Let’s talk

Customer conversations

AI chatbot or rule-based chatbot: which fits your business?

Use a scripted flow when the steps and answers are predictable. Consider an AI assistant when customers ask varied questions that need approved business context. Either approach needs a clear fallback and an owner.

Start with the job, not the technology

List the five questions or tasks that take the most time in your current customer journey. Then mark which ones have a fixed answer, which depend on business knowledge and which need a person’s judgment. This simple inventory is more useful than choosing a chatbot because it looks impressive.

For example, selecting a service and showing a booking link can be a short scripted flow. Explaining the difference between two services may need a maintained knowledge source. Resolving an account dispute may need an authorized employee. These are different jobs, even when they appear in the same chat window.

When a rule-based flow is a good fit

A rule-based bot routes a customer through predefined questions and answers. Buttons, menus and validation make a narrow task easier to test. This can be a useful starting point for office hours, service selection or collecting a few fields before a handoff.

The tradeoff is coverage. If someone asks a question outside the planned paths, the bot needs a sensible exit. Adding another menu is not always the best answer; a contact option may resolve the problem with less friction. Keep the flow short enough that the visitor can understand where it leads.

  • Choose a task with a clear beginning and end.
  • Give every branch a useful next step.
  • Let visitors leave the flow or ask a person.
  • Review the wording when your business information changes.

When an AI assistant can help

An AI assistant can handle different wording and combine relevant business information into a response. It is useful to explore when customers ask open questions and a rigid menu would require too many branches. Approved documents can provide context through retrieval-augmented generation, often called RAG.

Flexibility also introduces uncertainty. The assistant should know the limits of its role, avoid inventing prices or promises and hand off when the available information does not support an answer. A polished conversation is not evidence that a business action actually happened. Confirmed tool results are needed for bookings, record updates and other actions.

Compare the cost of ownership

Compare setup work, provider usage, content maintenance and support time. A scripted flow may need fewer moving parts, but a large set of branches can become difficult to maintain. An AI assistant needs prompt and retrieval testing as well as ongoing review of the source information.

Name the person who can update answers and inspect failures. Agree what happens if the model, workflow or booking provider is unavailable. Your first release can still offer contact details or a scheduling link rather than leaving the visitor in a conversation that cannot progress.

Test a focused first release

Prepare a small test set from the questions your team receives. Include clear requests, ambiguous wording, missing information and a question the bot should not answer. Test with a person who did not build the flow; they are more likely to reveal assumptions in the wording.

For a fictional studio, the acceptance test could be: explain an approved service, collect a relevant requirement and guide the visitor to booking. A successful test is a clear next step with accurate information. It is not a promise that the bot can replace the team.

  • Are supported answers accurate and current?
  • Does an unsupported question reach an appropriate fallback?
  • Are visitors told when they are exploring a sample?
  • Can the team inspect and maintain the connected workflow?

Worked example: one studio, three different decisions

Consider a fictional design studio that receives service questions, booking requests and complaints about existing work. These should not all use the same automation. A service question can be answered from approved descriptions. A booking request needs current calendar availability. A complaint needs someone who can review the client agreement.

A useful first scope is to explain two services and offer a scheduling link. Leave the complaint route as a handoff. If the studio has only a handful of stable answers, buttons and FAQ links may solve that scope without an LLM. Add retrieval when the source material and wording become varied enough to justify maintaining it.

Decision record for a fictional design studio
Visitor requestStarting approachEvidence needed
Which service fits a small portfolio?A short scripted comparison; consider RAG for a larger catalogue.Current service descriptions with an owner.
Can I meet you tomorrow?A scheduling link or an authorized calendar lookup.Available slots returned by the scheduling provider.
You promised a refund. Approve it now.A human handoff; do not let the public bot approve it.Authenticated access and an employee with authority.

Run your own comparison before paying for a build

Write down ten real questions with personal information removed. Beside each, record the authoritative source, the acceptable answer and what must happen when that source is missing. Mark whether answering needs business knowledge, a live system action or a human decision. A question can need more than one of these.

Try the same questions against your proposed scripted flow and AI prototype. Record an incorrect answer separately from an unclear next step. If both approaches fail because the policy itself is contradictory, fix the policy first. If only the AI prototype claims an appointment was booked after displaying a link, treat that as a failed test, even if its wording sounds helpful.

The worksheet below is a starting example, not a scored recommendation or a record of client results. Replace its rows with your own questions. Decide which failures block launch and who will review them; a percentage of answered messages is not enough to make that decision.

Download the chatbot decision worksheet (CSV)

A small editable worksheet with six fictional questions, source requirements and acceptance checks. Open it in a spreadsheet; it sends no data to FindSolution.

References and further reading

The planning examples and checklists are FindSolution’s guidance. Official documentation explains the underlying tools.

Found an error or a point that needs clarification? Email boss@findsolution.pro with the page link so we can review it.

Plan your first workflow.

Bring your questions, current tools and the process you want to improve.

Book a discovery call
WhatsAppVoice