The Margin Is in the Search: How We Built an AI Sourcing Engine for a Car Reseller

TL;DR: In car resale, the profit is decided the moment you buy, not the moment you sell. Our client was manually browsing dozens of listing sites every day, hoping to spot an underpriced car before someone else did. We replaced the browsing with an engine: it scrapes thousands of listings, cross-checks each one against wholesale value, scores the margin, and puts only the genuinely promising cars in front of him. It took four weeks to build. He now averages roughly $15K profit every two weeks — and, in his words, it let him buy his dad his dream truck.

Most people think a car reseller makes money by selling well. They do not. They make money by buying well.

By the time a car is sitting in your driveway, your margin is already fixed. Every skill you have on the sales side — the photos, the listing copy, the negotiation — is just you trying to realise a number that was locked in at purchase. Which means the real job, the one that actually determines whether the business works, is finding the underpriced car in the first place.

And that job is miserable. It is refreshing the same twenty listing sites over and over, all day, forever, hoping to be the first human to notice that someone has priced a car two thousand under what it is worth. Miss the window by an hour and it is gone.

That was our client's life. He is good at this — genuinely good — but he was spending his best hours hunting instead of negotiating, closing, and scaling. So we asked a different question: what if the hunting simply happened on its own?

//The engine, in plain terms

The system does four things in sequence, continuously.

It scrapes. Thousands of listings across the high-volume marketplaces where inventory actually shows up. Not a sample. The whole surface, on a loop.

It filters. This is where most naive versions of this idea die. A raw feed of every listing is not an advantage — it is just the same haystack delivered faster. So an AI filtering layer reads each listing the way an experienced buyer would: what is this car really, how is it described, what is the seller not saying, does this look like a genuine opportunity or a lemon dressed as one.

It prices. Every surviving candidate gets cross-checked against wholesale benchmarks. This is the step that turns "cheap" into "profitable," and they are not the same thing. A cheap car with no spread is a waste of capital. The engine estimates the gross profit on the spot.

It surfaces. Only the deals that clear the bar reach a human. He opens his notifications and sees a short list of cars with margin context already attached, instead of an ocean of listings with none.

The output is not "here are some cars." The output is "here are the cars worth your attention, and here is roughly what each one is worth to you."

//The hard part was deciding what not to show him

The instinct when you build something like this is to be comprehensive. Show everything. Cast the widest net. It feels safer — surely more options is more opportunity?

It is the opposite. A firehose of listings is functionally identical to the manual browsing we were trying to eliminate. If the system hands back 400 cars a day, he still has to evaluate 400 cars a day, and we have automated nothing except the scrolling.

The value of this system lives almost entirely in its willingness to throw things away. Aggressive filtering is the product. Every listing the engine discards is time it hands back to him, and the whole thing only works because it discards nearly all of them.

That is the general lesson, and it applies far outside car resale: an automation that surfaces everything has not automated the decision, only the retrieval. The judgment is the job. If your system will not exercise judgment, it is a search bar with extra steps.

//Speed is a feature, not a nice-to-have

Undervalued inventory has a shelf life measured in hours. Sometimes less.

This is what makes sourcing such a natural fit for automation, and why the ROI here is unusually clean. In a lot of business processes, doing the work faster is merely pleasant — the same outcome, sooner. In sourcing, doing it faster changes the outcome, because the opportunity genuinely evaporates. The car gets bought by someone else. There is no version of the deal available to you at 6pm that was available at 9am.

A human who checks the marketplaces three times a day is not doing a slower version of this job. They are playing a different, worse game, and losing most of the deals before they ever see them.

//What it actually changed

The headline number is the one the client volunteered himself: roughly $15K profit every two weeks, consistently, since the system went live. We built it in four weeks, from first conversation to live monitoring.

But the number is not really the interesting part. Three things changed underneath it:

• Sourcing became repeatable instead of opportunistic. It stopped depending on him happening to look at the right screen at the right moment. A business that depends on someone's attention is not a business, it is a habit — and habits do not scale.
• His reaction time collapsed. Deals that used to require luck now require a decision.
• He knew which listings deserved immediate action. Confidence, not just volume. That is what the margin estimate buys you.

And the outcome he cared about most had nothing to do with any of our architecture. He bought his dad a truck. That is the actual point of this work, and it is worth remembering when you are deep in scraper configs at midnight.

"Thanos helped me set up automation for my automotive business and it completely changed the game for me. Since implementing it we have been averaging around $15K profit every two weeks. Because of that, I was able to buy my dad his dream truck. If anyone is looking to automate parts of their business, he really knows what he is doing."

//Does this transfer to your business?

Probably — if your business has the same shape. Ask yourself three questions:

1. 1Is there an opportunity stream you cannot fully watch? Listings, tenders, RFPs, job postings, inbound leads, supplier catalogues, auction lots. Anything that arrives faster than a person can review it.
2. 2Is there an objective way to score an item's value? In cars it is wholesale benchmarks. In your world it might be margin, fit, urgency, or historical close rate. If you cannot articulate what "good" means, an AI filter cannot either.
3. 3Does being early actually matter? If the opportunity is still there next week, automation saves you time. If it is gone in an hour, automation makes you money. These are very different business cases, and the second one is far easier to justify.

If you answered yes three times, you are not looking at a chatbot problem. You are looking at a sourcing engine, and it is one of the highest-leverage systems a small business can own.

If you answered no to the second question, start there. The scoring logic is the system. Everything else is plumbing.

//The honest caveats

This is not a machine that prints money, and we would be doing you a disservice to sell it that way.

The engine finds opportunities; it does not close them. Our client still has to buy well, inspect properly, price correctly, and sell. His results reflect his own expertise operating with better information — not a system operating without him. Margins vary by market, by season, and by how much competition is running similar plays. And a sourcing engine is only as good as its scoring logic: point it at a market where you cannot reliably estimate value, and it will confidently surface garbage.

What automation did here was remove the bottleneck between his judgment and the market. It did not replace the judgment. It never does.

---

Want to know whether your business has a sourcing problem hiding inside it? Read the full automotive case study or tell us what you are trying to find.