A customer asks whether a custom cake can be ready on Saturday. The answer depends on the design, size, budget, pickup time, advance payment, and work already booked.
An AI tool can write a fluent reply before it knows any of those things. That is exactly why fluency cannot be the decision.
Start with the work before the words
A safe reply begins with the facts the business has approved: opening hours, products, service area, payment rules, lead times, and the questions staff must ask.
It also needs the current enquiry. Who is asking? What do they want? What is missing? Has a person already replied? What should happen next?
If the system cannot find those facts, it should not fill the gap with a confident guess. It should show what is missing.
The smallest useful AI flow
Kaiku Labs uses a simple pattern in its internal enquiry-intake demo. The demo uses fictional data. It is not a client result.
First, an enquiry is placed into one clear view. It may be pasted from WhatsApp or Instagram, or entered after a phone call.
The system extracts agreed details, classifies the enquiry, and prepares a draft from approved business information.
A staff member then edits, approves, copies, or rejects the draft. Nothing reaches the customer until that person decides.
Why approval is part of the product
A person knows when a regular customer needs a warmer answer. A person can notice that a promised date is no longer possible. A person carries responsibility for price, tone, exceptions, and promises.
Approval is not a temporary weakness that disappears when the AI becomes better. It is the point where business judgement enters the flow.
The system should make that decision easier. It should show the original message, extracted details, missing information, draft, source used, and previous history together.
What should be automatic
Routine preparation is a good fit: sorting enquiries, finding approved information, showing missing details, drafting a reply, and moving approved work to its next status.
High-risk decisions are not: inventing a price, promising availability, approving a refund, diagnosing a patient, committing stock, or sending a sensitive answer without review.
The line depends on the business. It should be written down before the first live message enters the system.
A chatbot is not the first requirement
If staff cannot see every enquiry in one place, know its status, or tell who owns the next reply, automatic writing will make a messy process move faster.
Fix the queue first. Give each enquiry a clear state, owner, and history. Then add drafting where repeated writing is genuinely slow.
Sometimes a saved reply template is enough. Sometimes an existing inbox tool already fits. AI is worth adding only when it removes real preparation without removing control.
A test before building
Collect ten real questions customers asked this week. Remove names and private details. Mark which answers were fixed, which needed current business data, and which needed judgement.
Fixed answers may belong on the website. Repeated answers may become approved templates. Data-dependent answers need a connected view. Judgement calls stay with a person.
That map tells you whether you need better information, a clearer inbox, an AI draft step, or no new system at all.
See the AI workflow starting point → Check whether an existing app fits →