Manufacturing has been running analytics on the shop floor since long before the current wave of AI. SPC charts, control limits, and reliability-centered maintenance are decades old, and plants that use them well already capture much of the value that vendors now promise under a new label. That history is useful context, because it explains why plant managers are skeptical and why the skepticism is often correct.
What has genuinely changed is the cost of working with unstructured and high-frequency data. Vibration traces sampled thousands of times per second, thermal imagery, acoustic signatures, and camera feeds from a line running at 400 parts per minute were previously either discarded or sampled down to a handful of summary statistics. The interesting failure signals lived in what was discarded.
The other thing that has changed is nothing at all: the plant still punishes anything that adds a step for an operator, and any system that cries wolf gets ignored within a fortnight. The framework that survives contact with a real plant is the one everyone on the floor already uses — availability, performance, and quality — so that is how this is organized.
Availability: Predictive Maintenance After the Pilot
Predictive maintenance is the most piloted and least scaled application in manufacturing. The pilot works because it is run on one well-instrumented critical asset with an engineer watching it. Scaling fails for reasons that have little to do with the model.
The first is failure data scarcity. A supervised model needs examples of failures, and a well-maintained pump may have failed twice in eight years. There is no training set. Plants that succeed either use anomaly detection against a healthy-state baseline rather than failure classification, or they use physics-informed approaches where degradation is modeled rather than learned from scratch. Expecting a classifier to learn a failure mode it has seen twice is the most common technical mistake in the field.
The second is that the alert has no owner. A prediction that a gearbox will degrade over the next six weeks is only valuable if it changes a maintenance plan, and maintenance plans are built around planned downtime windows, spare parts lead times, and technician availability. A prediction with a six-week horizon and a twelve-week spare lead time is information, not action. The integration that matters is not sensor-to-model; it is model-to-work-order, with parts availability checked before the alert is raised.
The third is asset selection. Not every asset deserves prediction. The screen is straightforward: does the asset constrain output, is its failure mode gradual rather than instantaneous, is the cost of unplanned failure materially higher than the cost of scheduled intervention, and is it already instrumented or cheaply instrumentable. Assets that fail without warning or that have redundancy do not qualify, regardless of how much the vendor demo emphasizes them.
Where it works, the return is unambiguous, because unplanned downtime on a constraint asset has a known hourly cost and every avoided event is countable.
Performance: Scheduling Against the Real Constraint
Production scheduling is where most plants leak the largest amount of recoverable capacity, and it is usually run in a spreadsheet by one person whose knowledge is undocumented.
The scheduling problems that AI and modern optimization handle well are the ones with expensive sequence dependencies. Changeover time that varies by the pair of products being switched between — light color to dark is cheap, dark to light requires a full clean — is the classic case. So is a heat treat step with a minimum batch size that forces campaign scheduling, or a line where tooling is shared between two work centers that cannot run simultaneously.
A human scheduler solves these with heuristics that are usually good and occasionally very wrong, particularly when demand shifts mid-week. An optimizer solves them consistently and, more usefully, can re-solve in seconds when a machine goes down at 2 a.m. The value is less in the first schedule than in the twentieth reschedule.
The prerequisite is honest routing and standard time data. Plants routinely carry standard times that were set at line commissioning and never updated, cycle times that ignore the operator-dependent step, and routings that omit the rework loop entirely. An optimizer fed those numbers will produce a beautiful schedule that the floor cannot execute, and the floor will go back to the spreadsheet. Auditing standards against actual machine data before scheduling work begins is not optional preparation; it is most of the project.
Throughput prediction is a lighter-weight entry point. A model that forecasts the shift's likely output from the first ninety minutes of run data gives supervisors a chance to intervene while intervention is still possible.
Quality: Where Vision Works and Where It Misleads
Automated visual inspection is the most mature AI application in manufacturing and the one where expectations are most often miscalibrated.
