Most business cases for AI in industrial operations are honest on the benefits and dishonest on the costs — not out of bad faith, but because the costs that don't show up in the pilot are the hardest to estimate in advance and the easiest to leave until later. Anomaly detection pilot on a production line delivers a 30% reduction in unplanned downtime within six months. This number goes on the board's slide. What doesn't go on the slide is the cost of integrating the model with the MES, of clearing the three years of maintenance history into text-free fields to create a training base, of retraining the model when the production process changes, of convincing the operations team to trust the system's alert rather than the senior operator's intuition. These costs exist, they are predictable, and an honest business case includes them.
What the pilot measures and what he does not measure
The industrial AI pilot was designed to answer a specific question: does the technology work in the company's technical environment? That's the right question for validating technical feasibility, but it's the wrong question for projecting scale ROI. The pilot operates under controlled conditions—carefully prepared training data set, vendor engineers available for real-time adjustment, scope of equipment or processes selected as most conducive to demonstration. The pilot result measures the potential of the technology under optimal conditions, not the expected result in normal operation with an in-house team.
The discount coefficient between the pilot result and the scaled result varies by project, but is rarely zero. In computer vision projects for quality inspection, it is common for the model to reach 98% accuracy in the pilot with the supplier's image set, and reach 80% in the first weeks of real operation because the lighting conditions, product variation and camera angle in the real environment are different from those controlled. This deviation is not a technical failure — it is the gap between test condition and production condition that any ML model faces. The business case needs to include the cost of the adjustment period and the expected outcome after that period, not the pilot's peak performance.
Costs that consultants do not include on the slide
Integration with existing systems is often the most expensive project item and the most underestimated. An AI model that runs on an isolated server and produces alerts on its own dashboard is not integration — it is just another system for the operator to check. For the anomaly detection model to generate operational value, it needs to create work orders in the ERP, appear in the maintenance management system, or pause the line via integration with the PLC. Each of these integrations has development cost, validation cost, and API license cost that is rarely in the initial budget.
Data: Most industrial AI projects underestimate the cost of data preparation by a factor of two to five. The argument the supplier presents is that the company already has the data — the SCADA system has been recording everything for years. What the argument leaves out is that this data is in the control system's proprietary format, with gaps for periods of maintenance or system failure, with inconsistent timestamps, with incorrect sensor values that were never detected because no systematic analysis was done on this history. Cleaning, structuring and labeling this history to create a training base is data engineering work that requires someone who understands both data and the industrial process — a rare and expensive profile.
Continuous retraining: ML models degrade over time when data distribution changes. In industrial operations, distribution changes because processes change—new products, new raw material suppliers, gradual wear and tear of equipment, process modifications. The model that worked well in the year of implementation begins to generate more false positives or false negatives a year later. The cost of regular retraining — collecting new labeled data, re-running the training pipeline, validating the new model — needs to be in the project's TCO.
Change management: the cost of convincing the operational team to trust and use the system is perhaps the most difficult to quantify and the most critical to return. An operator with twenty years of experience on a production line will not change his behavior because of a dashboard alert that he does not understand how it was generated. Projects that don't include structured training, follow-up time to build trust in the system, and an operator-to-model feedback process reach scale with a tool that no one uses. The ROI of a system that is not used is zero, regardless of the accuracy of the model.
The structure of an honest ROI calculation
There are two sides to industrial AI ROI. The quantifiable benefit includes reduced unplanned downtime costs — average downtime cost multiplied by the number of events avoided per year —, reduced scrap or rework, reduced energy consumption when applicable, and reduced spare parts inventory due to greater predictability of replacement. Each of these benefits needs to be estimated with a confidence interval, not as a specific number, because they all depend on assumptions about the accuracy of the model and the behavior of the operational team that have real uncertainty.
The cost includes software and platform license, edge hardware and additional sensors if necessary, integration with existing systems, preparation and structuring of historical data, periodic model retraining, training and change management with the operational team, and ongoing maintenance of the solution. In projects with an external supplier, it also includes the cost of post-implementation engagement for the first two years, the period in which most significant adjustments to the model take place.
The realistic payback for most industrial AI projects — when all costs are included — is between 18 months and 4 years. Projects that project payback below 12 months are often underestimating integration and change management costs or overestimating benefits based on pilot results. This is not a statement against investment — it is a calibration of expectations that prevents the project from being canceled for not delivering the promised return within an unrealistic time frame.
What a Responsible Project Approval Requires
An industrial AI proposal that reaches executive approval should include: description of the complete cost model by phase — pilot, scale, ongoing operation; the integration plan with existing systems including cost estimate; the data plane including the gap between what exists and what the model needs; the change management plan with adoption metrics; and the retraining and maintenance model with estimated annual cost.
The decision whether or not to approve the project should be made with these elements available, not afterwards. If the project only makes financial sense when integration and data costs are underestimated, the project does not make financial sense. Industrial AI projects that deliver consistent returns are those in which the business case was conservative in benefits and complete in costs — because when reality converges with the projection, confidence in the program is sustained and expansion is approved. Projects with optimistic business cases cancel in the second phase because the first return did not arrive when promised, and the organization loses appetite to continue.
Also read
- How much does Claude Code cost and when is it really worth it
- Biomanufacturing: when biology becomes a production line
- Inference chips and ASIC: when the specialized beats the generic
- Specialized chips and the end of the generic CPU era
- Neuromorphic and bio-inspired computing: when the chip starts to learn from the brain
- AI Energy and Sovereignty: What Governments Need to Plan Now