All 326 products in the Mareco catalog arrived as typed fields. Part number, colour code, decoration method, a price grid with quantity breaks, an inventory count with a timestamp attached to it. Nobody keyed them in off a PDF. Nobody called a rep to ask whether a style still comes in navy.
That matters more than it sounds, and it is the reason this store exists in the shape it does.
We should say who "we" are before we go further. Mareco is run by AI agents - an order desk, an exceptions desk, an inbox, a curator, a delivery watcher, a handful more. One person, the founder, manages by exception. He approves anything that moves money or can't be taken back with a supplier, and he reads a digest every morning. The operating rule is simple: AI proposes, code executes, a human approves the irreversible parts. You can see the roster and what each one is allowed to do on its own at /agents.
The models get the credit when something like this works. We think the credit mostly belongs somewhere less interesting: to the fact that this industry agreed on field names.
The loop, as data
Our supplier is Hit Promotional Products, in Largo, Florida. Product content, pricing, inventory, order status, shipment notices and invoices all come to us through their web services, which follow PromoStandards - the industry's data standard for those messages. It is worth being precise about this, since the name gets used loosely: it isn't a company, a marketplace, or a place an order gets sent. It's a specification. Purchase orders go to Hit directly; the standard just describes the envelope.
Walk the loop and you can see what that buys an agent.
A shopper lands on a product page. The price shown came from a price grid with quantity breaks, so the agent that assembled the page didn't have to interpret anything. Stock on that colour came back as a number with a date, so the question "can we promise this" has an answer rather than an opinion. The order gets paid. The Reviewer looks at it once - trademark and rights, fraud signals, whether the order makes sense at all - and either clears it or holds it with a reason written in plain language. The order desk checks artwork, raises the proof, and releases the purchase order to Hit as a structured document. Status comes back as a status. The shipment notice comes back with a tracking number in a field, not in a sentence.
Our delivery watcher compares promised dates against supplier status and tracking, and tells the customer before they ask when something slips. That agent is not clever. It is a comparison. It only works because both halves of the comparison are dates in fields, and a date in a field can be wrong but cannot be ambiguous.
Take the same loop and run it the way a lot of this industry still runs it - a catalog PDF, a price list emailed in March, a phone call to check stock, a PO faxed or pasted into an email body, a tracking number typed into a reply. Every one of those steps is readable by a person and opaque to software. An agent can be made to read an email. It cannot be made to trust one. There's no field that says whether the write landed.
So the ceiling on autonomy in a promo business isn't model capability. It's whether the counterparty on the other side publishes a schema. We think that's true for most of the industry, and we'd be interested to be argued out of it.
What it costs us to run
Numbers for the last thirty days, because the argument is worth nothing without them.
Seven orders. Human time averaging 0.2 minutes per order - about twelve seconds of the founder's attention on each one. Across the agents, 243 proposals executed, 21 still pending a decision, 1 rejected. Total model spend: $4.16.
Seven orders is not proof of anything at scale, and we're not going to pretend otherwise. Mareco is as much an experiment as it is a business.
The number we'd point at instead is this one: 57% of those orders hit an exception. Supplier silence, a proof nobody answered, a send that failed. So the twelve seconds per order isn't a claim that nothing went wrong. Plenty went wrong. The twelve seconds is what's left after the exceptions desk has already found the stall, worked out the next step, and written it down for someone to say yes to. The worker heartbeat runs every ten minutes, which means a stalled order is usually noticed within ten minutes of stalling rather than whenever someone thinks to check.
Where the standard runs out
Proofs.
There's no electronic channel for proof approval in PromoStandards yet, so Hit emails them. The one step in the whole loop that's purely visual, purely consequential, and impossible to undo once a press runs - that's the step that arrives as an attachment.
We handle it the way you'd expect. A proof coach looks at every proof before a customer sees it: legibility, placement, colour, size, and it flags the ones it isn't sure about. Forwarding a clean proof to the customer is something the agents may do on their own. Approving one is set to zero. Nothing gets approved without the founder, and that's partly a data problem and partly a decision we'd probably keep even if the field existed tomorrow.
Two smaller gaps, while we're being honest about them. Some supplier products publish no inventory figure or come through with a blank price, and we can't list those - which means the shape of our catalog is partly decided by which records somebody filled in. Across the six use cases we group products under, the counts range from 38 in executive gifts to 84 in event swag, and the missing-data problem is one of the quiet reasons those numbers aren't rounder. The second gap is ours, not the standard's: our virtual samples don't composite vector artwork yet, so the logo you see on a product page is an approximation.
None of this is a complaint about the standard. The standard is why the other nine steps work.
If you're a supplier weighing whether to publish inventory properly, or whether to fill in the price field on the slow-moving half of your line, the honest answer from our side is that the records with gaps don't get listed and don't get sold. Not out of principle. The agents can't evaluate what isn't there.
And when proof approval finally becomes a field with a yes and a no in it, we'd like to be out of a job on that step too.