DriveAgent

The context

A dealership gets the same question fifty times a week: whether the car in the listing is still available, what the monthly payment is, whether they take a trade-in. A salesperson answers them between two walk-ins, late and from memory, and half of Saturday's inquiries get answered on Monday, when whoever asked has already bought somewhere else. The problem isn't that they're short of people: it's that the first answer needs nobody, and the fifth one does.

What the system does

• The main screen is the conversation, not the inventory: it comes in over WhatsApp or the website and you see what was asked, what the agent answered and which record it answered from
• The agent answers off real inventory and not off what it remembers: if the car sold yesterday, it says so, which is the one answer that can't be improvised
• It books the test drive against the salespeople's calendars, so the slot it offers exists. An agent that promises times that later aren't there is worse than not answering
• It takes the trade-in details for the appraisal —model, year, mileage, photos— and leaves it set up for a person to put the number on. The part that can be automated is the form, not the appraisal
• When it doesn't know, it hands off to a human and says why: a specific financing question, a complaint, somebody angry. An agent that answers anything rather than hand off is the one that gets switched off within a week
• The dashboard doesn't count how many it handled: it counts how many it resolved, how many it handed off and how many went unanswered. The three together or the number means nothing
• And every answer shows where the data came from. An agent over a catalog that can't cite the listing is a chat that sounds right, and sounding right is exactly the problem