Poppy (Personal Agent Protocol): what happens when the customer’s AI shows up to buy the car

[dealerdata1]

Wanted to put this on people’s radar because I haven’t seen anyone in auto talking about it yet, but I sometimes live under a rock.


Back in October, Sierra and Meta released something called the Personal Agent Protocol, “Poppy” for short. It’s an open standard for how a customer’s personal AI agent works with a business. Think Meta’s Muse, ChatGPT, or whatever ends up on their phone next. Version 1.0 of the spec is out now, and the partner list is not small with both CarMax and Cox Automotive on it.


The plain version of how it works:

• The business puts one file on its website (/.well-known/poppy.json) that tells agents what they’re allowed to do there and how.
• The agent has to say it’s an agent and who it’s working for. No more bots pretending to be people on your lead forms.
• If a task needs the customer’s account or info, the customer signs in on your page and approves exactly what the agent gets.
• The business decides how agents get work done: browse the site, use an API, or talk to the business’s own agent.

That last one is what matters for us.

Right now, if a customer tells their AI “find me a white Tacoma under $45k out the door within 50 miles,” the agent scrapes VDPs and fills out forms. Your BDC gets a pile of leads that look like people who never answer the phone. Poppy gives that agent a front door. If the store has its own agent behind that door, the two can actually work the deal.

I run Dealer Data One, and we’ve built the Poppy 1.0 spec into the agent-to-agent system we’re working on. So I’ve spent more time in this spec than is probably healthy. Here’s what a setup like that should be able to do when it’s done right:

• Answer inventory questions from live inventory, not a stale feed
• Work price inside rules the managers set: what can be quoted, what needs a manager, which incentives can be mentioned
• Collect trade info and set expectations without pretending to appraise a car through a chat window
• Book service appointments
• Hand the customer to a person when it’s time, with everything already agreed in front of the salesperson
• Push the lead into your CRM flagged as agent-sourced, so it doesn’t land in the same follow-up cadence as a form fill
• Be clear about who’s talking. The spec has both sides identify whether a person or an AI is speaking.

Where it gets interesting is the rest of the store. Poppy isn’t only for chat. A dealer’s poppy.json can point agents to APIs, and the same customer session carries across all of them. So in theory:

• Service scheduling in the DMS becomes something a customer’s agent can book directly
• Digital retailing tools could share real deal structure, so the agent gets an actual payment worksheet instead of a “starting at” number
• Credit apps could run through the customer’s own sign-in and consent instead of an agent typing their info into a form
• Deposits or reservations could use the checkout extension that already exists for the spec
• There’s already an extension for approving exact terms before an action happens, which is what you’d want before anything gets signed or charged

I don't know of any store set up for this yet but would be curious to see if anyone has an example.

Curious how other folks would handle it. If a customer’s AI asked your store for an out-the-door number at 11pm on a Tuesday, what would you want it allowed to say?

 

[AdamRafael]

Great flag, Jake — this is the part most of us aren't ready for yet. Cox's AI in Auto Retail Tracker had 63% of buyers saying they'll shop with AI on their next purchase, and only 29% of dealers adjusting for AI search. So the agent showing up at 11pm on a Tuesday isn't a future scenario, it's the next version of the lead form.

If a customer's AI asked my store for an out-the-door number, I'd want it allowed to do three things, and only three, without a human in the loop: confirm that exact VIN is still in stock, quote the real out-the-door price for that VIN from live pricing — including taxes and fees, no "call for price" — and book a firm appointment or hold with a deposit, with the terms shown in writing.

What I would NOT let it do on its own: negotiate past a manager-set floor, appraise a trade, or promise a payment. Those need a person, and the agent should say so and hand over the full transcript so my team isn't starting cold.

The identification piece matters too. If I can't tell it's an agent, who it works for, and what the customer actually approved, I'm not comfortable letting it act. If I can, and the price it quotes is the same price my desk would quote in the morning, that's a fairer deal for the customer and a cleaner lead for us.