A useful chatbot starts with reliable answers and a clear handover

A useful chatbot helps someone make progress. It might explain how a service works, find the right guidance or help a colleague answer a recurring question. The best starting point is a clear job and a route through the whole enquiry, including what happens when a person needs to take over.

For a business or charity, that creates a practical opportunity: make everyday information easier to use while giving your team more context for the conversations that need their attention.

Choose a job you can describe

Start with one group of users and a set of questions your team recognises. An internal assistant for staff guidance and a public enquiry service have different audiences, information and responsibilities. Give each a clear purpose.

For an illustrative charity example, a visitor might want to understand a service's opening hours, referral process and what to bring to an appointment. The chatbot could guide them through approved information, then offer a contact route for a conversation about their circumstances. It should be clear which parts are general guidance and which require a colleague's judgement.

Agree what a successful interaction looks like before choosing software. It might be finding the right page, understanding the next step or reaching the appropriate team with a useful summary.

Give answers a dependable starting point

Choose the information the service is allowed to use. Identify an owner for each source, resolve conflicting versions and agree how changes will reach the chatbot. Include the less tidy details your colleagues know: exceptions, different service areas and guidance that has expired.

Where the tool provides source links, test whether those links support the answer and whether the user can open them. Connecting documents is a useful foundation, but it still needs testing. NIST's Generative AI Profile recommends checking generated citations and sources during evaluation and ongoing monitoring.

Our data and systems readiness work can help clarify what information is available, who maintains it and how it should be accessed.

Make the handover part of the service

Design the route to a person at the same time as the automated conversation. Agree when it should be offered, which team receives the enquiry and what happens outside staffed hours.

A visitor should understand whether they are entering live chat, requesting a callback or sending a message for a later response. Where the system supports transferring a summary, let the visitor check it and explain what will be passed on. Confirm what has actually been sent; a suggestion to contact someone is different from a completed referral.

The government's Service Manual guidance tells government service teams to make the use of AI clear and provide a human contact route. That is a useful design reference for businesses and charities too, rather than a claim that government service rules apply to every organisation.

Test complete conversations

Build a test set from representative questions, using synthetic or appropriately approved examples. Include the straightforward requests that should work well, questions with missing details, wording your team would not use and requests outside the agreed scope.

Check the whole journey: answer, source, clarification, handover and confirmation. Try changing a source document and confirm the service responds to the update. Check that one user cannot retrieve information intended for another.

Include the colleagues who will maintain the information and receive escalations. Their feedback can show where a helpful answer still leaves extra work for the team. Keep a record of what worked, what needs improving and what must be resolved before release.

Give the live service an owner

Plan who reviews feedback, fixes information and checks the service after a change. Decide how to pause a failing feature and keep another contact route available.

Human oversight also extends beyond individual replies. The UK Government AI Playbook discusses testing before deployment, checks during operation and user feedback for services such as chatbots. These are useful practices to adapt to the size and purpose of your service.

Bring a recurring enquiry, the guidance behind it and the point where a colleague normally gets involved. We can use those to explore a focused chat and knowledge service, connect it to your AI in Action priorities and plan the review arrangements needed to keep it useful.