AI Enablement: The First Issue Beyond the Refinery

Last issue closed the Product Data Refinery Series with a commitment: the next stretch of The Digital Roadmap would push into the territory the refinery makes possible. This is the first piece of that territory, and it is the question I have been asked more than any other since the series began: where does AI fit? I have stayed deliberately quiet on it while roughly every third post in your feed explained how AI changes everything, because I wanted to wait until there was something for it to stand on.

There is now.

Why the refinery had to come first

Start with the sentence I have been holding back since this series began.

AI does not solve undefined business problems. It executes defined ones at scale.

That distinction is the entire subject of this issue, and it is the reason this issue could not have been written first. AI is a spectacular implementer and a terrible strategist. It will fill a field faster than any team you could hire, but it cannot tell you what belongs in the field, because that answer lives in your business, not in the model. It does not know your unit of measure standard. It does not know whether your substitutes are allowed to cross brands. It does not know what a “one” is in your warehouse. And if you do not know either, the two of you are going to have a wonderful time being wrong together at remarkable speed.

Now look back at what all that refinery-building actually produced. Every station generated two outputs. The first is the refined product data itself, the fuel. The second output is quieter, and it turns out to matter more for this conversation: the rules that do the refining. The intake standard from Stage 1. The category rulebook from Stage 2. The packaging validation logic from Stage 3. The supplier-confirmed terms from Stage 4. The attribute dictionary and content standard from Stage 5. The relationship definitions and prioritization logic from Stage 6.

That second output is what AI consumes. Rules are criteria, and criteria are the difference between an AI that scales your standard and an AI that generates plausible noise.

Which means the Product Data Refinery has been an AI-readiness program the entire time. We never called it that, mostly because I did not want to write an AI series. I wanted to write a data series that would still be true after the model names changed.

What breaks when AI is pointed at an undefined catalog

Let’s highlight the problem first, because this failure pattern is running rampant in the market and it is worth recognizing before you fund it. We all need to acknowledge our chasing of this shiny object! (The first step of the 12 step program is… )

The most common AI deployment in product data right now is enrichment over an undefined catalog. An organization with no attribute dictionary points a model at forty thousand thin records and asks it to fill in the blanks. The model complies. It always complies. What comes back is confident, well formatted, and wrong in ways no human team could ever achieve, because no human team works that fast.

Plausible is the most dangerous grade of fuel there is. A blank voltage field announces itself. A generated voltage field that reads 120V, because most products in that category are 120V, passes every eyeball test on its way to the buyer who needed 240V. In a consumer catalog that is a return. In B2B distribution it is a stopped line, a failed pressure test, or a safety conversation, depending on the product. The data did not look dirty. It looked finished. That is the problem.

The deeper failure is what automation does to an undefined process. Automating a defined process scales the standard. Automating an undefined one industrializes the mess. Every gap in your criteria gets reproduced at machine speed, and the cleanup project that used to arrive every eighteen months matures early.

Then comes the familiar ending. The merchandising team spot-checks the generated content, finds enough fiction to stop trusting all of it, and quietly goes back to doing things by hand. The business concludes AI was not ready. It is the same trust decay we watched kill hand-built relationship tables in Stage 6, replayed with a newer tool and a much bigger cost.

The tool was never the problem. The absence of criteria was.

What good looks like: AI at every station

Here is the same walk we took through the series, at speed, with the machine bolted on. Watch for the pattern, because the pattern is the argument: at every station, a person designs the criteria and the machine does the drudgery. Six stations, one division of labor.

Stage 1: SKU Definitions

The intake filter was designed by people: category confirmation, UOM governance, UPC capture, vendor designation, intent classification. AI now reads the vendor’s five-thousand-line file, checks every line against that standard, drafts the records that pass, and flags the ones that fail with the reason attached. The intake valve is still a valve. It just stops being hand-cranked.

Stage 2: Regulatory Data

The category rulebook, which toggles apply to which categories, is a human decision and stays one. AI reads the safety data sheet, extracts the hazmat classification, pulls Country of Origin from supplier documentation, populates the toggles the rulebook demands, and flags the products whose certificates are missing or expired. AI will read a forty-page SDS without complaining, which is more than I can say for anyone I have ever worked with, including me. The gate itself stays human-designed. Nothing ships because a model guessed it was probably fine.

Stage 3: Operational Data

Someone already answered the foundational question, what is a one, and encoded the validation logic that goes with it: each weight times case quantity must equal case weight, within tolerance. AI extracts dimensions and weights from spec sheets, drawings, and packaging photos, runs them through that logic, and surfaces the physically impossible answers, like the six-pound each that supposedly ships two hundred to a four-pound carton. The machine does not get bored at row forty thousand. Your team did, somewhere around row three hundred, and the error rate showed it.

Stage 4: Fulfillment & Pricing Data

The standard here was always distributor-defined and supplier-confirmed, and that does not change. AI reconciles supplier confirmations against your defined terms, watches actual receipt performance against stated lead times, and surfaces the drift before your promise dates do. The drop-ship versus stocking decision stays exactly where it belongs, with the business, because that fork is strategy and no model gets a vote.

