Zero Impulse
Insights

Ideas for your next move

Digital systems & AI

A business chatbot needs a reliable route to a person

Plan an AI chatbot around approved answers and reliable human handoff. Set boundaries, protect account information and test failure cases before launch.

A customer support professional working with a headset and laptop
A customer support professional working with a headset and laptop · Stock photograph: Pavel Danilyuk / Pexels.

A visitor asks whether your team can support an event next month. The chatbot gives a confident answer, quotes an unapproved price and offers no way to reach the person who manages bookings. The conversation sounds helpful, but the business now has an expectation it never agreed to meet.

A useful chatbot starts with boundaries. Decide which questions it can answer from approved information, which details it can collect and when a person must take over. The handoff is part of the product, not an emergency repair.

Give the assistant a defined job

Begin with a manageable purpose, such as explaining services, helping a visitor find a package or collecting the information needed for a quotation. Avoid an instruction that effectively says “handle everything.”

Separate public information from account-specific information. A public assistant can explain the booking process. It should not disclose a customer's records because someone claims to be that customer in a message. Access decisions belong in the application, with appropriate authentication and permission checks.

Identify actions that require approval. Sending an enquiry to the right team is different from approving a refund, changing a contract or confirming availability.

Use a maintained source of truth

Prepare concise, current information about services, package boundaries, contact routes and common questions. Assign an owner to review it when the business changes an offer. A chatbot built on outdated pages will reproduce outdated promises more efficiently.

Make uncertainty explicit in the instructions. Where a price depends on scope, the assistant should explain what information is needed for a quotation. Where an answer is absent, it should say so and offer a next step.

Keep customer-facing language distinct from internal notes. A planning document describing an offer that has not launched should not become public product information simply because it was uploaded into a knowledge collection.

Define the handoff conditions

Escalate when the visitor asks for a person, the assistant cannot find an adequate answer, the conversation repeats without progress or the request requires authority it does not have. OpenAI's practical guidance on agents similarly treats human intervention as an important safeguard for difficult or consequential situations.

Make the handoff concrete. State whether someone is available now, what information will be passed on and how the customer will receive a reply. Do not display “connecting you” if the system merely saves a message for later.

Pass a concise summary and relevant details to the team, subject to the organisation's privacy arrangements. Requiring the customer to repeat everything undermines the value of the assistant.

Test with difficult conversations

Build a test set from genuine enquiries, removing unnecessary personal details. Include vague requests, spelling mistakes, unsupported pricing questions, complaints and requests for another person's information.

Check whether the assistant recognises its limits. Test attempts to make it ignore its instructions or invent an approval. A friendly tone is not evidence that the system is reliable.

Review both the answer and the action. The message may sound correct while the enquiry goes to the wrong inbox or the handoff lacks a contact method. Test the full route from visitor to responsible colleague.

Operate it as a service

Set usage limits, monitor failures and review a sample of conversations. Measure useful resolutions and successful handoffs, not only the number of messages exchanged. A long conversation may indicate confusion.

Give the team a way to correct the knowledge source and track whether the correction improves later answers. Keep a visible contact alternative if the assistant is unavailable.

The strongest chatbot makes routine information easier to access while preserving a dependable route to human judgement. That combination is more commercially useful than a system that tries to sound certain about everything.

Design a website assistant and customer handoff process with Zero Impulse.

Reference: OpenAI's practical guide to building agents.

More useful reading

Explore Digital systems & AI