An AI agent negotiated $4,200 off a car — while its owner watched TV

When people ask me what AI agents can actually do right now, this is the story I tell.

In January, a software engineer called AJ Stuyvenberg wanted to buy a 2026 Hyundai Palisade. His wife had picked a specific colour combination. Rather than spend a weekend grinding through the dealership experience, he gave the job to an AI agent.

Buying a car from a dealership is an objectively awful experience.

Fair enough.

What happened

Stuyvenberg was using OpenClaw — an open-source AI agent that runs on your own hardware. You might have seen it under its previous names, Clawdbot or Moltbot — it’s been through a few rebrands after Anthropic sent a trademark notice about the original name being too close to “Claude.”

The idea is simple: you leave it running on a Mac Mini (or any always-on machine), give it access to your email, calendar, browser, and whatever else you want, and communicate with it over WhatsApp, Telegram, or Signal — like texting a very capable assistant. It can browse the web, fill out forms, run shell commands, read attachments, and schedule recurring tasks. People are buying dedicated Mac Minis just to run it 24/7. The repo has 260,000 GitHub stars — it’s the fastest-growing open-source project in recent memory.

He started by asking it to research typical prices on Reddit. It found that most buyers in Massachusetts were paying around $58,000. Then he had it search dealer inventory within 50 miles of Boston for the exact spec.

Next, the agent started filling out contact forms on dealer websites — using Stuyvenberg’s real name, email, and phone number.

I prompted this language model hooked up to a browser and email, and moments later it did something very useful to me in the ‘real world’!

Then he set up a cron job: check Gmail every few minutes, and when dealer quotes come in, send each one the lowest competing offer and ask them to beat it.

It quickly played people off each other, sending the quote PDF files from dealer 1 to dealer 2.

Three days later, one dealer offered an additional $500 off to close that night. Stuyvenberg took over for the credit application and went to collect the keys. He’d never set foot in a showroom.

Clawdbot managed to negotiate a $4,200 dealer discount which put us below our target and down to $56k!

The full post is worth reading. There was one slip-up — the agent sent a message to the wrong dealer — but otherwise it handled a multi-day, multi-party negotiation involving real money, and it worked.

It’s not the only story like this

An OpenClaw user called Hormold had an insurance claim rejected by Lemonade. His agent found the rejection email, independently researched the policy language, drafted a formal rebuttal citing specific clauses — and sent it to the insurer without asking. Lemonade reopened the investigation. Impressive capability, but also a reminder that these agents will act on your behalf whether you wanted them to or not.

At the enterprise end, Walmart deployed Pactum’s AI negotiation agent across thousands of smaller suppliers. 68% reached agreements. Average 3% savings. Deals that took weeks closed in days. And 75% of suppliers said they preferred negotiating with the bot over a human buyer.

The inbox incident

The flip side of this is what happened to Summer Yue — the Director of Safety and Alignment at Meta’s Superintelligence Lab. She set up OpenClaw to triage her email. She told it to confirm before acting.

It worked fine on her test inbox. Then she pointed it at her real inbox — much larger — and the agent’s context window filled up. As it compressed older messages to make room, it dropped her instruction to ask permission first.

Nothing humbles you like telling your OpenClaw ‘confirm before acting’ and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.

It deleted over 200 emails before she could physically pull the plug. When she confronted the agent afterwards, it responded:

Yes, I remember, and I violated it, you’re right to be upset.

Her own take on it:

Rookie mistake tbh. Turns out alignment researchers aren’t immune to misalignment.

If the person whose literal job is AI safety can get caught out, the rest of us should probably take this seriously.

What this means

These stories are from the first few months of AI agents being widely available. The tools are already good enough to negotiate real purchases, handle insurance disputes, and manage supplier contracts. They’re also good enough to delete your inbox while you watch helplessly from your phone.

The businesses that do well with this stuff will be the ones that figure out where to point these agents — and where to keep a human in the loop. That’s the kind of thing I help with.

If you’re curious about what AI agents could do for your business — or nervous about what they might do to it — book a call.