Governance and Analytics: The Second Issue Beyond the Refinery

Last issue ended with a machine bolted onto the refinery and a warning that someone needs to watch the gauges. We will get to the gauges before this issue is done. But between that issue and this one, something kept happening in my calendar that changed how this piece opens, and I want to start there, because it is the clearest evidence I can offer that the industry has arrived somewhere new.

Over the past several months I have sat in conversations with manufacturers and distributors on both sides of the Atlantic, from mid-market operations to multi-billion-dollar global companies. These organizations have nothing to do with each other. Different products, different markets, different leadership, and, at least as they described them, different problems. And every one of those conversations, no matter what it started as, ended in the same place.

The same conversation, everywhere

One company has grown through acquisition for years and just paused that strategy. The deals still made financial sense. The problem was that every acquisition imported another catalog, another data model, another set of habits, and the organization finally admitted that each new deal was making an existing problem worse. They have since hired a leader whose actual mandate, stripped of the title, is to hold the business accountable to enterprise data practices that do not yet exist.

Another is a multi-billion-dollar operation with a different PIM in nearly every acquired business unit. They want to consolidate. One PIM is a strategy. Five is a museum of past strategies, and every exhibit was expensive.

A third ran a data modeling project on its top product families about a year ago. Good work, properly scoped, fully implemented. A year later they are back at the table, because the project was a success by every measure except the one where the data stayed fixed.

A fourth is scoping data lineage work: the discipline of proving where every field came from, who touched it, and what authority they had to touch it.

Four organizations. Four different doors. All of them standing in the same room. And here is the detail that ties it together: in every one of these conversations, the stated ambition, the reason the initiative has executive attention and budget behind it, is AI.

Not one of them asked for governance by name. They asked for consolidation, for accountability, for lineage, for an AI strategy. Governance is the word underneath all four requests. When a market keeps arriving at the same answer from different directions, without coordinating and mostly without naming it, you are not looking at a trend. You are looking at a wall, and everyone is reaching it.

Why AI forced the issue

The obvious question deserves a direct answer. AI was supposed to help with data. It does help, at exactly what last issue said it does: executing defined problems at scale. So why is the arrival of the most capable execution technology in a generation exposing data problems instead of erasing them?

Three mechanisms, and they build upon each other.

The first: AI removed the buffer that made ungoverned data survivable. For decades, bad product data was absorbed by people. The branch manager knew the real lead time regardless of what the system said. Customer service caught the unit of measure error before the invoice went out. The seasoned buyer read past a thin description because he knew the product line better than your catalog did. We watched that buffer collapse across the refinery series, first with e-commerce, then with the agentic buyer of Stage 5. AI finishes the job, because the machine now consumes the data directly on both ends of the transaction, with nobody in the middle compensating. Ungoverned data used to cost you friction. Now it costs you the outcome itself.

The second: AI amplifies the current state, whatever the current state is. That was the entire argument of last issue, and the organizations above are living the inverse of its promise. Pointed at a written standard, AI scales the standard. Pointed at five acquired catalogs with five data models, it scales the contradiction, at a pace and volume no human process ever achieved. The governance gaps in these companies are not new. They are decades old. But they used to leak at human speed and get mopped up at human speed. AI turns slow leaks into pressure failures, and pressure failures get executive attention.

The third mechanism is the one I think the industry has not fully articulated yet: the AI mandate is the first corporate initiative that audits your data on day one and cannot be faked past it. E-commerce limped along on mediocre data for years because humans compensated at every step. ERP implementations buried the problem inside a line item called migration cleanup. But the honest first step of any AI initiative is pointing the machine at your data. That is the moment decades of deferred decisions become visible all at once, in the first pilot, in front of the executives who funded it.

AI did not create the governance crisis. It ended the industry’s ability to defer it. The bill for thirty years of “we’ll clean that up later” finally arrived, and AI turned out to be the collection agent.

What breaks: the end of the project era

Look at how these organizations, and hundreds like them, have historically responded to data pain, because the pattern of response is itself the diagnosis.

The default response has been the project. Scope a cleanup, fund it, run it, declare victory, disband. The one-off data modeling engagement above is the pattern in miniature: the work was genuinely good, and a year later it had decayed, because a project ends and data does not. Products kept getting created. Suppliers kept changing packaging. Categories kept growing. The moment the project team walked away, the catalog went right back to drifting, because the daily habits that created the mess never changed. I have sat on both sides of these projects, and the decay is not a talent problem. Good teams did good work. The organization just never assigned anyone to keep it done.

The second response has been technology. Somebody buys a PIM, or an MDM platform, or now an AI tool, on the belief that the system will solve what the organization has not. It will not, and the company with a PIM in every business unit is the proof, purchased five separate times. A PIM without governance is an expensive place to store ungoverned data. The technology multiplies right alongside the problem, because every new leader and every new acquisition buys its own copy of the same hope.

