102 raw inventory photos → 22 live eBay listings. Now I’m trying to use less AI

I ran a second batch through the inventory system I’ve been building and it changed what I think the hard part of this actually is. For anyone who didn’t see the first post, Batch 1 was 40+ raw photos covering 11 random resale items. The goal wasn’t just to have AI describe pictures. I wanted to see how far I could get toward dropping real inventory photos into a system and getting usable eBay listings out the other side. All 11 items were identified, researched, turned into product records and pushed into Seller Hub. That first run exposed a bunch of problems that don’t really show up in demos. Shipping dimensions needed to be treated differently from product dimensions. Some categories forced different return policies. Marketplace fields could change after submission. One weird item could hold up everything behind it. I also realized pretty quickly that having the model say it was confident meant almost nothing by itself. Batch 2 was 102 photos covering another 22 items. All 22 were prepared and all 22 eventually made it through publishing. 17 were specifically identified and 5 were intentionally left more general because there wasn’t enough evidence to justify an exact identification. Only 4 needed outside research beyond what the system already had available. None required me to tell it what the product was. Getting all 22 live was useful because it gave me a much better picture of where this actually breaks. One listing had its price changed during the marketplace interaction. Another needed a category specific safety disclaimer. Shipping kicked an item into manual review. Marketplace behavior itself became something I had to verify rather than assuming the data I submitted was the data that ended up live. The more interesting failures were in the reasoning though. I’ve been deliberately testing whether the system can arrive at a correct answer for the wrong reason, because I think that’s one of the more dangerous failure modes in this kind of automation. One item had two identical physical units. The system ultimately produced the correct quantity of 2, but the evidence path it used to justify that count wasn’t actually sufficient. I counted that as a failure. That distinction has become pretty important to me. If the answer is right because the model guessed correctly, I don’t want that passing validation. Otherwise I’m basically training the system to get away with unsupported inference until the day the guess is wrong. I added a regression specifically for that case so a correct output reached through unsupported reasoning still fails validation. Across the 31 real items from these two batches, I currently have one known material escaped error created by the system itself, 1/31 or 3.23%. If I include the marketplace induced price mutation, it’s 2/31, or 6.45%. Small sample obviously, but I’d rather start measuring this now than just say the system “works.” That’s probably been my biggest finding so far. Getting an AI to identify an item correctly once really isn’t the hard part anymore. Getting it to do it for the right reasons, know when the evidence isn’t sufficient, survive marketplace weirdness and then repeat the process reliably at scale is much harder. The model also really likes to cooperate with whatever premise you give it. If you suggest an identification, it can be very good at finding evidence that supports you instead of actually trying to prove you wrong. Multiple passes help, but two passes can agree on the same bad assumption. So I’m moving more of the authority away from the model itself. If information can’t actually be supported, it stays blank or gets kicked back instead of guessed. Confidence can help prioritize review, but it doesn’t automatically grant permission to act. Things like reversibility, financial exposure, customer impact and account risk determine how much authority the system gets. I’m also starting to think using a general purpose model and browser for every step is the wrong way to scale this. The next version is moving the boring work out of the agent. Local barcode/OCR, exact record matching, cached research and rules, deterministic templates and state handling should deal with ordinary inventory cheaply. Models should get called when there’s actually ambiguity or reasoning required, and marketplace APIs should handle execution where possible instead of having an AI visually operate a website over and over. I’ve already started building that side. The current cached regression covers 33 items and 149 photos and is passing with 149 cache hits and zero false matches. More importantly, the regression now includes known failure cases instead of only testing whether previously successful items still work. The eBay API side is started too. I’ve deliberately parked it at machine local OAuth and read only GetItem access while I wait for eBay Developer approval. I’m using one existing J&J listing as the test target and nothing in the API work is allowed to modify the live listing yet. So I think my question has changed since Batch 1. It was basically “can I give AI a pile of random physical inventory and have it figure out what everything is and list it?” After 33 real items, 149 photos and getting the entire second batch published, I’m pretty convinced the answer is yes. Now I’m more interested in how little AI I actually need. How much of this can become deterministic, cached and cheap? When does the model actually need to reason? What evidence should give it authority to act? What failures should automatically kick something back to a human? And can I get the cost and error rate low enough that the same architecture makes sense across hundreds or thousands of items? That seems like a much more interesting problem than getting another demo to work once.