A customer fills in the contact form on a plumber's website: the water heater is leaking in the basement, can someone come this week? Twenty-one seconds later in the video, they have a booking confirmation for Thursday at one o'clock, and the owner has done exactly one thing: tapped approve.
The business in the demo, Maple Street Plumbing, is invented, and so is its customer. The workflow is not. It is the shape of the AI integrations we build, and the shape matters more than the AI.
Four steps, one decision
1. The request is screened before AI touches it. Anything that arrives from the public internet is untrusted: spam, nonsense, and occasionally text written to trick an automated system into doing something it should not. A plain, deterministic check runs first. Only a real-looking request moves on.
2. An AI agent does the research. It looks up whether this customer has been in before, checks the calendar for open slots, and reads the price list. Then it drafts a reply: here is what we think the job is, here is when we can come, here is what it is likely to cost. All of that is busywork a person would otherwise do while holding the phone.
3. The owner makes the call. The draft lands in the owner's inbox with two buttons: approve and send, or handle it myself. Nothing goes to the customer until a person has decided it should. The AI never sends anything on its own.
4. The customer gets a confirmation. Approve, and the booking goes on the calendar and the customer gets a clear message with the time and the estimate. Request to booking, no retyping.
Why the approval step is the point
It would be easy to remove step three. The system would be faster, and the demo would look more impressive. We would not build it that way, and we would advise you not to either.
A reply to a customer is a promise about price and time, made in your name. A model that reads your calendar and your price list is good at drafting that promise. It is not the one who has to keep it. Leaving the decision with the owner costs a few seconds per lead and buys something you cannot get back once it is gone: a customer's trust that what the business says, the business meant.
This is what we mean by practical AI. It does a specific, bounded job inside a workflow, with a person at the point where judgment is needed, and with every step logged so you can see what it did and why.
What it would take for your business
The pieces are ordinary: your contact form or inbox, your calendar, your price list, and a way to approve from your phone. The work is in wiring them together so the chain is reliable, the screening holds up, and the failure of any one piece raises an alert instead of silently dropping a customer.
What would you want handled like this? Tell us what comes in and what you wish happened next.