Every process plant has a distributed control system (DCS) keeping operations stable. Most have also invested in advanced process control (APC). Yet the gap between what these systems could deliver and what they deliver is often large. In many plants, APC applications that were commissioned with clear benefits now run degraded or are switched off, and the control technology the plant has already paid for returns a fraction of its potential value.
Understanding why means looking at what each automation layer does, where its authority ends and what falls through the gaps. The answer has more to do with how decisions are made at different timescales in a running plant than with the technology itself.
| Layer | Typical timescale | What it does | Where it stops |
|---|---|---|---|
| DCS | Second to second | Holds temperature, pressure, flow and level at setpoint through PID loops | Loops act independently and react after a disturbance |
| APC | Seconds to minutes | Coordinates dozens of interacting variables against constraints | Linear models drift as feed, fouling and catalyst change, and scope is usually one unit |
| RTO | Hourly or slower | Calculates economically optimal targets from process simulation | Slow updates, and it depends on APC to execute |
| Planning | Daily to weekly | Sets production targets and schedules | Built on commissioning-era assumptions |
What the DCS controls and where it stops
The DCS is the core of every process plant. Working second to second, it runs the PID loops that hold temperature, pressure, flow and level at their setpoints. It provides the operator interface, alarm management and continuous data collection that everything else depends on.
Its limits matter just as much. Each PID loop acts on its own. A temperature controller on a column knows nothing about the pressure controller on the same column, let alone what is happening upstream or downstream.
In practice this shows up as a tug-of-war between loops. Pressure, temperature and level controllers each do their job while the unit as a whole drifts away from its economic optimum. Operators often act as the coordinator, deciding which loops take priority during a disturbance.
Modern DCS configurations can include cascade, ratio, feedforward and override strategies. Those help, but they still encode local logic. They rarely represent plant-wide trade-offs such as throughput against energy, quality against recovery, or short-term stability against long-term fouling.
Reactive by design
The DCS is reactive. It responds after a disturbance has moved a variable off setpoint, so process units run with wider operating margins and throughput and yield stay below what the equipment could physically deliver. The largest economic decisions sit above the DCS by design. The control system can hold a setpoint precisely, but it does not decide which setpoint is economically best at that moment.
When feed properties shift, ambient conditions change or a downstream constraint tightens, the control system holds the line. People decide whether to push, back off or trade quality for rate. None of this makes the DCS inadequate at its job. Stable operation and optimal operation are different things, and the distance between them is where margin is lost.
Why APC value erodes after commissioning
Advanced process control, usually implemented as model predictive control, adds what the DCS lacks: multivariable coordination. Instead of managing loops one at a time, APC adjusts dozens of interacting variables together to move the process closer to its most profitable operating point.
In the control room, APC works like a disciplined move planner. At every control interval it predicts how the unit will respond to candidate moves, then selects the moves that respect constraints. That matters because the best operating point is usually at the edge of equipment limits, quality limits, emissions limits and downstream bottlenecks.
The value is real while the application stays healthy. Well-maintained APC can deliver higher throughput, tighter constraint management and more consistent quality. The problem is sustaining it.
APC relies on models identified during commissioning through step testing. As feed quality shifts, equipment fouls and catalyst activity changes, those models drift from the real process. Holding performance takes periodic re-testing and retuning, and without consistent maintenance many APC applications degrade or are switched off within a few years.
When operators stop trusting the controller
APC performance can also be limited by plant conditions that have nothing to do with the controller math. Sticky valves, analyzer drift, instrument noise and changing operating procedures all degrade prediction quality. The controller may still be running, but operators stop trusting it when its moves push into conditions that feel unstable.
APC also typically covers one process unit or a handful of related units. Coordinating optimization across a whole facility, where upstream decisions cascade into downstream constraints, is outside its scope.
Where RTO and planning fall short of real-time coordination
Real-time optimization (RTO) sits above APC in the hierarchy. It uses rigorous process simulation to calculate economically optimal targets, then passes them to APC to execute. In theory, RTO connects business economics to process control.
In practice, that connection has structural limits. Conventional RTO usually updates hourly or less often. Between updates, the process can move through several operating regimes the last run did not anticipate. RTO also depends on APC to execute its targets faithfully, so degraded APC directly limits RTO. And because RTO models need the same kind of maintenance as APC models, the resource constraints that starve APC of attention often affect RTO as well.
