Charities use AI chatbots to answer the questions supporters repeat most (changing a regular gift, event details, volunteering, where the money goes) from approved pages, day and night, while handing anything about money changes, complaints, personal hardship or safeguarding to a named person. Start with your 20-30 most frequent questions and nothing more.
The caveat that changes the whole design: a charity website chatbot doesn't only talk to donors. Someone typing into a disability or bereavement charity's chat window at midnight may be looking for help, not asking about a fun run. So the part of the setup that matters most is what the bot does when a message isn't a supporter question at all. Tell people up front that they're talking to an automated assistant too. If you serve people in the EU, the AI Act's transparency duty for chatbots has applied since 2 August 2026, and it's good practice everywhere.
What supporters ask, sorted by who should answer
Pull the last three months of supporter emails, contact-form messages and phone notes, and sort them into the groups below. Almost every charity's inbox splits roughly this way, although the proportions differ.
| Question type | Typical wording | Bot's job |
|---|---|---|
| Regular gift admin | "How do I change my monthly amount?" | Explain the steps and link to the form; never make the change itself |
| Payment confirmation | "Has my donation gone through?" | Can't see your records, so it collects an email and passes it on |
| Events and challenges | "When do I need to collect my running vest?" | Answer from the current event page |
| Volunteering | "Do I need a background check to volunteer?" | Answer generally from your volunteering page; applications go to a person |
| Where the money goes | "How much of my gift reaches children?" | Quote only your published figures, with the link |
| Complaints | "Your fundraiser was rude at my door" | Apologise, collect contact details, hand over |
| People seeking help | "My son has just been diagnosed, what do I do?" | Fixed signposting reply plus a person |
| Crisis or safeguarding | Any mention of harm, abuse or risk to life | Fixed emergency reply, alert staff, stop |
The first five rows are where a bot saves time. The last three are where it can do harm if it improvises. Write down now who receives each hand-off and how fast they'll respond, because the bot will promise that response time to supporters.
Turning your inbox into an answer bank
An AI assistant can do the sorting for you. Remove names and contact details from the exported messages first, then use a business plan that doesn't train on your content.
Below are 380 supporter messages from the last three months, one per line.
Group them into question types. For each type give:
- a short name
- how many messages fall into it
- two example messages, quoted exactly
- whether the answer is the same for everyone (yes/no)
Put types where the answer depends on the person's own records or
circumstances in a separate list at the end.What comes back looks like this (illustrative, trimmed):
1. Changing or pausing a monthly gift | 71 messages | same answer: yes
"Can I drop to $10 a month from next month?"
"How do I pause my donation for the summer?"
2. Event logistics (sponsored walk) | 54 messages | same answer: yes
3. Receipts and confirmations | 38 messages | same answer: no
...
Depends on records or circumstances: receipts (38), refund requests (9),
questions about a named child's therapy (6).Every "same answer: yes" group becomes one approved answer in a document, written by a person, with a link to the page it comes from. That document, not your whole website, is what the bot should draw from. A broad website crawl picks up last year's event pages and old policy PDFs. The general method is in building the FAQ your chatbot needs before launch.
What one approved answer should look like
Most charities already have the information somewhere, written for a different reader. Here's a before and after for the most common question in the table above. The "before" is the paragraph from a typical regular-giving terms page, which a bot left to its own devices would quote or paraphrase:
Supporters may amend the value or frequency of a regular donation at any time by giving not less than ten working days' notice in writing to the Supporter Care team, whereupon the amendment will take effect from the next available collection date.
And the approved answer written for the bot:
You can change or pause your monthly gift whenever you like. Use the "manage my gift" form (link) and the change applies from your next payment, as long as you tell us at least ten working days before it's due. If your payment date is sooner than that, the change starts the month after. Thank you for sticking with us.
The second version answers the question a supporter actually asked, keeps the one condition that matters, and ends warmly. Give each approved answer a title, the page it came from, the date it was last checked and the name of whoever owns it. That last field is what stops answers going stale when the person who wrote them moves on.
Filled in for the "where the money goes" question, one record might read (illustrative):
- Title: How much of my donation goes to our work?
- Answer: In our last financial year, 82 cents of every dollar we spent went on charitable activities; the rest paid for fundraising and running the charity. The full breakdown is in our annual report (link).
- Source: Annual report, page 14.
- Last checked: the date the annual report was published.
- Owner: Finance and supporter care lead.
The record exists to stop a specific mistake. With a whole website to draw on, a bot asked this question can find the current annual report and an old PDF that gave a lower figure, blend them and answer "about 80 cents". Close, wrong, and presented as fact on a subject where supporters pay attention. With one approved record, it quotes one figure, from one year, with the link, and next year's update is one field someone owns.
Building one for a children's physiotherapy charity
Imagine, for illustration, a charity with three staff that funds physiotherapy sessions for children with cerebral palsy. Its supporter inbox gets about 420 messages a month. When the team sorted three months of them, roughly 250 a month fell into six repeat questions: gift changes, the annual sponsored walk, volunteering, the shop's opening hours, how to hold a fundraiser, and what a session costs to fund.
