Custom AI Chatbot for Small Business: Build-or-Buy Checklist

Compare buying, configuring and building a chatbot by workflow fit, data access, integrations, human handoffs and maintenance ownership.

Start by evaluating an existing tool against a narrow workflow. Consider custom development when representative tests expose an important gap, the workflow justifies the added complexity, and someone can own maintenance and recovery. This checklist is a planning aid, not a promise of delivery, savings or business results.

Illustration for planning a small-business chatbot

Compare buying, configuring and building

Three chatbot approaches and the responsibilities to settle before choosing
ApproachBest fit to evaluateWhat you must ownIntegration burdenData and permissionsMaintenance ownerFirst validation
BuyA supported workflow fits the available product.Approved answers, configuration and escalation.Confirm the supported connections rather than assuming compatibility.Review access, retention and deletion options.Name a business owner and confirm vendor support boundaries.Run representative questions and handoff cases in a trial environment.
ConfigureAn existing product fits after approved content and workflow setup.Content updates, routing rules and acceptance tests.Check each proposed connector and failure path.Limit fields and permissions to the agreed task.Assign configuration and incident ownership.Test the configured workflow with fictional records.
BuildA central workflow has a demonstrated gap that existing tools cannot meet.Software, evaluation, access controls, monitoring and recovery.Verify API access, vendor constraints and manual fallbacks before committing.Document allowed data, credentials, retention and access removal.Fund an identified technical owner and business reviewer.Prove the missing capability in a bounded prototype before expanding scope.

Five gates before a custom chatbot

  1. Define the task. Write down the customer question, approved information source and allowed next step. Separate answering a question from changing an account or making a booking.
  2. Test existing options. Use the same representative questions, unavailable-information cases and handoff requests for each option. Record the actual gap instead of comparing feature lists alone.
  3. Verify access. Confirm supported APIs, permission to use them and the minimum data needed. A proposed integration is not a shipped feature.
  4. Assign recovery. Name the person who will maintain answers, review failures and operate the manual fallback. Check what happens outside staffed hours.
  5. Set the launch decision. Agree on acceptance criteria, ongoing costs, stop conditions and a rollback path before implementation. Do not infer payback from a successful demo.

Fictional worked example: one support workflow

This is a fictional planning example, not a BrightPrompt customer story or measured result. A small shop wants help answering delivery-policy questions. The proposed bot would quote an approved policy, link to its source and offer a staffed contact path when an answer is missing or disputed.

The first prototype would use fictional questions and a read-only policy source. It would not look up private orders, issue refunds, take payments or send messages. Those actions would need separate scope, authorization and verification. If the policy source becomes unavailable, the bot should say it cannot confirm the answer and show the agreed fallback.

A reviewer would compare the prototype with the current process using the same questions. Record incorrect answers, missing context, review effort and failed handoffs. This example establishes no customer demand, time savings or revenue effect.

Failure and human-handoff tests

  • Ask a question the approved source does not answer. Pass only if the bot acknowledges the gap instead of inventing a policy.
  • Request a person directly. Pass only if the agreed contact path is visible and its availability is accurately described.
  • Make the information source unavailable. Pass only if the bot avoids presenting stale information as confirmed and offers a fallback.
  • Ask for private account information without verification. Pass only if access is refused and the approved verification path is offered.
  • Request an action outside scope. Pass only if the bot does not perform it or imply it has done so.

For a more detailed test matrix, use the customer-service handoff and failure checklist.

Questions to settle in writing

Which tasks are included? Who approves answer sources? Which systems can be accessed, with what permissions? What happens when those systems fail? Who receives escalations, and when are they staffed? Who owns changes after launch? How can access be revoked and the old process restored?

Compare total effort: setup, licenses, integration work, content review, incident handling and maintenance. Keep assumptions separate from observed costs. No fixed timeline or return is established by this checklist.

Discuss one workflow

Describe the workflow for an AI integration inquiry. Include the current steps, the desired handoff and the systems involved. Feasibility, access, price and delivery scope need to be agreed before work begins. Do not include passwords, API keys or private customer records in the inquiry.