Above RTO, planning sets targets from commissioning-era assumptions on weekly or daily horizons. Operations compensates with conservative setpoints, and maintenance defers work that operations needs. Each function optimizes for its own priorities, with its own systems, KPIs and rhythm. The operating strategy ends up fragmented across teams, and the plant leaves margin unrealized even when everyone does their job well.
When quality arrives too late
Quality feedback shows how this plays out. Many plants rely on lab samples or slow analyzers, so the true quality constraint arrives after the process has already moved. Operators compensate with giveaway to avoid off-spec product, and APC respects a constraint it cannot see in time.
Meanwhile, a partly fouled exchanger or a drifting analyzer tightens constraints without triggering an alarm. Operations adapts by widening operating margins instead of working with planning and maintenance on the root cause.
No AI system substitutes for the pattern recognition that comes from decades at the board. But the coordination gaps between automation layers, and between the teams that manage them, hold recoverable value that individual expertise cannot capture alone.
How a learned layer fits above DCS and APC
AI-based process control adds an adaptive layer above the DCS and APC and works through them. It learns from plant historian data and is retrained to stay aligned with current plant behavior as conditions change. Where APC relies on linear models that approximate a narrow operating range, a learned model captures the nonlinear relationships that govern how a unit behaves across different feed rates, ambient conditions and constraint regimes.
This matters when the same move gives different results depending on where the unit is operating. A change that is safe and profitable at one throughput can become unstable near a constraint. Linear models struggle there, and that is where plants want to operate.
Integration usually runs through software gateways into the existing DCS, so existing loops, field devices and safety systems stay in place.
Building trust through advisory mode
Implementations that earn operator trust start in advisory mode. The system recommends setpoints and operators decide whether to act. Over time the question shifts from whether to trust it to when it performs well and when to override it. That is a sounder basis for adoption than forcing closed loop on day one.
Advisory mode also changes how knowledge spreads across a team. Senior operators sanity-check recommendations against decades of context, while newer staff see consistent strategies across shifts. When recommendations are logged alongside outcomes, the plant builds a record of what works, grounded in its own operating data instead of tribal knowledge.
As confidence grows, the plant moves to closed loop, where the system writes setpoints directly, inside limits set by plant engineers. That step depends on clear guardrails: well-defined constraints, rate-of-change limits and a straightforward path to manual takeover. Closed loop is the destination, and advisory mode is where the plant builds the operating discipline to sustain it.
Frequently asked questions
Can a learned optimization layer work with degraded or partly disabled APC?
Yes. It works as an overlay on the existing infrastructure and does not need the APC to be fully functional. Plants with underperforming APC often have the widest gap between current and optimal operation. The model learns directly from plant historian data instead of depending on step-test models that have to be maintained. For more on that difference, see a world without step testing.
Why do automation layers lose coordination as plant conditions change?
Each layer was designed for a different purpose. The DCS maintains stability, APC pushes against constraints, RTO calculates economic targets and planning sets weekly schedules. As equipment fouls, feed quality shifts and operator workarounds accumulate, the assumptions built into each layer drift apart from the real plant. Each function optimizes locally while the plant as a whole drifts from its economic potential. An AI adoption approach that addresses this disconnect can reconnect intent and execution across the hierarchy.
How does a learned optimization layer differ from the RTO already in place?
Conventional RTO uses rigorous process simulation to calculate optimal targets, typically updating hourly. A learned layer models dynamic, nonlinear relationships directly from operating data and moves setpoints minute by minute as conditions change. RTO also depends on APC to execute its targets, so underperforming APC limits it. A learned controller can write setpoints to the APC or directly to the DCS, which closes the execution gap that often keeps RTO targets from being realized. See how AI drives returns in smart manufacturing.
How Imubit applies this
Imubit applies Map, Model and Control across the automation stack in refining, chemicals and polymers, cement, and mining and minerals. Map traces the margin lost between layers to the recurring decisions behind it. Model uses Deep Learning Process Control® (DLPC®) to learn a nonlinear dynamic model of the unit from its own history, guided by first principles. Control takes the critical decisions to closed loop on top of the existing APC and DCS, inside limits set by plant engineers, backed by over 100 closed-loop controllers deployed worldwide. See how the platform works.