The third pattern is acceleration by acquisition. Every acquired business arrives with its own part numbering, its own units of measure, its own definitions of what a product even is. Without a governing standard to integrate into, each deal stacks another ungoverned catalog on the pile, until the day a leadership team pauses a growth strategy, an actual growth strategy, board-approved and capital-backed, because the data debt underneath it became unmanageable. If you want evidence that this topic has graduated from hygiene to strategy, there it is. Data governance just vetoed M&A.

And now the newest member of the family: the stalled AI pilot. The enrichment initiative that produced confident fiction. The agent that quoted the wrong product because the relationship data underneath it was assembled by hand years ago and never maintained. We covered the mechanics last issue. What matters here is the pattern joining the others: one more initiative that failed for the same root cause under a different name.

That is the claim I want to make plainly, because it is the thesis of this issue. Duplicate SKUs, PIM sprawl, decayed cleanup projects, acquisition indigestion, stalled AI pilots: these are not five problems. They are one problem with five symptoms. Standards that nobody owns, that no system enforces, and that nothing measures will regrow every weed you pull, no matter how good the pulling was. One of the organizations above already ran that experiment and paid for the results twice.

What governance actually is

Governance has a reputation problem, having spent decades as the name of a committee that meets monthly to produce more meetings. So let me define it the way this series has earned the right to define it.

Governance is the system of decisions that keeps refined data refined after the project ends.

It has four components, and if you have been following this series, you have already built two of them.

Ownership. Every standard has a named owner with the authority to change it and the accountability when it fails. Not a team, not a committee, a person. The refinery series settled where this lives: the business owns the rules, because the rules are business decisions wearing a technical costume, and IT owns the platforms the rules run on.

Standards. The rules themselves. This is the part the refinery already produced, station by station: the intake standard, the category rulebooks, the validation logic, the supplier-confirmed terms, the attribute dictionary, the relationship definitions. Last issue called these the refinery’s second output, the one AI consumes. Governance is what keeps that output alive.

Enforcement. Standards applied by systems, not by vigilance. The fill-rate gate that stops a thin product from loading to the web. The validation check that rejects a physically impossible case weight. The vendor-level designation that overrides an individual creator’s guess. A standard enforced by people remembering to check is a suggestion with good intentions.

Measurement. The evidence that the standard is holding, which gets the next section to itself, because it is the component almost everyone skips and the reason the other three quietly rot.

Notice what this definition does to the project-versus-program language the market is already reaching for. A project has an end date. The market keeps saying “program” because data work that ends is data work that decays. The companies asking for programs are asking for governance. Most of them just have not used the word yet.

The definition also reframes last issue. From Doers to Designers made the case that your people shift from entering data to designing the rules the machine executes. Designing the rules was half the job. Governing them, owning them, enforcing them through systems, and watching whether they hold, is the standing half of the job, and it never ends. The refinery refines the data. Governance keeps the refinery honest.

The gauges: analytics that make governance real

Here is where analytics enters, and why it belongs inside governance rather than off in its own issue. Governance without measurement is a policy binder: well organized, carefully reviewed, and ignored by the operation it claims to run. The gauges are what separate a governed refinery from a documented one.

Every data quality conversation I have with clients eventually lands on the same three words: completeness, consistency, and accuracy. Those three words are your quality gauges. The instruments below just make them measurable.

Completeness is fill rate against the attribute dictionary, tracked by category, because catalog-wide averages hide exactly the categories that are bleeding. This is the simplest gauge on the panel and the first one most teams stand up.

Consistency is agreement rate, straight from last issue’s pilot, and it deserves a careful definition because it is the least familiar gauge here. The machine applies a rule to a set of records. A person applies the same rule to the same records. Agreement rate is how often they land on the same answer. Say your description standard for cut-off wheels requires diameter, arbor size, and max RPM pulled from the supplier file. The machine processes five hundred records against that standard, your reviewer works the same five hundred, and they agree on 460 of them. You are at 92 percent, and the forty disagreements are worth more than the 460 agreements. Each one means the machine misread, the person misread, or, most often, the rule itself is ambiguous and two reasonable judges filled the gap differently. A rule that two judges apply identically is a rule that is actually finished being written. A low agreement rate is not a grade on the machine. It is a grade on the rule.

Accuracy gets two gauges, because you can measure it early or late. Exception rate is accuracy caught inside the refinery: how often the machine flags a record that fails a rule, while the fix is still cheap. Escape rate is accuracy that got past the gates and surfaced downstream, in a return, a mis-ship, a customer call. In my experience, the single best place to sit as a data steward owning governance is within the help desk tickets. They typically tell you all the gaps with the escape rate, one ticket at a time. And read the two accuracy gauges together: if escapes are rising while exceptions stay flat, your gates are checking the wrong things, and no single gauge would have told you that.

