A custom AI chatbot project needs a defined audience, an approved source of answers, and a usable route to a person. Start there before choosing a model or chat widget. A visitor asking about your services needs a different experience from an employee looking up a private account record.
When you compare chatbot development agencies, ask each one to describe the same complete conversation. Where does the visitor arrive, what can they ask, which information is collected, and what happens if the bot cannot help? Those details determine both usefulness and scope.
Related service: AI chatbot development services- 01Explain what the assistant can help with.
- 02Answer from approved, accessible sources.
- 03Ask only for information needed for the task.
- 04Escalate uncertain or unsupported requests.
- 05Deliver the context to the responsible person.
SECTION 01
Choose the chatbot's first job
A public website assistant can explain services and collect an inquiry. A support assistant might retrieve an authenticated customer's order. An internal knowledge bot might search approved staff documents. These are different products even if all three use a chat interface. Pick the first audience and completion condition explicitly.
A configured tool may meet a simple FAQ requirement. Custom development becomes relevant when the conversation must fit a particular website, preserve role permissions, collect structured information, or connect with a business system. Ask the agency to show which parts are configuration and which become software you maintain.
SECTION 02
Prepare answer sources and name their owner
Provide the pages, documents, policies, and product information the bot may use. Remove expired offers and resolve conflicting answers before import. Assign a person who can approve changes. A chatbot should not invent an answer because two departments maintain different versions of the same policy.
Define whether the answer needs a source link, when the assistant should ask a clarifying question, and when it should say the information is unavailable. For private content, source permissions must be enforced by the application and retrieval layer. A message telling the model to keep information confidential is not an access-control system.
SECTION 03
Worked example: a product support assistant
Imagine a small equipment supplier whose visitors repeatedly ask about product compatibility and installation documents. An illustrative first release searches approved manuals, links to the relevant document, and collects a product model when the question is ambiguous. It offers a support request when no verified answer is available.
An account-specific order lookup is a separate scope option. It requires identity verification, a permitted order API, and rules about which fields can be displayed. The public knowledge bot should not receive every customer's order history just to answer general product questions.
For handoff, the visitor confirms a short summary and contact details before submission. The support team receives the question, selected product, relevant conversation context, and any source already shown. A visible confirmation explains that the request was received; it does not promise an unapproved response time.
SECTION 04
Specify the interaction as carefully as the answers
Review the chat on a small screen and with a keyboard. Users need to find the input, send a message, read errors, and reach a person without losing their place. W3C's forms guidance emphasizes understandable labels, instructions, validation, and feedback. Apply those basics to chat intake fields as well.
Document whether transcripts are stored, who can review them, and how long they are needed. Do not ask for unnecessary personal information. If the assistant can create a ticket or update a CRM record, give that action a narrow permission boundary. OWASP's guidance on excessive agency supports limiting tool access and high-impact actions.
SECTION 05
Use conversation-based acceptance tests
Prepare realistic questions with expected behavior, not just a list of topics. Include spelling mistakes, follow-up questions, an unsupported request, conflicting documents, an attempt to obtain another customer's data, and an unavailable integration. Check answer quality and handoff completion separately.
Ask the agency to demonstrate a source update, a failed ticket submission, and the method for disabling the bot. In a remote engagement, agree who reviews conversations, how feedback becomes a change request, and which launch checks need your team present. Match language and support hours to actual operating capacity.
Keep the first release affordable by reducing channels, source collections, or optional actions—not by leaving the human handoff undefined. Contact us for pricing. Use the budget guide below when comparing a configured assistant with a more integrated custom build.
SECTION 06
Frequently asked questions
What information should we give a chatbot development agency?
Share the audience, common questions, approved sources, handoff team, channels, existing systems, and any actions the bot may take. Include examples where it should refuse or ask a person for help.
Can an AI chatbot answer every customer question?
No. Its scope and sources determine what it can support. A useful chatbot explains uncertainty and routes unsupported requests to a person instead of inventing an answer.
Can a website chatbot connect to our CRM?
It may be possible when the CRM exposes a suitable API and grants the required access. Define the fields, validation, duplicate handling, permissions, and failure behavior before including the connection.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- Accessible forms: labels, instructions, and feedbackW3C Web Accessibility Initiative
- Excessive agency: permissions, functionality, and autonomyOWASP
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy