After three years of conference keynotes promising that AI will transform every industrial process, the actual returns are settling into a clearer pattern. A small set of applications produces measurable, durable value within twelve to twenty-four months of deployment. A much larger set produces talking points, pilot projects, and PowerPoint slides. The difference between the two sets is not a function of how interesting the technology is. It is a function of how well the use case fits the structure of the underlying problem.
What is worth understanding, for an operator allocating capital and attention against AI initiatives, is which categories belong in which bucket. The short list is genuinely short. The skepticism list is genuinely long. Most companies are currently funding the wrong proportion of the two, in part because the vendors are loudest about the categories that work least well in practice. The signal-to-noise ratio in the market is poor, and the cost of getting the allocation wrong is significant.
This piece sorts the use cases by where they sit on the curve of demonstrated, repeatable, near-term return. Nothing here is theoretical. The pattern matches what has been visible across roughly two hundred deployments the author has observed or worked on across industrial categories from 2022 through 2025.
- Across 200+ industrial AI deployments observed since 2022, roughly five categories account for approximately 80 percent of measured P&L impact.
- The remaining 20+ categories collectively account for the other 20 percent, with median payback exceeding 36 months and high failure rates.
- Industrial companies that concentrate their AI investment in the working categories report 2.4x higher returns on AI spend than companies that spread investment evenly across categories (BCG Industrial AI Survey).
The five that work
Predictive maintenance on rotating equipment. The original AI-in-industry use case, and the one with the most consistent track record. Sensors on pumps, motors, compressors, turbines feed time-series data to models that predict failures three to thirty days before they happen. The economics are durable because the cost of unplanned downtime is high, the cost of unnecessary preventive maintenance is also high, and the AI model is genuinely better than the existing rule-based maintenance schedule. Payback is typically 9 to 18 months. The category is well-developed enough that buying is now the right answer for most operators; building rarely produces enough additional value to justify the additional complexity.
Computer vision for quality inspection. Replacing or augmenting human inspectors on production lines with vision systems that detect defects. Works particularly well in high-volume, repetitive production where the defect categories are well-defined and the inspection task is genuinely visual. The economics work because the false-negative rate of human inspection is consistently higher than people believe, and the productivity gain from automated inspection compounds across high-volume operations. Payback typically 12 to 24 months. Some categories (semiconductor wafer inspection, pharmaceutical packaging, automotive paint defects) are now considered essentially solved problems.
Dynamic pricing in distribution and aftermarket. Models that adjust prices based on inventory position, competitor pricing, customer segment, demand signals, and willingness to pay. Works for distributors, parts businesses, and companies with large catalogs and frequent transactions. The value is in capturing the margin that existing pricing structures leave on the table, and it is substantial when the catalog is large and the transaction frequency is high. Payback typically 6 to 12 months, which is unusually fast. The category is also where the most uplift remains available, because most industrial pricing is still set by judgment and convention rather than by signal.
Demand forecasting at the SKU level. Replacing the conventional supply chain forecasting approach (which is usually one of a handful of statistical methods applied uniformly across the catalog) with models that learn item-specific patterns. The improvement over baseline is typically 15 to 30 percent reduction in forecast error, which translates into either lower inventory at the same service level or higher service level at the same inventory. Both produce material P&L impact in working-capital-intensive businesses. Payback 12 to 18 months. The category requires good data discipline; companies with poor data hygiene struggle.
Customer churn and expansion signals. Models that predict which B2B customers are about to leave and which are ready to expand, scoring accounts in real time and routing them to the right intervention. Works because the signal is in data the company already has (usage patterns, support tickets, billing changes, contact behavior) and the response (account team outreach) is cheap relative to the value of preventing churn or capturing expansion. Payback typically 9 to 15 months. The category requires the action layer to actually act on the signals, which is where many implementations fail.
The longer list to be skeptical of
The categories that get the most attention right now, and that produce the least demonstrable return, are recognizable. The pattern is that they have a strong narrative, a thin track record, and a high failure rate that gets discussed quietly when it gets discussed at all.
