Conversations that
finish the job.

A chatbot that answers questions is a search box with worse ergonomics. What earns its place is an agent that completes the task: takes the booking, captures the deposit, checks availability against the system that actually knows, and writes the result back. That requires tool access to real systems, which is where most chat projects stop and where this one starts.

Unattended operation is the requirement, and it is a demanding one. It means deciding in advance what happens when the customer is ambiguous, when the request is outside scope, and when the model is not confident — so handoff to a human is designed as a normal path from the beginning, not added after the first bad transcript is screenshotted to a group chat.

WhatsApp specifically has rules worth knowing before you commit: business-initiated messages outside a service window use approved templates, and messaging is metered. Those constraints shape the conversation design, so we design within them rather than discovering them at launch.

This is a system we run ourselves. Mesa handles restaurant reservations on WhatsApp unattended, around the clock, including deposit capture — so the failure modes described here are ones we have met rather than anticipated.

What you get

• Conversation design covering the task, not just replies
• Tool access to the systems that hold real answers
• Confidence thresholds with defined human handoff
• Channel setup, templates and rate limiting
• Transcript logging and per-conversation cost monitoring

Not included

• Model inference costs, billed at provider rate
• Training bespoke foundation models