Guides · AI at work
A Chatbot for Your Nonprofit: Answer 5 Board Questions First
By Josh Croom · October 27, 2026 · 10 min read
A chatbot for nonprofits answers the routine questions on your website, like your hours, your services and how to give, at any hour of the day. Before yours goes live, write down answers to the five questions your board will ask, test the chat with hard messages, and read every chat for the first two weeks.
A question after the office closes
11:48 PM. A mom finds your food pantry's website late on a Sunday night. She wants to know if she can come in tomorrow without a photo ID, and your office closed hours ago. So she types her question into the little chat window in the corner of the page.
Your board will want to know what the chat says next. A good chat tells her it's a bot and offers a text from a person in the morning, because it only knows what's on your fact sheet. A bad one guesses in a sure voice and can send her across town on a rule your pantry never made.
Can I come to the pantry tomorrow if I don't have a photo ID?
I'm an automated helper, not a person. The pantry is open Monday and Thursday from 9 AM to noon.
Our fact sheet doesn't say anything about ID, so I won't guess. Would you like someone from our staff to text you the answer in the morning?
Yes please. This number is fine.
Got it. Someone from our office will text you after 9 AM.
What a nonprofit chatbot is good at
Many of the messages a nonprofit gets ask the same few questions. People want your hours, your address, where to park, what you offer, who it's for, how to give and how to volunteer. Your staff answers them by phone and email all week, and each one takes a few minutes they could spend with the people you serve.
A chatbot can answer those questions at any hour, from a fact sheet you control. That's the right job for it. It shouldn't decide who qualifies for help, give medical, legal or money advice, or handle a person in crisis on its own. When a question lands outside the fact sheet, a good chat says so and offers a person.
What the chat answers, and what goes to a person
Draw this line before you set anything up, because it decides what goes in the fact sheet and who gets the alerts.
Send these to a person
- Whether someone qualifies for help in their situation.
- Anything medical, legal or about money.
- A person who says they're hurt, in danger or in crisis.
- A complaint, or a question about a past visit.
Let the chat answer these
- Your hours, your address and where to park.
- What you offer, and what you don't.
- How to give, and where to send a check.
- How to volunteer, and who to call next.
Why your board will ask hard questions
People type real questions into search, and many of them don't fully trust AI yet, so the chat has to answer correctly and hand off to a person when it can't.
- 19%of clicks on the pregnancy center search ads we run came from a search typed as a questionsearches that start with words like how, what, can or am
- 60%of U.S. adults would be uneasy if their own health provider leaned on AIPew Research Center, a survey of 11,004 adults
Two times a chatbot got it wrong
In November 2022, after his grandmother died, a man asked Air Canada's website chatbot about the lower fare the airline offers after a death in the family. The chatbot told him he could buy a full-price ticket and ask for the lower fare within 90 days. The airline's own policy said the request had to come before the trip, so it turned him down. He took it to a Canadian tribunal, and in February 2024 he won.
The airline had argued that its chatbot was a separate legal entity, responsible for its own actions. The tribunal member called that argument remarkable and turned it down. His ruling held the airline responsible for the information on its own website, and the chatbot's answers were part of that website.
In 2023, the National Eating Disorders Association took its chatbot offline after people reported that it gave out weight-loss tips. Researchers had first built the chatbot to give only prewritten answers. The association said the company running it added AI answers without telling it, while the company said the change was part of their contract. Either way, the people the nonprofit served got advice it never meant to give.
Your organization owns every word its chat says, as the tribunal told Air Canada. Someone on your side also has to know exactly how the chat answers and who can change it. Get that in writing from whoever runs your chat.
Five steps before the chat goes live
Do these before the board meeting, because they produce the answers your board will want to read.
Step 1
Write the fact sheet
List your hours, address, services, what you don't offer and your phone number on one page. The chat should answer from this sheet alone, so anything you leave off the sheet is something it can't answer.
Step 2
Name the alert person
Pick one person who gets crisis alerts and requests to talk to a person during the day, and one more for nights and weekends. A chat that promises a callback needs a real name behind it, or it's making a promise nobody keeps.
Step 3
Test it with hard messages
Before launch, type the messages you hope never come: someone in danger, someone who is hurt, a medical question and a question about money. Add easy ones about cost and hours, try each one in English and in Spanish, and write down exactly what the chat did.
