There is a gap between what enterprise software vendors are promising and what their customers are actually experiencing. Vendors are talking about agentic futures and AI-first platforms. But for many organizations on the ground, the reality looks very different. An AI assistant is bolted onto a legacy system, doing light tasks in isolation, with no real sense of context.
This gap matters, and not enough vendors are being honest about it.
The reason the gap exists is architectural. Most AI added to enterprise software today is not reasoning. It is routing with better language. Understanding the difference, and what it actually takes to close it, is the question worth asking.
The expensive filing cabinet
Think about what an ERP system actually is at its core. It records everything, every transaction, every invoice, every approval, but it has no knowledge of what it is recording. It is, in essence, a very expensive filing cabinet. Extraordinarily reliable. Extraordinarily precise. And completely unable to tell you what any of it means.
For an algorithm to understand what it is looking at, you have to teach the system to reason. In practice that means adding a layer of information about the information, meta-information that tells the system what everything is. What is an invoice? What is a supplier? What is a credit note? What are the attributes that constitute each of those things, and what do they mean? Not just in a database schema sense, but in a language that can be understood both by a human and by a large language model.
That is what a semantic layer does. An ontology is the foundation of it, the structured account of what the objects are and how they relate. But the ontology is the skeleton, not the body. The real value is the meaning layered on top of it. It captures not just that an invoice relates to a purchase order, but what that relationship means in this industry, under these obligations, against these rules. That higher-order meaning is the part that takes years to build and the part that is hard to copy. Without it, an LLM processing your enterprise data is pattern-matching over records it does not understand. With it, reasoning has something real to stand on, and the deeper the meaning, the further the reasoning can go.
Three verbs. One progression.
Enterprise software has always moved work forward. But the way it moves work forward has changed in stages, and each stage represents a fundamentally different capability.
Workflow systems route. They move a task from one box to the next along a path a designer fixed in advance. Semantic systems interpret. They understand what the task is and how it relates to everything around it. Reasoning systems evaluate. They determine what should happen when the path was never drawn and the answer is not certain. Each capability contains the one beneath it. Routing alone is brittle. Interpretation alone is inert. It is evaluation that finally acts.
Consider what looks like a trivial instruction - "approve the invoice if everything looks correct." A routing system cannot act on it. It can present the invoice, validate the fields, and confirm the approver has the right permission. What it cannot do is form a view about whether approving is correct. So it hands the judgment back to a person, every time, and calls the handoff a workflow.
A reasoning system treats the same instruction as a decision. It reads the state of the invoice, the policy thresholds that apply, the history of the supplier, the relationship to the purchase order. It weighs those factors. If its confidence clears a threshold the organization has set, it approves. If not, it escalates, with the reasons attached. The invoice clears no faster than before. What has changed is that the system now forms a view before it acts.
Confidence is not the same as correctness
This is where the argument turns, because evaluative capacity introduces a problem that deterministic software never had to face. A system that evaluates can be confident and still be wrong.
Confidence is the system's own estimate of how likely its decision is to hold. Correctness is whether the decision actually held. The two are different properties. A system can read a duplicate invoice as routine, because every field matches a familiar pattern, assign the decision high confidence, and approve a payment that should never have been made. The pattern was real. The conclusion was false. Organizations that treat high confidence as proof of correctness will automate their mistakes faster than they ever made them by hand.
The answer to that gap is not a higher confidence threshold. It is a deterministic floor beneath the reasoning. A reasoning system proposes under uncertainty, but nothing it proposes commits until it has cleared the business's hard rules. The invoice must not be a duplicate. The entries must balance. The purchase order must exist and match. Those are facts, checked the same way every time, no matter how confident the reasoning was. Confidence belongs to the reasoning. Correctness belongs to the deterministic business layer beneath it. The two must never stand in for each other.
Governance is the third guarantee, and it asks a different question again. Not whether an action is correct, but whether the system is allowed to take it at all. This is the job of the control plane. It is not a feature to be bolted on later. It decides whether this agent, or the person it stands in for, may see this data and make this move under the organization's access rules and the regulation that applies, and it holds the boundary between what the system may do unattended and what it must escalate. Autonomy without it is not capability. It is exposure. This is the part we are building now, and the part most vendors are not addressing at all.
Reasoning, in the end, is not the removal of uncertainty but its management. The old systems eliminated uncertainty by refusing to operate inside it. A reasoning system stays in motion, acting where confidence clears a threshold the organization has calibrated, and surfacing the cases where it does not. Handled this way, reasoning becomes a disciplined way of acting inside doubt instead of waiting it out.
What most vendors are not talking about
Most of what passes for AI messaging in enterprise software today focuses on what the AI can do. A list of tasks, a count of agents, a demo of a workflow that used to take ten clicks and now takes one. Almost none of it addresses the architecture beneath. Few vendors are talking about the semantic layer. Fewer still are talking seriously about governance.
Data sovereignty is the clearest example. Organizations processing payroll, project contracts, and employee records have a legitimate concern about where that data goes when it is fed into an AI system. In ERP, routing sensitive data through an external provider is not a trade-off most organizations will accept, particularly those operating under strict regulatory frameworks. The answer is not to avoid cloud infrastructure. It is to be deliberate about where data lives and who can access it. For enterprise buyers in regulated industries, hosting in EU data centers with access restricted to EU-based governance is not a nice-to-have. It is a baseline expectation.
Governance will come back to bite the vendors who are not taking it seriously. The world of LLMs makes this more acute, not less, because there is less control over where data goes. That is not an argument against using LLMs. It is an argument for building the architecture that keeps them inside the right boundaries.
Why vertical depth is both a barrier and a differentiator
A reasoning system can never reason past the depth of meaning it has been given. Its judgment is bounded by the richness of the semantic model beneath it, and that richness is not generic. The meaning of a contract, a project, or an obligation shifts from one industry to the next. Professional services, public sector, higher education, each has its own objects, its own relationships, its own rules.
There has been a discussion in the industry about whether large language models could handle the verticalization themselves, scanning enough data to infer the nuances of a given sector. They cannot. Not because the ambition is wrong, but because the data that would make it possible sits in proprietary systems inside private data centers that LLMs cannot reach. Vertical knowledge is not something you can download. It is something you build over years, through deep engagement with the customers who operate in those verticals.
That makes vertical depth both a barrier and a differentiator at once. A barrier, because anyone entering a vertical faces the same steep climb. A differentiator, because those who have already made that climb hold something that cannot easily be copied.
The shift that was always the point
I have always believed that software should carry the load instead of handing it back to the people using it. For years, enterprise software recorded what people did and discarded why they did it, leaving the reasons in human heads and the system blind to them. Intent was the missing infrastructure.
A reasoning layer is what becomes possible once that infrastructure is finally laid. When a system can hold not only the state of the work but the purpose behind it, it can do the thing intent was always meant to enable. It can decide. Not freely, and not without governance, but within the boundaries we set, and with the meaning we have given it.
The organizations that will succeed with enterprise AI are not the ones that bolt on the most features. They are the ones that invest in the meaning layer and build reasoning on top of something real. That is the shift from a system of record to a system of reasoning. And it reorganizes everything above the ledger.
This article was originally published on diginomica.com.
