The One-Minute Test for Whether a Vendor's AI Is Real
The call on September 4th ran 107 minutes. The man on the other end runs an e-commerce company and was looking at replacing his inventory system sometime next year, and for the first stretch we talked about warehouses. Wave picking, put-away carts, picking one bin for many orders at once. He said he’d be tempted to build his own handheld picking app instead of buying one.
Then he mentioned that his company has run its purchasing through a portal he wrote himself since 2019. He enters the order in the portal and it pushes the purchase order into his inventory system through that system’s API. I said everything in our front end runs on our own API too, and that if you open Chrome’s console you can see the calls just from clicking around: “There’s literally nothing that isn’t an API.” He said the API on his current system was good, which was one of his two reasons for picking it. I said the API is always the first thing I look at.
Later he told me what it had cost him. That API is big, and when he wanted an agent working against it, the documentation was too large to hand over whole. So he condensed it into something an agent could read. If the vendor had published that version, he said, he would have used it.
He had an opinion about AI inside software, too. He named a project-management tool that keeps pushing its own assistant at you and, he said, is impossible to use from Claude Code. “It’s the worst pattern in applications and the ones that I think are going to die are the ones that do that.” I told him what I tell everybody about ours: “We have an embedded AI assistant. And I tell people that is like the minimum power that you’re going to get because it’s not running agent.” My programmer built it. It’s good at pointing you to the right documentation, and it can fix some data if you ask. I wouldn’t say vendors with an assistant and a thin API are doomed, though. Plenty of buyers will never run an agent, and for them the assistant is fine.
What I care about is the other direction, so I told him what I do. A supplier had emailed me a PDF that morning of products I’d never sold, with names, prices and a color here and there. I didn’t open a spreadsheet. I told Claude Code to build a product catalog in SKU.io from that email. It made the products, and it pulled out an attribute that wasn’t a column in the PDF: the titles followed a pattern, and it read the pattern as data. I’d had to iterate on the skill to get that. It built a saved view of everything it had just added. When I asked for a purchase order from the same products, on consignment, it worked out the payment term and created the PO. I go into the app afterward, to look.
On the pages of the app where one of those skills applies, there’s a small icon with a count on it in the top bar, and it lists them. I’d had the idea for about a year and built it a month before the call. He asked whether we were the only software that did that. I said I’d never seen it anywhere else.
The test I’d give a buyer came out of that conversation. Have the vendor open the browser’s developer console next to their product, go to the Network tab, and click through a purchase order. A call to a real endpoint on each click means an agent can do the same thing without the screen. A couple of calls to load the page and then nothing, while a chat panel does the work, means the AI lives inside the product. It can mislead in small ways, since a vendor can route a lot through one batched call, and it says nothing about what API access costs, which the buyer’s checklist covers. The piece on running my own business through an agent is the terminal side. I’ll open the console with you on a call, on any vendor, ours included: book a demo.