A business chatbot needs a defined audience, permitted knowledge, supported tasks and a usable escalation path. Decide whether it only explains information or may also initiate actions. The chat window is one part of the product; content ownership, access control, uncertainty handling, operational support and evaluation determine whether people can rely on the complete workflow.
Let’s plan your business chatbot.
Bring your workflow and the decisions you need to make. We’ll help you define the requirements, examine the trade-offs and plan a practical next step.
Write the service boundary
List the questions or tasks the first release supports and the conditions that require a person. Specify users, languages, channels and information sources. Avoid an opening promise that the assistant can answer anything about the business if its actual evidence covers one product line.
Distinguish public documentation from customer-specific records. An answer about a standard return policy and an answer about a particular customer's refund require different sources and access rules.
Define the knowledge and response contract
| Requirement | Decision to make | Acceptance example |
|---|---|---|
| Source ownership | Who approves and updates content? | Replaced policy stops being treated as current |
| Access | Which user may retrieve which records? | Another account's information is unavailable |
| Evidence | When must an answer show its source? | Citation supports the actual answer |
| Missing information | When should the bot ask a question? | Ambiguous product request is clarified |
| Uncertainty | When should it stop answering? | Conflicting policies route to review |
| Escalation | Who receives unresolved work? | Ticket contains relevant context |
| Action | What can change outside chat? | Sensitive change follows required authorisation |
Adapt these tests to the specific service. A citation or an escalation button is not evidence that the supporting process works.
Walk through a support question
Suppose a customer asks whether a replacement can be delivered tomorrow. The chatbot may have a shipping policy but no confirmed stock or carrier commitment. It should not turn a general policy into a specific promise.
Define the next step: request required details, retrieve permitted live information, or pass the request to staff. If a human must check availability, preserve that unresolved state. A friendly answer is not a substitute for a verified commitment.
This hypothetical scenario illustrates a requirements boundary, not observed performance of a particular model or a customer deployment.
Treat documents as information, not authority
Retrieved pages or messages can contain instructions that attempt to redirect the system. Enforce permissions and tool boundaries independently of generated text. OWASP's prompt-injection guidance describes malicious instructions embedded in external content and the need for layered controls.
Test whether the system follows an instruction hidden inside a support document to reveal another record or call an unrelated tool. A prompt asking the model to behave safely does not replace independently enforced access rules.
Make escalation operational
Define the receiving queue, required context, ownership and response expectations. The customer should know what has happened and whether they need to provide more information. Track unresolved handovers rather than counting every chat closure as a successful resolution.
Measure completed supported tasks, unsupported answers, useful escalations and staff correction effort. Review failures by task type so easy questions do not hide recurring problems with particular workflows.
Questions buyers ask
Should the chatbot take actions immediately?
Only within a clearly authorised and tested boundary. A knowledge assistant can be useful without the ability to change customer records or make commitments.
Will adding more documents improve answers?
Only if those documents are relevant, current, accessible to the user and retrieved appropriately. Conflicting material can create additional work.
Prepare a chatbot brief
Bring representative questions, approved sources and escalation examples. Use the RAG comparison guide and discuss AI development around a bounded first service.
