How a nonprofit gets started with AI

Note: general recommendations only, not professional advice. You are responsible for checking before use. The full text is at the end.

Who this is for

For a nonprofit CEO, an operations manager or a board chair who wants to start using AI in an organized way. No technical background needed. You need a Claude or ChatGPT account, a Google Sheets or Excel spreadsheet, and one hour a week for a month.

The problem in three lines

Everyone on the team is already trying something with AI, each on their own and in a different tool. Nobody knows what worked, what went into those tools, or whether someone pasted a donor list in there. And in the end you're left feeling "we tried AI and nothing changed", because nobody chose in advance what to save and how to measure it.

The method in 6 steps

Step 1: A week-long task log

What to do: For one working week, every team member writes down their recurring tasks. Not one-off tasks, only what comes back every week or every month.

How: Open a shared sheet with the following columns, and ask everyone to fill in a row at the end of each day:

• Task: in plain words. "Writing a thank-you email to a donor", "Summarizing the team meeting", "Updating the volunteer list".
• How long: an estimate in minutes.
• How often: daily, weekly, monthly.
• Type of work: writing, summarizing, finding information, numbers, coordination, calls.
• Does it involve personal data: names, phone numbers, health, financial situation, details about minors.
• What happens if there's a mistake: it can be fixed before it goes out, or it goes straight to a donor, a client or a funder.

Explain to the team that the goal is to find work that can be made easier, not to check who works how much. Otherwise the log fills up with the "right" answers.

Common mistake: Skipping the diagnosis and choosing a process because it looked impressive in a demo. That way you pick something shiny that happens once a quarter, instead of something boring that takes three hours every week.

Step 2: Sort the tasks into three groups

What to do: At the end of the week, merge all the rows into one table, and put every task into one of three groups.

How:

You can let the AI sort the table (prompt 1 below). Before you paste it, delete people's names from the table. The tasks themselves are enough.

Common mistake: Putting whatever takes the most time into the first group, even if it's sensitive. At the start it's better to save less and not put people's data at risk.

Step 3: Choose three first processes

What to do: From the "Good to start" group, choose three processes. No more.

How: For every candidate process, answer five questions. A process that gets "yes" on all of them makes the list:

1. Does it recur at least once a week?
2. Does it take at least an hour a week in total?
3. Does it have a single owner on the team who wants to try?
4. Can the result be checked before it goes out?
5. Can it be done without putting sensitive personal data into the tool?

Three processes that fit almost any nonprofit:

• Meeting summaries: record or jot down points, and the AI arranges minutes with decisions, tasks and owners.
• Communication drafts: a post, a newsletter, an email to partners. The AI writes a first draft from material you already have.
• Turning long material into short: an annual report into a post, a program document into one page for the board, volunteer guidelines into a bullet list.

Common mistake: Starting ten processes in parallel. Two weeks later none of them has been checked properly, and you have nothing to show the board.

Step 4: Working rules on one page

What to do: Before you start, write a short document that the whole team reads and approves. It answers three questions: what may go into the tool, what may not, and who checks what goes out.

How: The rules we recommend starting from:

• Never put in: ID numbers, credit card and bank details, medical information, identifying details about service recipients, details about minors.
• Fine to put in: texts already published, general internal documents, your own drafts, tables without names.
• Organization account: work from the nonprofit's account, not a personal one. That way you can close access when someone leaves.
• Privacy settings: check in the tool's settings whether your content is used to train models, and turn that off if you can. Terms change, so check again once a quarter.
• A person checks: anything that goes out (to a donor, a funder, a client, social media) is read by a person. The AI writes a draft, a person approves.
• Facts and numbers: every number, name and date in the text is checked against the source. AI makes up details that sound credible.

If you have a donor or volunteer database, also read the privacy law (Amendment 13) checklist in our playbooks, and check with a lawyer.

Common mistake: Writing a ten-page policy. Nobody will read it. One page, in plain language, with examples.

Step 5: A two-week trial for each process

What to do: Each of the three owners runs their process with AI for two weeks, and keeps a record.

How:

1. Before you start, write down how long the process takes today. That's your starting point.
2. Write a fixed prompt for the process, and save it in a shared document. Don't invent a new prompt every time.
3. After every use, log one row: how long it took, how many corrections were needed, and what the AI missed.
4. Once a week the owner improves the prompt based on what was logged.

If the tool lets you create a project with fixed instructions (Claude and ChatGPT both have this), put the working rules and a short description of the nonprofit there. That way you don't have to explain it again in every conversation.

Common mistake: Judging the process by the first attempt. The first use is always slow. Measure the second week.