It works extremely well for defects with consistent visual signatures under controlled lighting: surface scratches, missing components, print defects, dimensional checks on rigid parts, weld bead inspection. It reliably outperforms human inspectors on repetitive checks late in a shift, which is when human inspection accuracy declines most.
It struggles where the defect definition is subjective, where lighting or part presentation varies, and where defects are rare enough that a training set cannot be assembled without deliberately producing bad parts. Cosmetic grading on textured surfaces is the perennial disappointment — two experienced inspectors will disagree with each other, and a model trained on their combined labels learns the disagreement.
The practical questions to settle before installing a camera are what the false negative cost is, whether the line can tolerate the false positive rate at the required throughput, and who adjudicates the disputed parts. A system tuned to catch every defect will reject good product at a rate that quality managers will not accept, and a system tuned the other way will let escapes through. That trade-off is a business decision, not a model parameter, and it should be made explicitly.
The underrated quality application is not inspection at all. It is root cause analysis — correlating defect occurrence against process parameters, material lots, ambient conditions, shift, and tooling age. That analysis exists in most plants as a quarterly engineering exercise. Running it continuously across the full parameter space surfaces relationships that nobody thought to chart.
Supplier Quality and Incoming Material
Defects that originate upstream are the most expensive because they are discovered latest. Incoming inspection is typically sample-based and light, and the first real signal comes from a line stoppage.
Two applications are worth building. The first is supplier risk scoring that combines delivery performance, certificate-of-analysis variance, historical defect rates by lot, and change notifications into a signal that adjusts incoming inspection intensity — inspect more from suppliers trending badly, less from stable ones. The second is document extraction on the certificate and test report stream, which arrives as PDFs in a hundred different layouts and is currently keyed by hand or, more often, filed unread.
Reading those documents automatically and checking the values against specification limits catches out-of-spec material before it reaches the floor. That single check has stopped more line stoppages than most predictive maintenance programs.
Safety on the Floor
Safety analytics divides into two categories with very different maturity.
Leading-indicator analysis is well established and undervalued: near-miss reports, observation cards, work order backlogs on guarding, overtime hours by department, and contractor mix are all predictive of incident rates, and the data usually exists in systems that have never been joined. Overtime concentration in particular is one of the most consistent precursors in the data, and it is visible in scheduling records months before anything happens. Plants that already run structured workforce management have the fatigue and coverage data required for this without any new collection.
Computer vision for real-time hazard detection — PPE compliance, restricted zone intrusion, forklift proximity — is technically capable and organizationally sensitive. It works where it is framed as an intervention system rather than a monitoring system, where the output goes to an area alarm rather than to an employee file, and where the workforce is involved in scoping before installation. Deployed without that groundwork it generates grievances faster than it generates safety improvements.
The Data Layer Under All of It
Most manufacturing AI projects that stall do so at the same place: the operational technology data is trapped. Historians with proprietary interfaces, PLCs with no network path, MES data that records what was planned rather than what happened, and quality records kept in a system that shares no keys with production.
Sequence the work accordingly. Establish a reliable path from the machines to a queryable store, reconcile asset and part identifiers across MES, ERP, and maintenance systems, and confirm that downtime reason codes mean what the codebook says they mean. A plant where 40% of downtime is coded "other" cannot support availability modeling regardless of what it spends on models.
What a Realistic First Year Looks Like
Pick one line or one cell, not the plant. Instrument its current performance for a full month before changing anything, including the reason-code distribution and the actual versus standard cycle times. Deploy one application against the largest identified loss — usually either changeover time or a single dominant defect mode. Prove it in production for a quarter, then replicate the pattern rather than the project.
Manufacturers who treat this as business process automation connecting maintenance, quality, scheduling, and procurement generally get further than those who buy four disconnected point solutions, because the loss on one line is rarely caused by only one function. The manufacturing solutions overview covers how those pieces fit together, and a short review of your current downtime and defect data will usually make the first target obvious.
