Rules-based chatbot
Fits predictable questions and tightly controlled responses.
Approved static answersAI chatbot development · proposed service
AI Chatbot Development for Useful Answers and Clear Human Handoffs
Help customers find approved information without making unsupported promises on your behalf.
MTI Tech's proposed customer-assistant service starts with the questions your team can answer reliably, the sources it approves and the situations that require a person. The right solution may be a rules-based assistant, an AI-assisted chatbot or a combination designed for a specific task.
M
Support assistantAutomated assistant · not a person
Do you build websites?
Yes. Web development is one of the published services.Source: Services page
What would my store cost?
I can't quote a price. I can pass your question to the team.
↪ Handed to a person
Choose the right kind of assistant
Match the design to the need, the risk and the budget.
Your questions are mostly…
AI-assisted chatbot
May suit varied wording and a larger approved knowledge base.
Approved sources, flexible wordingAction-taking agent
A separate design with additional permissions and approval requirements.
Permissions and approvalsThere is no requirement to add an LLM where approved static answers already do the job.
Start with a focused support scope
Decide what it answers, and what it must not.
Useful starting questions may concern published services, product guidance, onboarding steps or a documented support process. Define what the assistant must not answer and what it should do when the available material is incomplete.
Good starting scope
Published servicesProduct guidanceOnboarding stepsDocumented support process
Never inventA priceA policyA delivery promiseAn account status
An assistant should not invent a price, policy, delivery promise or account status to keep a conversation moving.
Build the human handoff into the experience
Make escalation visible.
Specify when the assistant should stop, what information it may collect and how the issue reaches the responsible team.
01The assistant stops
At an agreed trigger, not when it runs out of words.
02Collects what it may
Only the information agreed for the handoff.
03Reaches the right team
The issue arrives with the responsible people.
SoftwarePerson
The customer should understand whether they are speaking with software or a person.
Where account-specific information is involved, authentication and permissions must be designed separately from a public FAQ experience.
Test more than fluent answers
Sounding natural is not the test.
Evaluate whether the response is useful, supported and appropriately escalated, not simply whether it sounds natural.
One fluent but unsupported answer
Sounds natural
Supported by a source
A launch decision should rely on a documented test set and review process. Changes to the source content should trigger relevant retesting. Test examples are illustrative.
What a pilot should include
Agree the ground rules first.
- Channel
- Approved sources
- Question types
- Escalation path
- Evaluation method
- Conversation-data handling
- Cost controls
- Team responsible for updates
Separate scope items, not implied extras
Voice callingOutbound messagingAccount-changing actions
Each needs its own agreement before it becomes part of the build.
Chatbot questions
Before you build an assistant.
Not automatically. This is a proposed client service. MTI Tech's own approved-data, rules-based website chatbot remains a separate implementation and does not need an LLM or an external AI API.
No. It needs an explicit boundary and a useful fallback when evidence is missing.
Only under an agreed architecture with appropriate authentication, access controls and data-handling requirements.
Track supported answer quality, successful handoff, resolution where measurable, inappropriate answers and cost. A high conversation count alone is not success.
Start with your real support questions
Bring the questions your team already answers.
Bring an authorized sample of recurring questions and the information your team uses to answer them. We can compare the appropriate assistant options before committing to a build.