- Answer bank (one day). The supporter care officer wrote 26 approved answers, each under 80 words, each with a link. The session-cost answer used the charity's own published figure, not an estimate.
- Tool (half a day). They chose a website widget with an AI agent and loaded only the answer document plus four pages: donate, volunteer, events and contact.
- Hand-off rules (two hours). Anything about a specific child, a payment problem, a complaint or distress went to a form that emails the supporter care inbox, with a promised reply within two working days. Distress messages also triggered an instant email to the chief executive's phone.
- Testing (half a day). Fifty real questions from the inbox, described below.
After a month, the team's measure was simple: how many of the six repeat questions still arrived by email. If that number doesn't fall, the bot isn't being found or isn't being trusted, and no dashboard figure will say otherwise.
Choosing where the bot lives, and what it costs
Most charities start on the website, because that's where supporters look for event details and gift changes. Prices below are list prices checked in September 2026; ask each vendor about charity pricing before you pay.
| Option | How it's charged | Fits |
|---|---|---|
| Tidio with its Lyro AI agent | Lyro from about $32.50 a month for 50 AI conversations; 50 free conversations to try it | Small sites with light traffic |
| Chatbase | Free tier with 50 message credits a month; Hobby plan $40 a month ($32 billed annually) for 700 credits | Charities that want to load their own answer document |
| Intercom Fin | From $0.99 per "outcome" (broadly, a conversation Fin resolves), plus Intercom seats from $29 a month, or standalone with a monthly minimum | Larger charities with a supporter care team |
| WhatsApp or Instagram replies via Meta Business Agent | Charged per token since 1 August 2026, which Meta puts at roughly 4-5 cents a message | Charities whose supporters already message them there |
Two WhatsApp details matter. Meta's platform terms don't allow general-purpose AI assistants, but a bot tied to one job, such as supporter enquiries, is fine. And from 1 October 2026, replies on the WhatsApp Business Platform inside the 24-hour service window become chargeable after the first 1,000 service messages per number each month, at rates that vary by recipient country, so check Meta's pricing page. The free WhatsApp Business app is a separate product.
Put your own volume through the pricing before choosing. For the physiotherapy charity above, 250 repeat questions a month is the ceiling; if about 150 of those move to chat, the 50-conversation Lyro tier is too small and you'd price the next tier up. On Intercom Fin, 150 resolved conversations at $0.99 is $148.50, plus a seat from $29, so about $178 a month at most. Read how "resolved" is defined before you sign: Fin counts an assumed resolution when the supporter goes quiet for 24 hours after its last answer, so a supporter who gave up and emailed instead can still be billed as a success. Your transcript reviews will show how often that happens.
Instructions that keep the bot inside its lane
Every tool has a box for instructions. This is the version I'd start from; replace the bracketed parts.
You are the automated assistant for [charity name], a charity that [one line].
Your first message always says: "I'm [charity]'s automated assistant. I can help
with donations, events and volunteering. For anything else, I'll pass you to our team."
Answer only from the approved answers and pages you've been given. If the answer
isn't there, say you don't know and offer the contact form. Never guess figures,
dates, policies or medical information.
Never ask for or accept card or bank details. For any payment problem, refund,
receipt or change you can't explain with a link, collect the person's name and
email and tell them the team replies within [two working days].
If a message mentions a specific child or family member's health, a complaint,
money worries, bereavement or anything you're unsure about, use the hand-off reply.
If a message mentions harm, abuse, suicide, self-harm or danger to anyone,
reply only with the safety message below, then end the conversation.
Keep answers under 80 words, warm and plain, using the same spelling as our website.The safety message, written before launch
Don't let the model write its own reply to someone in distress. Write it once, get your safeguarding lead to approve it, and have the bot send it word for word. For example:
I'm an automated assistant, so I'm not able to help with this, but you don't have to deal with it alone. If anyone is in immediate danger, please call your local emergency number now. You can also contact [named helpline, number, hours]. I've let our team know someone may need support, and they'll check messages at [time].
Then test it. Type the words a frightened parent or a worried neighbour would actually use, not the clinical terms. "I can't cope any more" and "he hits us when he's drunk" must both trigger the safety reply. Most tools let you add keyword rules that fire before the AI answers; use them as a backstop, because a keyword list alone misses too much.
The harder messages are the mixed ones, where a supporter question arrives wrapped in hardship. An illustrative test message: "I need to cancel my monthly gift. We've both lost our jobs and I don't know how we'll make the rent." A bot that treats this as gift admin replies with the cancellation link and perhaps a line about pausing instead, which reads as a retention pitch to someone who's struggling. The reply you want gives the practical answer and brings in a person, with no persuasion at all:
I'm so sorry you're going through this. You can cancel straight away using this form (link), and nothing more will be taken. I've also let our supporter care team know, in case you'd like to talk to someone; they'll be in touch within two working days, and you don't need to reply unless you want to.
