At the AFPM Annual Meeting, our team watched a room of executives answer the same live poll questions and reach different conclusions. The data was the same, but the realities behind the answers were different.
Executives agree on what AI promises. Operators are clear about where its value shows up, which is at the moment of decision. The two groups do not yet agree on how to get from one to the other. That gap between ambition and execution is becoming the central challenge for industrial AI, and it is widening.
Vision has outpaced reality
Digital transformation has dominated boardroom conversations and conference agendas for years. AI is now presented as the next competitive lever, with the promise of wider margins, safer operations and a more productive workforce.
Leadership expectations are clear and increasingly measurable. Executives no longer ask whether AI can work. They ask whether its return can be proven against margin, and whether it can scale across sites. That is a much higher bar.
The frontline works in a different reality
Inside the plant, the conversation is different. Operators are not asking for transformation. They are asking for something more immediate and more practical: help making a better decision now, on this shift.
A common objection we hear is that a site does not have the data, or does not have it in the right place or format. Usually the data exists but is hard to reach. It is scattered across systems, and finding it and pulling it out of the context it was recorded in eats people's time, which leaves a heavy cognitive load on the people who have to act on it.
So the problem is one of timing, context and cognitive load. Day to day, the frontline works under constraints that most transformation plans ignore:
- limited time
- fragmented systems
- constant interruptions
- serious consequences for getting a decision wrong
Their questions are about what is in front of them:
- Can you help me make a better decision this shift, and consistently across every shift?
- Can you help me find what I need faster?
- Will this work within my constraints?
Today the honest answer is often no. In one discussion, people said they spend a large share of their working time searching for information. That is attention lost at the moments that need it most.
Two measures of value
The poll made the tension visible. Asked for leadership's primary expectation from AI, about 51% of respondents chose productivity gains and safer operations. Asked where people actually feel AI making a difference, the answer was better operating decisions.
The two are connected, but they are not the same. The first is strategic, measured over quarters and across sites in aggregate outcomes. The second is tactical, felt in seconds inside a single decision made under pressure. The strategic outcome does not arrive until the tactical problem is solved.
Until AI shows up at the moment an operator needs it, its value stays theoretical. For the frontline, value comes down to one question: did this help me make a better decision right now?
Most systems were not built for this
Most AI systems were not designed for this environment. However sophisticated, they are built on the wrong abstraction. They model individual units, generate isolated insights and depend on workflows that sit outside the moment of action.
A plant does not run that way. When a crude unit is pushing for margin, a downstream unit is approaching its constraint and a shift change is under way at the same time, they form one plant-level decision that has to be resolved in real time.
The mismatch shows up in practice as static models that degrade, dashboards that go unread and pilots that stall. Operators have the information, but it sits in silos that are optimized independently, handed off in sequence and never reconciled in real time. That is what breaks down, and the industry is starting to name it.
Three reasons come up again and again when companies explain why AI has not delivered: pilots that never scale, poor data quality and a lack of alignment between teams. They are different expressions of the same underlying problem, which is systems built to optimize the parts rather than coordinate the whole. The difficulty lies less with AI itself than with putting it into operation, because most systems still sit outside the moment of action instead of working inside it.
A market that punishes delay
All of this is happening against a volatile global backdrop. Energy markets are tightening, supply chains are shifting and geopolitical risk is redrawing the windows in which a plant can make money. Margins are no longer stable, so decisions matter more and have to be made faster.
Closing the gap between the boardroom and the control room
Closing the gap between executive ambition and operating reality takes more than better models. It takes systems built to work inside the decision loop itself. Two kinds of AI are now ready to help.
Large language models as an interface to plant knowledge
LLMs have spread through the public domain at unprecedented speed, raising real concerns about opaque reasoning, leakage of intellectual property and loss of control over how knowledge is used and interpreted. In a plant, their role is narrower and more grounded. They serve as an interface to industrial knowledge, and that is already delivering value.
They remove a major source of friction. With an LLM interface, operators do not have to hunt across systems, documents and tribal knowledge for answers to basic but critical questions:
- What does it mean when a flare turns green?
- What happened the last time this constraint was hit?
- Where is the SOP?
Instead, they ask and get an answer in context. LLMs can surface relevant knowledge across systems, rebuild fragmented context and shorten the time to a decision, which cuts into the time spent searching for information. That improves productivity and lays a foundation for better decisions in the moment. It does not change how decisions are executed, and that is where most systems fall short.
Operations AI, from insight to action
The second kind of AI is where most efforts break down. It is built on your own data, inside your own environment, and coupled tightly to how your plant runs. Because of that it does not generalize, and it is not meant to.
Better insight does not by itself produce better outcomes. In a plant, decisions are constrained by upstream and downstream units, depend on conditions that keep shifting, and are executed by several roles across a shift. Yet most AI systems still behave as if a single recommendation were enough.
What is emerging instead are systems that work inside the decision loop. They:
- reconcile constraints across units continuously
- adjust as conditions change, rather than hours or days later
- coordinate actions across roles, rather than informing one user
Consider a crude unit pushing throughput to capture margin. At the same time, a downstream unit is close to exceeding its constraint limits, energy costs are spiking and a maintenance issue is reducing flexibility. A traditional system flags each of these separately. An operations-focused system resolves them together and guides the plant toward the best coordinated action in real time.
The aim is coordinated execution, where the plant's operating strategy reconciles every constraint, unit and role in real time and keeps people involved where it matters.
The AFPM audience pointed the same way. In the poll, two answers each drew about 37%: real-time autonomous optimization, and systems that continuously learn. The question is moving on from "what should I do?" to "how do we execute the best decision, consistently, across the plant?"
A reality check
Plants that run with no one in the loop are not a realistic goal, and the industry knows it. The aim is to make operators and engineers more effective. Where AI runs in closed loop, it runs inside limits set by plant engineers. ARC Advisory Group describes this as a human on the loop, where the AI calculates and executes and the operator moves from manual controller to high-level supervisor.
What real progress looks like
The companies making real progress share one trait. They stopped treating AI as a technology initiative and started treating it as a change to their operating model. Digital and operations are not separate workstreams. Success is measured in margin and throughput rather than model accuracy. And each proof point is designed to scale, instead of leading to another pilot. An AI project that cannot scale does not deliver a lasting result.
The next phase
Industrial AI is no longer early, but it is not yet mature. Asked in the poll how they see their own organizations, about 40% of respondents said early stage, about 47% said moderate and about 10% said advanced.
The companies that win the next phase of industrial AI will be the ones with the clearest operating picture and a system that helps every role act on it consistently, shift after shift. In our experience, that gap can be closed.
Imubit works on this problem in refining, chemicals and polymers, cement, and mining and minerals, with over 100 closed-loop controllers deployed worldwide. The Imubit Platform maps plan-to-plant gaps to the recurring decisions behind them, models the causality with learning process models, and takes the critical decisions to autonomous execution, inside limits set by plant engineers, on top of the existing APC and DCS.