Generative design. The promise is that AI can produce engineering designs that humans cannot. The reality across most industrial categories is that the generated designs require so much downstream engineering review and validation that the productivity gain over conventional CAD-driven design is marginal. There are specific categories (topology optimization in aerospace, parametric design in some additive manufacturing applications) where this works. Most industrial design contexts are not those categories.
Autonomous operations. The promise is that AI can run plants, warehouses, or logistics operations without human supervision. The reality is that the long tail of edge cases in industrial operations is much longer than the autonomous systems can handle, and the safety regulations require human oversight in any case. The category does produce value at narrow, well-defined tasks (autonomous material movement in defined zones, robotic pick-and-place in structured environments). It does not produce the wholesale operating model transformation that vendors describe.
"AI strategy" consulting and tooling. The category includes platforms and consulting engagements that promise to "make your company AI-ready" or "build your AI roadmap" without producing a specific deployed application. The value is overwhelmingly in the specific applications, not in the readiness work. Companies that invest in readiness without specific use cases tied to specific operational changes are buying a deliverable, not a result.
Generative AI for knowledge work. The promise is that large language models will dramatically improve productivity of engineers, salespeople, and operators by acting as on-demand assistants. The early data is genuinely mixed. The productivity gains are real but small (5 to 15 percent on specific tasks) and the deployment cost is non-trivial. The category is not nothing, but it is also not the transformation that vendors describe. It belongs in the "interesting, watch carefully, allocate modestly" bucket, not the "concentrate investment" bucket.
The use cases that pay are the ones where the model output triggers a specific operational change. The ones that do not pay are the ones where the model output enters a slide deck.
Why the working set is what it is
The five categories that work share three structural features that the skepticism set does not.
First, the decision the model is improving is one the company was already making, on a clear cadence, with measurable outcomes. Maintenance scheduling, quality inspection, price setting, demand forecasting, account prioritization. These are existing decisions. The AI is replacing or augmenting an existing rule or judgment, not introducing a new decision category. This means the baseline is measurable and the improvement is attributable.
Second, the action triggered by the model is bounded and routinized. Schedule a maintenance task. Reject or accept a part. Set a price. Reorder inventory. Reach out to an account. Each of these is a small action that an operator can take without requiring organizational change. The model produces a recommendation; the existing operational structure consumes it. No coordination cost, no role redesign.
Third, the data needed to train and run the model already exists in the company's operational systems. Sensor data on equipment. Production line images. Transaction records. CRM and usage data. The companies do not have to build new data infrastructure to deploy the model; they have to connect existing data to a model that already knows how to use it.
The skepticism set fails one or more of these tests. The decision being improved is new (or not a decision at all). The action is unbounded (requires organizational change). The data does not exist in usable form. Any one of these gaps adds 12 to 24 months to payback. Two of them push the payback past the point where the program survives a budget cycle.
What the operator should do
Three diagnostics help an operator allocate AI capital well.
First, concentrate. The companies that produce returns are not the ones running fifty pilots. They are the ones running three to five deployments in the categories that pay, with enough investment in change management and integration to actually capture the value. The portfolio approach to AI is one of the most expensive mistakes in the category.
Second, buy before build. The five working categories are mature enough that competent vendors exist. The decision to build a custom model in these categories should be justified by a specific reason (proprietary data, unusual operating context, integration complexity), not by a default preference for in-house. Most companies that build in these categories produce results that are roughly equivalent to what they could have bought, at three to five times the cost and twice the timeline.
Third, demand the action layer in every business case. If a proposed AI investment cannot describe, in specific terms, what operators will do differently the day the model is in production, the investment should not be approved. The most expensive mistake is funding a model that produces signals nobody is required to act on. The second most expensive is funding a strategy initiative that produces no model at all.
The AI that works in industry is the AI that improves a specific, existing, repeated decision, with a specific, bounded action, on data the company already has. The AI that does not work is everything else, and most of the spend in most companies is currently going to the everything else.