What a Magento operator learned that AI consultants get wrong.

Fourteen years of pricing rules and B2B tier logic and warehouse routing teaches you what AI is actually for. Most of our competition didn't go to this school. They went to a different one, the one where you learn to raise a demo in a weekend and a round on the demo, and the difference between the two educations shows up in everything the two kinds of consultants ship.

Let me describe the school first, because it does not sound like an AI credential and that is the point. From 2012 on, my company built and ran Magento stores: storefronts, ERP integrations, B2B pricing engines, fulfilment workflows, about 250 mid-to-large projects across fourteen years, for clients whose largest group moved 340 million euros in GMV. Half of my core team contributed to Magento core itself. None of that involved a model. All of it involved the exact material AI projects are made of: business rules, dirty data, integrations, and people who need the system to work at seven in the morning.

The first thing the school beats into you is that most of a business is rules, and rules want to stay rules. A B2B pricing engine is customer group tiers stacked on contract prices stacked on currency logic stacked on promotions, and every layer is deterministic because money is deterministic. When a distributor's biggest account gets quoted the wrong tier, nobody wants to hear that the system is right ninety-something percent of the time. The AI-native consultant looks at that engine and sees a place to put a model. The operator sees a place to absolutely never put a model. The model belongs at the edges of the rules, on the corner cases the rules cannot reach, not in the middle of logic that must be right every single time. I have written the long version of that argument in the rule-based essay; the short version is that fourteen years of pricing disputes teach it faster than any whitepaper.

The second lesson is that the data is never clean, and the mess is not a phase before the real work. The mess is the work. Every store we ever ran ate vendor files with columns nobody could explain, SKUs that meant different things in different warehouses, attributes typed by hand by someone in a hurry three years ago. The demo-first consultant builds on sample data and shows you magic. Then production data arrives and the magic produces confident nonsense, because the demo was trained on a world that does not exist. Operators start from the opposite assumption. The input is hostile until proven otherwise, the pipeline needs validation at every seam, and the interesting question is never can the model do it on good data. It is what happens on the data you actually have.

Third lesson: the Tuesday morning test. Warehouse routing taught me this one personally. A routing rule that looks elegant on the whiteboard meets a picker at 7 a.m. who knows the elevator on the second floor is broken and routes around your system with a pen. If the software does not let her override it, she overrides the software, and now your data is fiction. Any system that cannot absorb a human override is not an operational system, it is a slide. AI consultants who never ran operations design for compliance: the model decides, the human obeys. Businesses do not work that way on any morning, and the AI that survives inside them is built for override from the first sketch, which is why we build ours that way. It is the same failure I described in the manifesto, seen from the warehouse floor instead of the boardroom.

Fourth lesson: errors have a price tag you can read. When a checkout breaks during trading hours, you can watch the revenue not arriving, per hour, in a dashboard. Running stores wires that meter into your nervous system. So when an AI vendor says the extraction is about 85 percent accurate, an operator immediately asks the operator questions. Which 15 percent? Detected how? Who fixes them, at what cost per fix, and does the fixing cost eat the saving? A demo-first consultancy hears 85 percent and says transformative. A store operator hears 85 percent and says: that is 15 wrong rows per hundred flowing into an ERP that orders real inventory with real money. Neither reaction is wrong in its own world. Only one of them survives contact with yours.

Fifth lesson: integration is the project. In fourteen years I do not think we ever shipped anything that lived alone. Everything talked to an ERP, a PIM, a WMS, a payment provider, a shipping API, usually several at once, usually versions of them the vendor stopped documenting years ago. The AI-native instinct is to treat integration as plumbing someone else will do after the interesting part. The operator knows the plumbing is where the months go and the interesting part is a week. An AI feature that has no place in the integration map is not a feature. It is a toy with an API bill attached.

So what is AI actually for, according to this school? Almost exactly what the demo school uses it for last.

• The corner cases. The rules handle the predictable bulk cheaply and perfectly. The model handles the long tail you could never afford to hand-code: the weird vendor file, the malformed address, the product description in the wrong language.
• The error catching. Pointing a model at a thousand mapped rows and asking which cells look wrong is one of the highest-value uses I know. It is also unglamorous, which is why it rarely makes the demo.
• The drafts. Descriptions, mappings, classifications, produced by the model, approved by a human whose override is a first-class citizen of the workflow, not an embarrassment to it.
• The translation between mess and structure. Taking the unstructured thing a business actually has and proposing the structured thing its systems need, with a confidence signal honest enough to route the doubtful cases to a person.

Notice what is missing from that list: the autonomous agent that runs your company. The school teaches you why it is missing. Every process I ever automated had an exception path worn smooth by use, and the exceptions were not noise around the process. They were the process, the accumulated judgment of everyone who ran it before. You do not replace that with a model. You give the model the bulk and the judgment the exceptions, and you measure the seam between them relentlessly.

The demo-first instinct is not stupid, and I want to be fair to it, because the people who hold it are often technically better with models than my team was on day one. It is optimized for a different game. A demo needs to be impressive once, in a controlled setting, to an audience deciding whether to believe. An operation needs to be boring ten thousand times, in an uncontrolled setting, for an audience that stopped watching. The instincts that win the first game, hide the edge cases, maximize the wow, defer the plumbing, are precisely the instincts that lose the second. Nothing in an AI-native education punishes those instincts, and everything in an operator's education does. That asymmetry, not any secret model knowledge, is the actual moat.

There is a version of this essay that ends with my company frozen in amber: the proud Magento shop, suspicious of the new thing. The truth is the opposite and slightly un-sexy. We stopped being only a Magento shop years ago, the way any operator adapts when the market moves; we built Shopify and BigCommerce capability the same way we later built AI capability, with a month of business analysis before a single pitch, formulating exactly what the new tool can and cannot do compared to the one we knew. That is the transferable skill. Not loyalty to a platform. The habit of mapping a technology's real limits against a real business before recommending it, so the recommendation survives the client's Tuesday morning and not just the pitch meeting.

So when you evaluate an AI consultant, ask them my school's entrance questions. Ask what they would refuse to automate, and why. Ask which part of your data they expect to be lying to them. Ask where the human override lives in their design, and what happens to their accuracy claim when it meets your worst quarter's data. The graduates of the demo school will answer with capabilities. The operators will answer with consequences. Fourteen years of pricing rules, tier logic and warehouse routing is one long lesson in consequences, and it turns out that lesson, not the model, was the qualification all along.