Process manufacturers do not have a shortage of digital capability. Historians, PLCs, SCADA systems, laboratory databases, condition-monitoring tools and increasingly capable analytics are already common. Yet a plant can know pressure, flow, tank level, valve position and motor state continuously while still discovering an expensive physical condition too late.

That distinction helps explain an apparent contradiction in industrial digitalization. McKinsey reported that 72% of process-industry respondents said their digital transformation had stalled before achieving at-scale impact. Deloitte's 2025 survey of 600 manufacturing executives found only 29% using AI or machine learning at the facility or network level, while another 23% were still piloting it. Almost two-thirds, 65%, ranked operational risk as a first- or second-priority concern when implementing smart-manufacturing initiatives. [1,2]

At the same time, focused sensing projects can create large operating gains. In one 2026 beverage application reported by Processing, more than 160 CIP cycles were evaluated over six weeks. The plant identified average water savings of 1,586 liters per cycle, approximately 4 million liters per year, while shortening CIP by about 10%. A separate Processing article described an average 25% reduction in process-cycle time from real-time spectral sensing and AI, with individual supplier-reported operating opportunities ranging from about $100,000 to $2 million. [3,4]

The conclusion is not that industrial AI lacks value. The harder problem is deployability: reproducing value without turning every use case into another integration program, hardware project and specialist workflow.

Deployment architecture determines whether a good use case can scale

Three common deployment patterns each solve part of the problem. Data-first software can move quickly against historian, SCADA, laboratory and sensor data. Its hard limit is observability: software cannot reliably create a physical state that the plant never measured. Sensor-first point solutions can create that missing signal, but every new use case can become another hardware, calibration and interface project. Transformation-first platforms can unify large operating domains, but broad integration scope, staffing and change management may arrive before the first use case proves its economics.

A brownfield plant usually needs a narrower sequence. Start with one expensive problem. Use the data already there. Determine which physical state remains invisible. Add the minimum sensing required to expose it. Diagnose the condition in the context of the rest of the plant. Keep the operator and established controls authoritative. Verify whether the intervention changed the outcome. Then reuse the same integration for the next problem.

This sequence matters because it keeps the first deployment smaller than the problem while preserving a path to plant-wide reuse. It also prevents a common failure mode: buying more data before defining which decision the new data is supposed to improve.

The first hard limit is observability

Consider product loss in a food or beverage process. A flowmeter can show liquid moving. A PLC can show a valve opening. A tank-level signal can show inventory changing. A historian can preserve all three. A laboratory can later characterize a sample. None of those necessarily answers the immediate economic question: what material is moving through this line right now?

If that state is not measured, another analytics layer cannot manufacture the missing physics. The plant needs an additional perception modality. For liquid processes, UV/Vis spectroscopy can expose changes in optical and organic signatures, product transitions and abnormal states. When the missing state is more compositional, near-infrared sensing can add sensitivity to O-H, C-H and N-H combinations associated with water, carbohydrates, proteins and lipids. Quantitative outputs still depend on representative reference samples, matrix conditions and validated models; NIR is not a universal direct-reading chemistry sensor. [5,16]

Two food and beverage examples are enough to show the principle. In starch processing, a calibrated spectral model can distinguish product-rich washout from normal rinse water, making carryover and washout a continuously observable state. In dairy, NIR literature supports models for fat, protein and lactose, which maps to milk-to-water transitions, residual product during CIP, cream or whey carryover and composition drift. [5,6]

The more important lesson is architectural: the sensor should follow the blind spot. If existing data already exposes the first problem, new hardware is unnecessary. If the plant is physically blind, targeted sensing should enter the same context and decision layer rather than becoming a separate island.

Wastewater and data centers expose the same gap in different ways

Industrial wastewater offers a larger set of direct examples. Conventional instrumentation may continuously expose flow, pH, dissolved oxygen, conductivity and tank level while leaving the operator without a continuous view of organic loading, suspended solids, nitrate or the signature of an abnormal influent event. Published work has demonstrated in-situ UV/Vis and NIR models for COD, nitrate and TSS in activated sludge, including detection of induced process disturbances. A 2025 Water Research study separately demonstrated real-time online COD monitoring in an industrial wastewater process using NIR. [7,8]

Continuous optical information can also become part of treatment optimization rather than merely monitoring. A 2024 study of tertiary ozonation used real-time UV/Vis measurements to adjust ozone dosing; the model reported 94.29% dosing-prediction accuracy while measured COD removal varied from 16% to 71%. The point is not that one calibration transfers everywhere. It is that a physical state that once arrived through periodic testing can, with the right validation, become part of an operating loop. [9]

Data-center liquid cooling presents the same observability problem in a different physical system. Temperature, pressure and flow describe heat transport, but they do not fully describe coolant chemistry. Current liquid-cooling guidance identifies pH, conductivity, contamination and water-glycol composition as important to corrosion, fouling, heat-transfer performance and reliability. NIR is relevant where composition itself is ambiguous: a 2025 study using a miniaturized NIR analyzer reported a 0.026% detection limit for trace water in ethylene glycol. In a cooling loop, the useful operating state is not 'spectrum changed.' It is closer to 'coolant mixture is drifting after top-up' or 'unexpected water ingress is developing,' correlated with CDU flow, temperature, pressure and workload context. [10,11]

These applications look different, but they share the same design requirement: use existing instrumentation for context and add a new modality only when the plant lacks the physical signal needed for the decision.

The second blind spot is time

A plant can also have the correct measurement and still lack timely visibility. Grab samples and laboratory testing remain essential for compliance, calibration and ground truth, but they observe a dynamic process at discrete points in time. The standard EPA BOD method, for example, uses a five-day incubation. That result can be critical for validation and reporting; it cannot function as a real-time operating signal. [12]

The cost appears in the decision window between the physical change and the confirmed result. In CIP, end-of-cycle swabs or grab samples can leave plants choosing between chronic overwashing and the downtime of repeating a cleaning step after a failed result. In the 2026 beverage case cited above, continuous liquid analysis detected transitions that the existing conductivity measurement did not resolve. ABB's survey of 3,215 industrial organizations reported median unplanned-downtime costs of approximately $84,681 per hour in food and beverage and $98,222 per hour in water/wastewater. Those are sector-wide downtime figures, not costs caused by laboratory testing, but they illustrate why unnecessary hours of uncertainty matter. [3,13]

The practical architecture is therefore not 'sensor versus lab.' The laboratory should remain the reference layer. Continuous sensing closes the blind interval. Existing plant signals provide context. Analytics determine whether the change matters now.

Context and verification turn a signal into operating value

A spectral change is not automatically an anomaly. During a product transition, it may be expected. During CIP, it may mark the endpoint operators are waiting for. During stable production, the same change may indicate a developing problem. Context determines meaning.

Processing recently summarized this problem directly: AI alone does not fix reliability; context does. Another 2026 Processing article argued that industrial AI's next value step is translating insight into action on the plant floor rather than ending at another dashboard. The same principle applies to process sensing. A useful system has to connect the physical signal to the operating state, route an understandable action, and then measure whether the action improved the result. [14,15]

That creates a simple operating sequence: detect the developing condition; diagnose likely causes; predict consequences where the evidence supports it; route an operator-approved response; verify the result. Verification is especially important because it converts an AI recommendation into measurable plant economics without requiring autonomous control on day one.