The chatbot was the wrong mental model for enterprise AI. It made everyone look at the conversation. What did the user ask? What did the model answer? Was the answer accurate, safe, polite, grounded, compliant? Those are real questions, but they are not the main questions once AI starts doing work.
The main question becomes: what procedure is now executable?
Most companies are full of work that is already halfway to software. A request arrives. Someone checks a source. Someone applies a rule. Someone asks for missing information. Someone drafts a response. Someone gets approval. Someone updates a record. Someone files the evidence. The work is not fully deterministic, because judgment and exceptions live inside it, but it is not mystical either. It is a procedure with handoffs, authority, tolerances, and a memory of what happened last time.
That is where enterprise AI gets interesting. Not when a model gives a better answer inside a chat window, but when a recurring operating procedure becomes something the system can execute, inspect, and improve.
The SOP becomes the product.
This sounds dull, which is probably why it matters. Enterprise software is usually sold as an application, a dashboard, a copilot, a data platform, or a transformation programme. But the actual value sits in a much less glamorous place: the repeatable way work happens. A good SOP says what counts as valid input, what sources must be checked, who can approve which action, what evidence must be retained, when to escalate, and what cannot be decided locally. That is already a product spec. It just usually lives in a PDF, a SharePoint page, a training deck, or the heads of three experienced operators.
AI changes the status of that document. The procedure is no longer just guidance for humans. It becomes an execution surface.
That does not mean removing humans. The lazy version of this story says the model does the work and the human moves up to strategy. That misses the actual control problem. Humans do not disappear from the workflow. They become more explicit. Their roles have to be named: approver, exception owner, policy interpreter, quality reviewer, accountable manager, customer-facing decision-maker. A human in the loop is not enough. The workflow has to say what the human is in the loop for.
This is why the model is not the unit of enterprise AI. The workflow is.
A model can be excellent inside a bad workflow and still produce bad work. It can have the wrong source, the wrong permission, the wrong escalation path, the wrong definition of done, or the wrong audit trail. Conversely, a merely competent model inside a well-designed workflow can be useful because the limits are visible. The system knows what it checked, what it failed to check, what it is allowed to change, and where it must stop.
That is also why model inventories are too thin as the main governance artefact. They tell you which model is being used. They rarely tell you what work the model is embedded inside. For agentic systems, the useful inventory is closer to a workflow map. What procedure is being executed? Which systems can it read? Which systems can it write to? Which steps are deterministic? Which steps require judgment? Where does approval happen? What evidence is retained? What happens when the source is stale, the tool fails, or the case falls outside policy?
Those questions sound operational, not futuristic. Good. That is the point. Enterprise AI becomes real when it stops being an impressive sidecar and starts inhabiting the boring machinery of work.
The companies that get this right will not just buy better models. They will turn their best operating procedures into living systems. The procedure will run. The evidence will accumulate. Exceptions will teach the next version. Controls will move from policy language into the path the work actually travels. Improvement will become less like a quarterly process review and more like maintaining a product.
This is a different job from prompt engineering. It is closer to process design, product management, control design, and institutional memory. You have to know what good work looks like before you can ask an agent to do it. You have to decide which parts can be automated, which parts can only be assisted, and which parts must remain accountable human judgment. The system will not find that boundary for you. It will happily execute the wrong procedure at scale if the procedure is badly drawn.
The enterprise AI question is therefore not “which model should we use?” That question still matters, but it comes too early and carries too much prestige.
The better question is: which SOP is ready to become a product?