The Handover Is the Product: Designing Charity AI That Knows Its Limits

There is a particular kind of silence at the end of a bad chatbot conversation. Someone has already told the bot they are struggling with rent, or caring for a parent, or trying to find a food bank that is open on a Saturday. The bot has offered three links, one of which is broken, and then a cheery prompt: “Is there anything else I can help with?” That silence is not efficiency. It is a door quietly closing.

AI chat and triage tools are turning up in UK charities for understandable reasons. Advice lines are overwhelmed. Small teams are answering the same questions again and again. Funders like the sound of innovation. And there are genuinely useful things a well-designed bot can do: signpost to opening hours, collect basic details, translate a first query, or help someone find the right form. But the most important design decision is not the welcome message or the tone of the typing indicator. It is the handover. The moment the bot stops being useful and a human being takes over.

If that moment is badly designed, the whole thing becomes a more polished way of leaving people on their own.

Start with the exit, not the entry

Most chatbot projects begin with a flow diagram that maps the first question. That is the wrong end. Start by mapping every point where a person should be able to leave the bot and reach a human. Then make those exits visible, early and repeated.

Some exits are practical: someone needs to change an appointment, make a complaint, or ask about a grant. Some are ethical: someone discloses abuse, mentions suicidal thoughts, or describes a situation that clearly needs a trained adviser. Some are simply human: the person has tried twice, the bot has misunderstood, and they are getting angrier by the second.

A good rule for 2026: if a bot cannot confidently help within two exchanges, it should offer a person. Not after five more menus. Not after a feedback survey. Within two.

Transparency that sounds like a person wrote it

People do not need a legal essay before they type. They do need to know what they are talking to, what it can and cannot do, and what happens to their words. Put that in plain English at the start, and keep it reachable throughout.

Try something like: “I’m an automated assistant. I can help with basic questions and find the right team. I can’t give advice about your specific situation. You can talk to a person at any time.” Then make “talk to a person” a real button, not a phrase buried in a help centre.

Be honest about waits. If the team replies within two working days, say so. If the phone line is open 10am to 2pm, say that too. A bot that implies instant human support and then delivers a three-day silence is not transparent. It is just a queue with better manners.

The handover has to land somewhere

A handover is not a link to a contact form. It is a warm transfer of context, consent and expectation. If I have already told the bot that I am a carer in Leeds looking for respite support, the human should not begin with “How can I help?” as if we have never met.

Design the handover so that the person does not have to tell their story twice. Pass on the conversation, with the person’s knowledge. Show a realistic next step: a callback time, an email reply window, a phone number, or an out-of-hours route. If no one is available, say what happens next and when. “We’re closed now. Leave a message and we’ll call you back tomorrow before noon” is a hundred times kinder than a dead end dressed up as a chat window.

This is where charity AI projects often reveal their real budget. The bot gets the investment. The human queue gets the leftovers. But the handover is the product. If it is understaffed, undertrained or invisible, the bot is just a more efficient way of saying “not now”.

When the bot gets it wrong

Every bot will get it wrong. The question is whether the organisation has designed a dignified way to recover.

Make it easy to say “that’s not right” or “I need a person”. Do not make people argue with a machine to prove they deserve help. If the bot gives incorrect information, acknowledge it, correct it, and route the person to someone who can help. Log the failure without turning the person into a test case. And review transcripts with consent, care and a clear purpose, not just to improve “engagement”.

Safeguarding triggers need special attention. A bot should not attempt to counsel someone in crisis. It should recognise certain language, stop the usual flow, and offer a human and relevant emergency information immediately. That is not a technical feature. It is a duty of care.

A short audit for your next sprint

Before you add another flow, ask these questions:

• Can I reach a human within two clicks, from any point in the conversation?
• Is it obvious that I am talking to a bot, and what it can and cannot do?
• Are waiting times and opening hours described honestly?
• Does my context travel with me to the human, with my consent?
• Is there a clear out-of-hours route that does not leave me stranded?
• Can I opt out of the bot without losing access to support?
• Are people in distress or at risk routed to a human quickly and safely?
• Is the language plain, warm and free of jargon?
• Does the bot avoid giving advice it is not qualified to give?
• Who owns escalation, and what happens when they are on leave?
• Is there a kill switch if the bot goes badly wrong?
• Are we measuring good handovers, not just bot “resolution” rates?

That last one matters. If the only metric is how many conversations the bot closes, you will design a bot that closes conversations. If the metric is how well people get to the right human support, you will design something more honest.

Efficiency with a conscience

AI can take pressure off a charity, but it cannot replace capacity. The most purposeful use of a chatbot is not to keep people away from your team. It is to help them arrive better prepared, at the right door, with their dignity intact.

So design the exit before you design the entry. Staff the handover. Test it with disabled supporters, older people, people with low digital confidence and people who have actually used your service. Then keep listening after launch.

A chatbot that knows when to stop is not less ambitious. It is more trustworthy. And for a charity, trust is not a nice extra. It is the whole point.