Step 4
Read every chat for two weeks
Someone on your staff reads every conversation for the first two weeks. When an answer is wrong, fix the fact sheet, because that's the only change that stops the same mistake from coming back tomorrow.
Step 5
Write down the data answers
Record where the chats are stored, who can read them and when they're deleted. Put those answers in the board minutes, so the decision and the reasons for it sit in one place.
The five questions your board will ask
Adapt each sample answer to your own chat, check that it's true and fill in the blanks from your own testing before you hand it to your board. If any answer comes out vaguer than a yes or no, a name, a number of days or a test result, the chat isn't ready yet.
- Does it tell visitors it's a bot? Yes. The chat window tells visitors they're talking to an automated helper, and it shows them how to reach a person.
- Can it give medical, legal or money advice? No. It tells the visitor it isn't a nurse, a lawyer or an advisor, and it points them to our staff.
- What does it do when someone is in danger? When we tested danger messages on (date), the chat (what it did), and (name) got the alert within (minutes).
- What happens when it doesn't know the answer? It answers only from our fact sheet. When the sheet doesn't cover a question, it says so and offers a call or a text from a person.
- Who can read the conversations, and for how long? (Names) can read them, and each conversation is deleted after (number) days.
How to tell if the chat is working
After the first two weeks, look at four counts each month. Count the chats, and how many came in after your office closed. Count the questions the fact sheet couldn't answer, because each one is a line to add. Count the handoffs to a person, and check how soon someone followed up. Then count the wrong answers you found, and watch whether that number drops as the sheet gets better.
A chat also shows you how people describe their needs before they reach you. If your nonprofit runs the Google Ad Grant, the questions people type into your chat are a list of words to check against your ads and your pages. Our Google Ad Grants guide explains how the Grant works and what Google expects from it.
When the chat promises a callback, someone has to make it. The five-minute callback routine works the same way for a chat handoff as it does for a missed call: one named person, one phone and a fast reply.
Is your chat ready for the board?
Answer five quick questions. Your answers stay on your device.
Does the chat answer only from a fact sheet your staff wrote?
Is one person named for alerts during the day, and one for nights?
Have you tested it with danger, medical and money messages, and written down what it did?
Will someone read every chat for the first two weeks?
Can you say who reads the chats and when they're deleted?
You're ready to bring it to your board. Put the five answers on one page, and attach your test notes so the board can see what the chat actually did.
You're close. Fix the first question you answered No before the meeting, because that's the one your board will ask about first.
Hold the launch for now. Start with the fact sheet and the alert person, since every other answer depends on those two.
Before the chat goes live
Tick each one off as you go. Your ticks stay on this device.
0 of 7 done
Talk it through with Miguel
We build websites for nonprofits and run Google Ad Grants, and our AI agent can answer website chat from what we teach it about your organization. If your board is weighing a chat, book a call with Miguel. He'll ask how questions reach you today and who reads them at night, and then you can decide together what fits.
Questions people ask
Should a nonprofit use a chatbot on its website?
A chatbot makes sense when your staff answers the same questions every week and someone can read the alerts at night as well as during the day. If nobody can own the alerts yet, wait, because a visitor who's promised a person and never hears back stops trusting you.
What should a nonprofit chatbot never do?
It should never give medical, legal or money advice, decide who qualifies for help, or handle a person in danger on its own. It should say it's a bot, answer only from your fact sheet and hand everything else to a person.
Who is responsible when a chatbot gives a wrong answer?
Plan as if your organization is. In a 2024 case in Canada, a tribunal held an airline responsible for what its website chatbot told a customer. Ask your own lawyer how the law applies to you, and keep the answer on file.
How long should we read every chat?
Read every chat for at least the first two weeks. After that, read a sample each week, plus every chat that ended in a handoff or a complaint.
Can a chatbot help someone in a crisis?
It can point them to help fast, but it shouldn't try to counsel them. Give the chat 911 and the 988 Suicide and Crisis Lifeline, and set it to alert a named person right away.
See the websites we build for nonprofits
Read next
- AI for Churches: Let AI Draft the Week So Your Staff Has Time for PeopleWhere AI fits in a church office, what stays with people, and three prompts to start with.
- AI for Nonprofits: How a Small Team Can Hand Off the PaperworkHow a small nonprofit team can hand AI the reports and letters, and what stays with a person.
- AI for Pregnancy Centers: Use It for the Office, Never for Her FileWhere AI saves a center's staff hours, and the line that keeps every client's story out of it.