Walk through a metals plant on a Tuesday morning and watch what happens when the phones start ringing. A customer calls about an order. The CSR pulls up the system, scans for the part number, and finds three slightly different entries. One has the right grade but the wrong dimensions. One has the right dimensions but a legacy customer code. One was created by a sales rep who left in 2019. She picks the correct one in under three seconds because she remembers the conversion that happened when this customer renamed their procurement system in 2022.
That decision is your moat.
Most industrial companies have been told the AI question is whether they have a clean data moat. It is the wrong frame. The mid-market manufacturer with 40 years of operational scar tissue does not have clean data and never will. What they have is something more valuable: a workforce that has internalized thousands of exceptions, workarounds, and unspoken rules that make the business run.
AI will commoditize businesses built on generic intelligence. It will deepen the advantage of companies that can encode their specific operating judgment. In legacy industries, the moat is the exception logic your best people use every day without realizing it.
The Shape of the Actual Asset
Spend any time in a logistics dispatcher's queue and the pattern is everywhere. A load comes in marked as a single shipment, but the dispatcher knows it is three pickups because that broker always books it that way. A vendor portal kicks back an invoice because line items do not match the PO, and the AP clerk knows to recode line 7 to a different GL account because that vendor splits freight differently than everyone else. A new sales order references an anchor customer order from six months ago, and the operations manager knows that means inheriting routing preferences nobody has ever written down.
These are not edge cases. They are the work. The system of record captures the official transaction. It rarely captures the judgment that made the transaction correct. The rest lives in habits, side conversations, and the kind of pattern recognition that takes a decade to develop.
Where the Value Lives
The common failure mode is not consulting itself. It is treating AI deployment as a data exercise instead of an operating system redesign. Clean up the official transaction and you still miss the judgment that makes the workflow correct.
Why Model Commoditization Makes This More Valuable, Not Less
There is a popular argument that AI will flatten competitive advantages because every operator gets access to the same frontier models. That is true for businesses built around generic workflows. The more easily a workflow can be described, demoed, and packaged, the more easily the model layer or the incumbent software layer can absorb it.
Industrial operations have a different shape. The work is not hard because the language model cannot reason. It is hard because the business is not legible to the model. As models commoditize, the scarce input becomes the context that makes them useful inside a specific operation. A frontier model with no operational context is a confidently wrong intern. The same model wired into the decision logic of your dispatch desk is something else entirely.
The moat is not the model. It is also not the data, narrowly defined. The moat is the encoded understanding of how your specific business handles its specific exceptions.
That encoding is what compounds. Every time a workaround gets written down, mapped, and made queryable, the next system built on top gets better. The dispatcher's tribal knowledge becomes routing logic. The AP clerk's coding instincts become invoice classification rules. The CSR's customer memory becomes a context layer any agent can reach into.
Why This Is Suddenly Urgent
The first wave of AI companies wrapped generic intelligence around generic workflows. That advantage is fragile. If the workflow is easy to describe, easy to demo, and easy to package, the model layer or the incumbent software layer can absorb it.
Industrial operations are different. The data exists, but it is not agent-operable. It is scattered across ERP fields, emails, spreadsheets, tribal memory, vendor portals, customer-specific rules, and habits nobody has written down. Until that context is captured, even the best model is operating on a cartoon version of the business.
Most companies technically have data. Very few have agent-operable data.
Agent-operable data is data with enough structure, context, permissioning, history, and workflow connection that an AI system can safely act on it. A PDF SOP is not agent-operable. A spreadsheet full of customer exceptions is not agent-operable. A veteran dispatcher's memory is not agent-operable.
The transformation work is converting operational knowledge into a form agents can use. That is why the opportunity is not “use AI on your data.” The opportunity is to make the business understandable to agents.
Even the frontier labs are revealing the bottleneck. OpenAI launched the OpenAI Deployment Company to embed forward-deployed engineers into customer operations. Anthropic and Deloitte expanded their alliance around implementation frameworks, technical support, and moving pilots to production. The labs are not just selling models anymore. They are building deployment arms and systems-integrator partnerships because enterprise value lives in implementation, workflow redesign, and adoption.
Operational Relationships, Not Just Embeddings
Embeddings help an AI system find related information. Ontologies help it understand operational relationships.
In this context, an ontology is a map of how the business treats customers, parts, vendors, orders, rules, and exceptions. It encodes that this part number is the successor to that one, that this customer's “rush” means same day while that customer's “rush” means within the week, that this vendor's invoices always need a manual freight adjustment.
A retrieval system over your SOPs will find the document. It will not understand why three different SOPs contradict each other and which one is actually followed. An ontology built from the exceptions your best operator knows can.
This is unglamorous work. It does not get keynote slots at conferences. It determines whether your AI deployment survives first contact with the real business.
| Embeddings | Ontologies |
|---|---|
| Find things that look similar | Encode why two things are the same |
| Surface candidate documents | Resolve which document actually governs |
| Cheap to build, generic by design | Slow to build, defensible by design |
How to Capture This Without Boiling the Ocean
The trap is to commission a six-month data and process discovery effort. By the time it finishes, the business has moved, the consultants have left, and you have a 200-page document nobody reads.
A better pattern, drawn from what actually works in mid-market deployments:
Pick one workflow that hurts. Not the most strategic one. The one where your best person is the bottleneck and you cannot hire your way around it. Invoice coding, order entry, dispatch triage, claim review, returns auth — whatever is currently held together by the institutional memory of two or three people.
Sit with one of those people for a week. Not to interview them. To watch them work and write down every decision that surprises you. Every “oh, this one we always do differently.” Every keyboard shortcut, side spreadsheet, and Slack message they send to confirm something the system already showed them.
Build the smallest possible agent or workflow that handles 70% of routine cases cleanly and routes the other 30% to the operator with full context attached. Watch what they do with the weird ones. Feed those decisions back. Repeat.
After six months you do not have a transformation. You have one workflow where exceptions are mapped, one operator who is no longer the bottleneck, and the beginning of an ontology the next workflow can borrow from.
That is how the moat actually compounds.
The Bet
Everyone will have access to the same models.
Almost nobody will do the hard work of making their business legible to those models.
The next decade will not be won by the companies with the cleanest data warehouse. It will be won by the companies that turn tacit knowledge into agent-operable workflows before their competitors do.
The exceptions are the moat.