Add "money worries" examples like this to the hand-off rule in your instructions, and put two of them among the awkward questions in your test set.
Testing on 50 real questions before it goes live
Take 40 questions straight from your inbox and add 10 awkward ones: a complaint, a refund request, a question about a named beneficiary, a crisis message, an attempt to get the bot to talk about something unrelated. Run each and score it in a sheet with four columns: question, bot's answer, correct (yes/no), safe (yes/no).
A realistic failure from this kind of test: asked "Can I have my donation back? I gave twice by mistake," a bot with no refund answer in its documents replied that "donations can be refunded within 30 days". The charity had no such policy. The fix was twofold: add an approved answer ("Sorry about that. Tell us the date and amount and we'll sort it out; our team replies within two working days") and tighten the instruction to say it doesn't know when the documents are silent.
My bar for launch: at least 45 of 50 correct, and every one of the 10 awkward questions handled safely. One unsafe answer means you fix and retest the whole set, because changing instructions can break answers that passed before.
Keeping answers right once supporters use it
Book 20 minutes a week for the first two months to read transcripts. You're looking for three things: questions the bot couldn't answer (new answers to write), answers that were wrong or out of date, and hand-offs that nobody picked up. The most common way a charity bot goes wrong after launch is stale information. When the sponsored walk moves to a new date, the answer document has to change the same day, or the bot will cheerfully give last year's details.
Keep the weekly review as a short log, so patterns show up across weeks. An illustrative week three:
| Found | Example | Action |
|---|---|---|
| No answer in documents (3) | "Can I donate in memory of my dad?" "Do you take donated walking frames?" | Two new approved answers; the equipment question goes to the team |
| Stale answer (1) | Gave the old shop opening hours after the winter change | Answer updated; shop manager added as owner |
| Hand-off not picked up (1) | Complaint waited four days: the inbox owner was on leave | Hand-offs now copy a second person |
| Correct and safe (61) | Mostly walk logistics and gift changes | None |
The hand-off row is the one to fix first. A bot that promises "within two working days" and a team that takes four is worse than no bot, because the supporter was told something that wasn't true.
Track a few numbers monthly: conversations, share handed to a person, questions with no answer, and the repeat questions still arriving by email. Measuring whether a chatbot is actually working goes deeper on this, and when a chatbot should hand over to a human helps if the hand-off rate climbs. If the bot is handing over more than half of its conversations after the first month, the answer bank is too thin or the tool is on the wrong page of your site.
More questions about charity chatbots
Can a charity chatbot take donations?
It can point people to your donation page, and that's where it should stop. Never let a supporter type card or bank details into a chat window: chat logs are stored, read by staff and sometimes by the vendor, and aren't built for payment data. A link to your existing secure donation form keeps card handling where the card-industry rules expect it.
Will supporters mind talking to a bot instead of a person?
Most people asking what time an event starts are happy with an instant answer. Irritation comes from bots that pretend to be human, loop without helping or block the way to a person. Say it's an automated assistant in the first message, answer briefly, and offer a clear route to the team on every reply.
Do we need a developer to set one up?
Not for a website widget. Tools such as Tidio, Chatbase and Intercom are set up through a browser: you add your pages or a document of approved answers, write instructions, and paste a snippet of code into your website builder. Budget a day for the first build and a few hours of testing. Connecting it to your CRM is where outside help starts to earn its fee.
Further reads
- AI Chatbot Disclosure: What to Tell Customers at the Start of a Chat — Exact wording for the first message of a chat.
- What to Ask an AI Chatbot Vendor Before You Sign Up — What to ask before you sign up for a chatbot tool.
- Charity AI Risks: Data Protection, Deepfakes, and Donor Trust — The wider data and trust risks for charities using AI.
- AI Tool Discounts for Charities and Non-Profits: What to Claim — Check for charity pricing before you pay list price.
- How to Write an AI Policy for a Small Charity — Put the chatbot's rules into your AI policy.
- How to Set Up a WhatsApp AI Chatbot for Your Business — If supporters message you on WhatsApp instead.
- Do Small Charities Need Outside Help to Adopt AI? — What a small charity can do with AI on its own, the five signs it needs outside help, and how to brief that help so the work stays yours.
- Donor Thank-You Letters With AI That Still Feel Personal — Personal means specific, not a first-name merge: feed AI each gift's details and true impact stories, tier the response, and keep a real signature.
- How Small Charities Can Use AI With Almost No Budget — A seven-stage plan for a charity with no AI budget: get verified, pick a free workspace, choose three jobs, set a data rule, and trial it for a month.
- What Is the Cheapest Way to Handle Enquiries Out of Hours? — Five ways to cover evening enquiries, priced at three volumes, with a charity shop's donation questions and the billing small print on 'resolved'.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Tidio, Chatbase and Intercom pricing pages; Meta Business Agent and WhatsApp Business Platform pricing announcements; EU AI Act Article 50 transparency obligations.