Bookings made and kept, without the phone tag.

Under the hood

All three channels are built to feed one booking queue. Each request is matched to a customer record by phone number or email address, so a returning client is recognised and someone who asks on two channels gets one booking, not two. The website books through the same rules as WhatsApp, so the result is the same whichever way a customer comes in.

Under the hood

The language model turns the message into a structured request: one booking type from the list you set up, a place and a time window. Each booking type carries its own length, buffer, travel allowance, deposit rule and approval rule, so the model only picks the type and never invents a duration. Times are worked out in your time zone and stored with it, so a visitor booking from abroad sees your local time with the zone named. A request that fits no type gets one clear question, not a guess.

Under the hood

The agent reads free and busy times from Google Calendar or Outlook, not the details of your other appointments. Your hours, the buffer each booking type needs and the travel allowance between areas are taken off before a slot counts as open. An offered slot is held for a short time you set, so two people chatting at once are not offered the same hour. The hold lapses on its own if the customer goes quiet.

Under the hood

Each booking type carries the rule you set for it: book straight away, wait for a deposit, or wait for your approval. The rule is checked by fixed code, not by the language model, so no wording from a customer can talk it past one. The agent sends your deposit details but never takes or handles money, and the slot stays provisional until you mark the deposit as received. In the first period every booking waits for you, whatever its rule says.

Under the hood

Just before writing, the calendar is checked again, because it may have changed since the offer. Only then is the event created, carrying a booking reference. That reference makes the write idempotent: if the connection drops and the step is retried, it finds the event it already made instead of adding a second one. The agent only moves or cancels events it created. Your own entries are read as busy and never touched.

Under the hood

Reminders go out at the times you set before each booking. On WhatsApp they use message templates approved in advance, because that is the only way a business may start a WhatsApp conversation. A reply such as "can we make it Thursday" is read by the language model and answered with open slots under the same rules. A no-show is recorded, and you can set a rule that the next booking from that customer needs a deposit.

Under the hood

A cancellation frees the slot and works down the waiting list in the order you keep it. Each person is offered the slot with a short window to accept, one at a time, so nobody is told a slot is theirs after it has gone. The person who accepts is booked through the same rule check and calendar write as any other booking. Each offer and reply is kept with the booking, so you can see who was asked and when.