The scale gauges are the second family, and they tell you whether the operation is growing. Throughput: records refined per week, now that the machine does the drudgery. Time to onboard: how long from receiving a vendor’s file to products ready for sale, which is the number your merchandising and purchasing teams feel in their bones. And rule coverage: what share of the catalog is governed by written, enforced rules versus tribal knowledge. Coverage is the quiet one, and it might matter most, because it measures how much of your refinery actually has gauges at all.

Watch only one family and you fail in the direction of whichever one you picked. All speed and no quality is last issue’s failure mode with better reporting: you have industrialized the mess and can now prove how fast. All quality and no speed produces a beautifully curated catalog that cannot keep pace with the business it serves. The two families exist to hold each other honest.

It’s best to start small and build as you develop the right methodology with your team. Three to five gauges on one page, reviewed weekly by the people who own the rules. Forty metrics reviewed by nobody is wallpaper, not instrumentation.

The executive story

The gauges are for the control room. Leadership needs something different, and the difference is where most data teams lose the funding conversation.

An executive does not need your fill rate. An executive needs to know what the fill rate bought. The translation is the skill: fill rate against the dictionary becomes products that can be found and trusted by the AI agents doing your customers’ buying. Escape rate becomes returns prevented and stopped lines avoided. Time to onboard becomes revenue reaching the market weeks earlier. Agreement rate becomes the confidence to hand the machine the next category. Throughput becomes the one leadership cares about most, whether it knows it yet or not: hours returned to the team, redeployed into designing the next standard instead of keying the next spreadsheet.

That last translation is the change management story, and it decides whether this whole arc survives contact with a budget cycle. Last issue argued that the return on AI is redeployed design capacity, not harvested headcount. The gauges are how you prove the redeployment is producing something. One translated story per month, tied to a number leadership already watches, will protect this work better than any dashboard ever built. Executives fund narratives backed by evidence. They defund dashboards they were never taught to read.

The leadership mindset: governance is a growth capability

For leadership, the reframe this issue asks for is short: governance has graduated from hygiene to growth capability, and the market has already ruled on it.

A growth strategy paused because the data underneath it cannot absorb another acquisition is a board decision, and its cost is measured in deals that did not happen. Companies that govern their data integrate acquisitions faster, onboard suppliers faster, extend catalogs faster, and point AI at all of it with confidence, because the standards the machine needs are written, owned, enforced, and measured. Companies that do not will keep buying technology and running projects, and the wall will keep being exactly where they left it.

Which means the executive question changes. “What is our AI strategy” is the wrong first question, and by now most leadership teams quietly suspect it. The right first question is: who governs the data our AI strategy depends on? If the answer is a committee, a vendor, or a shrug, the AI strategy is a press release with a budget.

The ownership split does not move. The business owns the rules and the priorities. IT owns the platforms, the security, the integration. What leadership owns is the standing nature of the thing: governance is a function you operate, like safety or quality, not an initiative you complete. It appears in the org chart and the budget every year, or it does not exist.

Where to start: the four-part test

Take the domain where your AI ambition is highest. Not the weakest station, not the loudest complaint, the place where the machine is headed first. Now ask four questions about the data standard there.

Is it written down? Does a named person own it? Does a system enforce it, or does vigilance? And can you see whether it held last month?

Any “no” is your starting point, and the order of the questions is the order of the work. A standard that is not written cannot be owned. An owned standard that is not enforced is a suggestion. An enforced standard that is not measured is an assumption. Most organizations discover they fail at question one, which is genuinely useful information, because it means the AI initiative was about to automate something that does not exist.

Then instrument the pilot from last issue. That one rule, one category, one month of a person reviewing every exception: turn its agreement rate into your first standing gauge. Add two to four more across completeness, consistency, accuracy, and scale, put them on a single page, and review them weekly with the rule owners. At the end of the month, translate one of them into a business number and walk it into a leadership meeting. That is a governance program in miniature: owned standards, enforced by systems, measured weekly, and translated for leadership. Everything past that point is scale.

What comes next

You now have a machine executing your rules and gauges telling you whether the rules are holding. Some of those gauges are going to do something interesting over the next few months: they are going to go quiet. Agreement rates that settle in the high nineties. Exception rates that flatten near zero. Escapes that stop escaping. Quiet gauges are the interesting ones, because they mark the parts of the refinery that have earned the right to run without a hand on the valve. Next issue closes the arc: automation, and how to grant that right only to the processes that have proven, on instruments you trust, that they deserve it. You measured. Now we can talk about letting go.