A Marketing Agency Gave an AI Agent Access to Its Own Payment Stack. It Found a Bug Nobody Saw.

The gap between an AI agent demo and an AI agent doing production work is enormous, and almost nobody publishes the second thing.

So this is worth looking at. On September 9, a Nevada marketing agency pointed Claude at its own live WooCommerce store — not a sandbox, not a client’s site — and asked it to handle sales tax configuration and get payments working. They logged the whole thing and published it.

The interesting part isn’t that it worked. It’s what kind of work it turned out to be good at, and where it stopped.

It Started With Research, Not Configuration

The agency operates across three states. Before touching a setting, the agent went and read tax law in all three — and came back with a structural observation rather than a list of rates.

Nevada, California, and Arizona all exempt professional services. But they get there differently, and the difference has teeth.

Nevada exempts services not tied to tangible property, but taxes fabrication labor — work that produces a physical good. California’s Regulation 1540 exempts agency fees only when they’re separately stated on the invoice, making them nontaxable outright. Bill them in a lump sum alongside taxable work and you fall into a rebuttable presumption splitting the combined charge 75 percent nontaxable and 25 percent taxable — which you then have to document your way out of.

Arizona was the sharp one. Its Transaction Privilege Tax doesn’t list advertising or consulting, and Proposition 126 froze new service taxes at their 2017 scope. But Arizona taxes activities, not objects — and Job Printing and Publishing are named, taxable classifications. An agency that starts fulfilling print orders can land inside a taxable classification without ever deciding to become a printer.

One rule underneath all three: pure service, exempt. Something tangible changes hands, taxable.

Then It Read the Database Instead of the Dashboard

This is the part engineers will recognize as the actual value.

The agent pulled the store’s stored configuration directly rather than trusting the admin UI — which, it turned out, was displaying an expired session and would have misled a human working from the screen.

What it found: the store’s base address was completely empty. WooCommerce anchors every tax calculation to that address. Any rate that calculated was untrustworthy in both directions.

It also found two competing tax plugins installed and neither configured. The recommendation to remove one hinged on a detail buried in a settings page: Stripe Tax overrides WooCommerce tax classes entirely, while WooCommerce Tax respects them. Since the whole exemption design depended on tax classes, that settled it.

A 404 That Was Actually an Architecture Problem

Apple Pay wasn’t working. The domain verification file returned a 404, which looks like a file problem and gets treated like one in most support threads.

It wasn’t. The store had three payment gateways live at once, and all three wanted Apple Pay. Apple’s own developer forum confirms that a single domain cannot host multiple merchant domain-association files — one domain, one provider, with subdomains as the official workaround.

No amount of re-uploading the file would have fixed that. The constraint reframed the problem from “fix this file” to “choose a payment architecture.”

The Defect Sitting in the Database

The payoff came during test setup.

Checking whether products were checkout-ready, the agent found all six service products flagged taxable with no tax class assigned — placing them in Standard. These were tax-exempt services.

It was inert at the time, because automated tax calculation was switched off. But the moment anyone enabled it, every service product would have started charging sales tax it should never charge, on real client invoices.

The research phase had predicted exactly that failure mode. The test phase found it waiting.

Where It Stopped Is the Real Story

The agent didn’t make the consequential calls. Which gateway should own Apple Pay, whether to consolidate processors, whether to test with real money — each came back as a question with tradeoffs laid out, not a decision already executed.

That boundary is doing a lot of work, and it’s the thing missing from most agent demos. The failure mode people fear from autonomous agents isn’t bad research. It’s confident, unreviewed action on production systems.

What scaled here was verification: reading actual stored state, checking primary sources, cross-referencing a vendor constraint that reframed a problem. What didn’t scale — and wasn’t asked to — was judgment.

For anyone evaluating agents for operations work, that’s the useful benchmark. Not whether it can do the task, but whether it knows which parts aren’t its call.

---

This piece is adapted from an internal working log kept by MCNM Marketing, which documented the session on its own store. Tax treatment described here reflects the regulations as read at the time of the session and varies by facts; nothing in this article is tax or legal advice.

---