Stage 5: Marketing & Merchandising Data

The taxonomy, the attribute dictionary, and the title and description discipline are yours. AI normalizes raw manufacturer content against the dictionary and drafts titles and descriptions inside the discipline, at a pace no content team can match. Without the dictionary, that same capability is a fiction generator with good grammar. And the fill-rate gate still decides what loads to the web, exactly as Stage 5 built it. The machine fills faster. The gate still gates.

Stage 6: Product Relationships

If Stage 6 read like it was written with something in mind, it was. The rows-versus-rules maturity arc was the AI conversation before this issue existed. You write the definitions and the prioritization logic. AI generates the candidates: matching competitor part numbers to your equivalents at a scale no cross-reference team could staff, pulling component lists out of service manuals nobody ever keyed. The definitions decide which candidates qualify. The prioritization ranks them. The live data maintains them. You govern rules. The machine produces rows.

Read back through those six and notice what the person did every single time. Designed. Decided. Reviewed. Now notice what the machine did: the part nobody ever listed under “what I love about my job.”

From doers to designers

That division of labor raises the question underneath every AI conversation, the one people ask carefully: what happens to the team?

The honest answer is that the work changes shape, and it changes in the team’s favor. What AI absorbs is the portion of the job your best people already resented: keying dimensions off a spec sheet, copying descriptions between systems, auditing a spreadsheet row by row to find the six lines that drifted. Nobody built a career dream around cell-by-cell auditing. It was the tax we paid because the work had to be done and there was no other way to do it.

There is now another way to do it, and it moves the person up a level. The data specialist who spent forty percent of the week entering attributes becomes the person who designs the validation logic the machine runs, reviews the exceptions it raises, tunes the rules when the exceptions reveal a gap, and decides what the standard should be for the next category before the machine touches it. The auditor becomes the exception reviewer. The enterer becomes the standard designer. The title may not change for a while. The job already has.

I am aware that “data management is interesting” is a sentence with a credibility problem. This discipline earned its eye-roll honestly, one million-row spreadsheet at a time. But the eye-roll was always about the drudgery, and the drudgery is precisely the part that just got delegated. What is left is design: encoding how the business thinks about its products into rules a machine can execute. That is judgment work. That is strategy work. And the people who get good at it, who can design rules over governed data, are going to be some of the most valuable people in the building over the next decade. If you are early in a product data career, this is the skill to run toward while everyone else is still asking whether AI will take the job.

The leadership mindset: scale is a resource allocation decision

For leadership, this issue reduces to one decision about hours.

AI does not shrink the product data function. It changes what the function’s hours buy. Every hour the machine absorbs from entry and auditing is an hour that can be redeployed into designing standards, extending the rulebooks into new categories, and tuning the relationship logic that Stage 6 put in leadership’s hands. That redeployment is the entire return. The organizations that instead harvest those hours as headcount savings will discover they cut the design capacity that made the machine useful, and their standard freezes at whatever it was the day they cut it.

The moat argument from the end of the series gets sharper here, not weaker. Your competitors can rent the same models you can, tomorrow morning, at the same price. What they cannot rent is your attribute dictionary, your category rulebooks, your relationship definitions, or the people who know how to write and tune them. The models are becoming a commodity. The criteria are the differentiator. Fund accordingly.

And the ownership split from the capstone carries straight through. IT connects the model, secures it, and operates the infrastructure it runs on. The business designs the rules the model executes, because the rules are business decisions wearing a technical costume. Hand AI enablement to IT alone and you will get a working integration pointed at an undefined problem, which is exactly what this issue opened by warning you about.

Where to start: pilot where your rules are strongest

The capstone told you to diagnose your refinery and start with the weakest station. For AI, invert that advice, on purpose.

The diagnostic finds the weakest station because that is where the operating model needs attention. An AI pilot starts at your strongest station, the one with the most rules actually written down, because AI amplifies whatever it is pointed at. Pointed at a written standard, it scales the standard. Pointed at ambiguity, it scales that instead.

So pick the station where your criteria are most defined. Pick one rule and one category. Let AI execute against that rule, and have a person review every exception it raises for a month. Track how often the machine and the reviewer agree. That agreement rate tells you whether the rule is defined well enough to trust at scale, and where it is not, the exceptions will show you exactly which part of the standard was living in someone’s head instead of in the rulebook.

If you go looking for a station with written-down rules and cannot find one, that is also an answer, and it is not answered with a software purchase. Go back to the diagnostic. Build the station. The machine will still be there when the criteria are.

What comes next

You may have noticed the pilot ends with a measurement, not a go-live. That was not an accident. Before any of this scales, someone has to be able to see how it is performing: the quality of what the machine produces, the pace of what it enables, and the story the executive team needs in order to keep funding it. You cannot govern what you do not measure, and you should not automate what you have not measured. Next issue, we take on governance and the analytics that make it real. The machine is bolted on. Now someone needs to watch the gauges.