Key takeaways
- Build the baseline cost of the problem from your own records and agree unit costs with finance.
- Isolate the share of the problem the solution can actually address before applying any improvement estimate.
- Use conservative, expected and optimistic ranges, and count one-time, recurring and acting costs.
- Run a sensitivity check and design the pilot to measure the assumptions that move the answer most.
Start from the cost of the problem
Credible AI business cases start with what the problem costs today, measured from your own records, not from a vendor's claimed improvement percentage. If you cannot state the current annual cost of unplanned downtime on a line, scrap on a product family or manual inspection labor at a station, you are not ready to estimate a return.
Build the baseline from the bottom up. For downtime, that means stop hours by cause multiplied by the margin lost per hour, plus labor, overtime, expediting and any secondary damage. For scrap, it means units scrapped multiplied by material, processing and disposal cost, plus rework labor. Use at least a year of data if you have it, so seasonal patterns and rare events are represented, and agree the cost-per-hour or cost-per-unit figures with finance before going further.
Estimate the addressable share
No AI system addresses all of a problem. A predictive maintenance model on rotating equipment does nothing for downtime caused by material shortages or changeovers. A vision system catching surface defects does not help with dimensional ones. Go through the baseline and separate the portion the proposed solution could plausibly influence.
Example: suppose a hypothetical line logs 400 hours of unplanned downtime a year. Cause codes show that 160 of those hours come from failures of motors, pumps and gearboxes that develop over time. The rest come from jams, material issues and electrical trips. The addressable baseline for a rotating-equipment monitoring project is 160 hours, not 400.
Apply an improvement range, not a single number
Estimate what share of the addressable problem the solution might remove, and express it as a range: conservative, expected and optimistic. Base the range on pilot results where you have them, on your engineers' judgment about how many events were detectable in advance, and on how reliably people will act on outputs. Avoid adopting a vendor's headline figure; their best customers are not your plant.
Example: continuing the hypothetical line, the team reviews past failures and judges that some showed warning signs and some did not. They set a conservative case of 20 percent of addressable hours avoided, an expected case of 35 percent and an optimistic case of 50 percent. At an agreed cost of $2,000 per downtime hour, chosen only for illustration, that gives:
| Case | Hours avoided | Annual gross benefit |
|---|---|---|
| Conservative | 32 | $64,000 |
| Expected | 56 | $112,000 |
| Optimistic | 80 | $160,000 |
Add other benefit streams separately, such as lower spare parts expediting or reduced overtime, each with its own range and its own source.
Count all the costs
AI costs are often understated because the software subscription is easy to see and everything else is not. Include:
- One-time costs: sensors and hardware, installation, network work, data integration, configuration or model development, validation, training and project management, including your own staff time.
- Recurring costs: software subscriptions, cloud compute and storage, hardware maintenance, model monitoring and retraining, support contracts, and the internal time to review alerts and maintain the system.
- Costs of acting: more inspections or planned repairs triggered by alerts, some of which will find nothing.
Example: the same hypothetical project has one-time costs of $90,000 and recurring costs of $40,000 a year including alert review time.
Calculate payback and net value
Net annual benefit is gross benefit minus recurring costs. Simple payback is one-time cost divided by net annual benefit.
| Case | Net annual benefit | Simple payback |
|---|---|---|
| Conservative | $24,000 | 3.8 years |
| Expected | $72,000 | 1.25 years |
| Optimistic | $120,000 | 0.75 years |
For larger projects, finance will usually want net present value over three to five years using the company's discount rate, plus any capital treatment rules. Simple payback is still useful for a quick conversation on the floor.
Run a sensitivity check
Identify which assumptions move the answer most. Change each one in turn while holding the others steady: cost per downtime hour, addressable share, improvement percentage, recurring cost and adoption. In the hypothetical example, the conservative case barely pays back, which tells you the decision hinges on how many failures are genuinely detectable in advance. That is exactly the question a pilot should answer.
A few rules keep the analysis honest:
- Count a benefit only if it will show up somewhere finance can see, such as more output that can be sold, lower spend, or avoided capital.
- Capacity freed up on a line that is not constrained has limited cash value unless you can sell more or remove a shift.
- Do not count the same benefit twice under two names, for example avoided downtime and increased throughput from the same hours.
- State every assumption and its source in the business case.
Benefits that are hard to price
Some benefits are real but resist a clean dollar figure: fewer safety exposures from emergency repairs, better knowledge retention as experienced people retire, faster onboarding of new technicians, improved customer confidence after fewer quality escapes, or better data that makes the next project cheaper. Do not force a number onto these. List them separately in the business case, describe how you would observe them, and let decision makers weigh them on their own terms.
The same applies to risk reduction. If a single failure of a critical asset could stop a plant for weeks, the expected annual cost may look small while the consequence is severe. In that case, present the scenario plainly, for example the cost of one such event and the estimated likelihood range, rather than burying it in an average. Executives are used to weighing insurance-type decisions; give them the inputs to do so.
Finally, compare against alternatives. The relevant question is often not whether the AI project pays back, but whether it pays back better than spending the same money on a spare motor, an extra technician or a conventional automation fix.
Use the pilot to replace assumptions with data
Treat the first ROI estimate as a hypothesis. Design the pilot so it measures the assumptions with the most leverage, then rerun the calculation with real numbers before deciding to scale. When you do scale, track the same measures for a year after rollout so the business case for the next project rests on your own results.
Related use cases
Maintenance & reliability
Predictive maintenance
Use sensor and maintenance history to spot developing equipment failures early and schedule repairs before they stop the line.
Quality
Visual quality inspection
Cameras and computer vision models inspect parts on the line to catch surface, assembly and dimensional defects consistently.
Operations & production
Process & yield optimization
Models recommend or automatically adjust process setpoints to keep quality on target while improving yield, throughput and energy use.
Operations & production
Energy & utilities optimization
Measure and model energy use by line and utility system, then cut waste through better scheduling, setpoints and load management.