The shortest version of what we do would be "AI lists your parts for you". It is a better line than the one we use. We do not use it because it is not true, and because the difference is the entire product.
What the agent is good at
Writing a used part listing is mostly assembly. The donor vehicle gives you the fitment. The part gives you the category and the search terms. Your own history gives you the price band and the phrasing you use for grades. None of that is creative work, and all of it takes twenty minutes a listing when a person does it by hand for each channel.
That is a good use of a language model: read the structured facts, produce the prose, do it consistently, do it in seconds.
What it is not good at
Being sure. A model asked for a part number will produce a part number, and it produces one with the same confidence whether it read it off the casting or inferred it from something that looked similar. In a business where the wrong interchange number means a return, a negative, and a part that comes back scratched, "confidently approximate" is the exact failure mode you cannot tolerate.
You cannot prompt your way out of that. You can only decide where the human sits.
Where we put the human
At the end, before anything is public. The agent produces a complete draft and flags what it was unsure about instead of smoothing over it. You read it, change what is wrong, and approve. The correction is remembered for your yard, so the same mistake gets made once.
This costs a few seconds per listing. In exchange, nothing you did not read has your yard's name on it. We think that is the right trade, and we would rather say so plainly than describe it away in a shorter headline.