake. If the savings do not repeat after a month, the practice tasks were wrong, not the idea; swap them for jobs that recur weekly.
In month two, teach the next layer: summarising long documents and spreadsheets, meeting notes with consent, and drafting grant sections against a funder's own criteria. Turn the rules card into a written policy at the same time, so that the habits from week one are on paper before the riskier tasks begin.
Questions managers ask about AI training for a small charity
Do volunteers need the same training?
Anyone who will use AI on the charity's behalf needs the data rules and the review habit, and both are short enough to teach in one session. Volunteers who use their own accounts raise extra questions about who can see the chats and who owns the output, so settle those in your AI policy before asking volunteers to draft anything with it.
Should we start with an online course instead?
Courses from large vendors and sector groups are a good second step because they explain features and general concepts. They do not know your donor data rules, who reviews what, or which tasks you want to try. Run the first week from this plan, then use a course to fill gaps in month two.
What if someone already pastes donor data into a personal account?
Stay calm and do not blame them; the rules were never written down. Stop it, note roughly what went in and when, check the account's training setting, delete the chats if that is appropriate, and tell the person responsible for data protection. Ask your data-protection adviser whether anything must be reported.
How long before staff count as trained?
A week starts the habits, but competence builds across a month of real tasks. Re-check each person at 30 days with the five questions at the end of this plan, and again after any change of tool or any near-miss. A short written log of who learned what, and when, is more useful than a certificate.
Further reads
- AI Literacy Requirements: What Your Staff Need to Know — What staff need to know, and where the literacy duty comes from.
- AI Implementation for Nonprofits: A Phased Plan for a Small Team — A phased plan once the first week has gone well.
- How Much Time and Money Does AI Staff Training Take? — What staff training takes in hours and money.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI and Anthropic nonprofit programme pages (as recorded in October 2026); OpenAI Services Agreement and Terms of Use on training and credentials; text of Article 4 of the EU AI Act; a grant-making trust's published statement on disclosing AI use in applications.