Step 6: Decide, share, and add a process

What to do: At the end of the two weeks, sit down for half an hour and decide for each process: continue, change or stop.

How:

• Continue: if time was saved and quality held. The prompt goes into the shared library, and one more person on the team learns to use it.
• Change: if time was saved but there were many corrections. Usually the prompt is missing information or an example.
• Stop: if no time was saved, or checking takes longer than writing. That's a good result too: now you know.

Then present one page to the board or the team: what we tried, how much time was saved compared with the starting point, what we learned. Only then choose the next process from the "Good to start" or "Maybe later" group.

Common mistake: Rolling out to the whole organization after one success, without passing on the rules. Everyone who joins reads the rules page first.

What not to do

• Don't paste donor, volunteer or service-recipient lists into an AI tool, even "just to tidy them up".
• Don't send out AI text that no person has read.
• Don't let AI make decisions about people: who gets a service, who is accepted as a volunteer, who gets a scholarship.
• Don't use AI to reply to people in distress.
• Don't buy a subscription to a specialized tool before you've tried the process in a general one.
• Don't measure success by "how many people used it". Measure time saved and quality.

Ready-made prompts

Replace whatever is in square brackets. Before you paste a table, delete names and identifying details from it.

I run a nonprofit in the field of [field]. Here is a table of recurring tasks the team logged over one week, without names: [paste table].
Sort each task into one of three groups: "Good to start" (recurring, mostly text, no sensitive personal data, a person checks before it goes out), "Maybe later" (touches personal data or numbers), "Not now" (decisions about people, advice to clients, money, or goes out without a check).
For each task write the group and a one-sentence reason. At the end, sum up how many minutes a week are in each group, based only on the data in the table.

From the tasks in the "Good to start" group: [paste].
For each task answer yes or no to five questions: 1) Does it recur at least once a week? 2) Does it take at least an hour a week in total? 3) Does it have a single owner on the team? 4) Can the result be checked before it goes out? 5) Can it be done without sensitive personal data?
Recommend three tasks at most, with a short reason. If there isn't enough information to answer, write "unknown" and don't guess.

Write a one-page draft of working rules for AI tools for the team of a nonprofit in the field of [field], with [number] staff and [number] volunteers.
The rules should cover: what may not go into the tool, what may, working from an organization account, checking privacy settings, a person who checks everything that goes out, and checking facts and numbers.
Plain language, short sentences, one example per rule. No legal terms. At the end add a line: "For questions about personal data, contact [name of the person in charge]".

Here are the points noted at a team meeting on [date]: [paste].
Arrange them into minutes: topics discussed, decisions, tasks (what, who is responsible, by when), and open questions.
Don't add decisions or tasks that aren't in the list. If it isn't clear who is responsible or by when, write [to complete].

Here is [a report / a program document / guidelines]: [paste].
Write from it [a Facebook post of up to 120 words / one page for the board / a bullet list for volunteers].
Use only what is written in the document. Don't add numbers, quotes or achievements. Keep the nonprofit's tone: [warm and simple / formal and to the point].

Here is the log of the trial for the process [process name]: [paste rows: date, time, number of corrections, what it missed].
Starting point: [time before the trial].
Summarize: how much time was saved on average according to the log, which mistakes recurred, and what should be added to the prompt. Recommend: continue, change or stop, with one reason. Don't round numbers in your favor.

One-page checklist

Diagnosis week

• Shared sheet with columns: task, time, frequency, type, personal data, what happens if there's a mistake
• The team knows the goal is to make work easier, not to check up on them
• All tasks sorted into three groups

Before you start

• Three processes chosen, each with an owner
• Working rules page written, read and approved
• Organization account, and privacy settings checked
• Starting point recorded for every process

During the two-week trial

• A fixed prompt for every process, saved in a shared place
• A row in the log after every use
• Prompt improved once a week

At the end

• A decision for every process: continue, change or stop
• A summary page for the board or the team
• The next process chosen

How you know it's working

• Time: how many minutes a week each process takes, compared with the starting point you recorded.
• Corrections: how many corrections each output needs. If the number drops from week to week, the prompt is getting better.
• Steady use: whether the owner keeps using the process a month after the trial, without a reminder.
• Data incidents: how many times someone put data into the tool that the rules forbid. The goal is zero, and every case is a reason to sharpen the rules, not to punish.
• Decisions: whether at the end of the month you have a clear decision for every process. "Stop" is a good decision too.

Disclaimer

This content is general recommendations only, not professional, legal, accounting or tax advice. Every organization is different. You are responsible for checking that it fits your situation, and for checking the prompts and outputs before use. Outside tools, prices, privacy settings and terms of service change, so check them at the source. For questions about personal data and privacy protection, consult a lawyer.