The question "how quickly will this pay back?" comes up in every conversation about automating quality inspection — and rightly so. The problem is that an honest answer cannot be given up front. Ready-made calculators with sample numbers show someone else's process, not yours. So instead of promising a specific result, here is the method: which costs to collect, how to put them together, and where assumptions have to be replaced by measurement.
Start with what manual inspection costs today
The starting point is not the AI system but the current state. Most plants know what their inspectors cost, but less often count the other components — and those usually decide the outcome.
- Labour: how many person-hours per shift go into visually inspecting this one product.
- Escaped defects: what a complaint, return, sorting at the customer or contractual penalty costs when a defect leaves the plant.
- False rejects: how many good parts go to re-inspection, rework or scrap only because they were judged too cautiously.
- Inconsistent judgement: whether the result depends on the person, the shift and fatigue — and what resolving disputes over what counts as a defect costs.
- Missing data: what you do not know about your process today because manual inspection leaves no record.
The two middle items are worth listing separately, because they pull in opposite directions. Reducing one kind of error usually increases the other — so you need to know the cost of each before setting a decision threshold.
Then work out what the system really costs
On the investment side there are four groups of costs. Each can be reduced by decisions made before deployment.
- Software licence — quoted individually, depending on scope.
- Hardware: camera, optics, lighting and the station computer. If the existing setup passes qualification, part of this item disappears — see the article on retrofit.
- Deployment and line integration — with us or with your own team. With an in-house deployment you do not pay for integrator hours.
- Maintenance: engineer time for collecting difficult cases and new model versions. Automatic training means you do not need a machine-vision specialist for it.
Put both sides into a simple calculation
The annual benefit is the sum of savings in labour time, in the cost of escaped defects and in the cost of false rejects — minus the annual cost of running the system. The payback period is the one-off investment divided by that net annual benefit. Nothing more is needed: where the numbers you put into the formula come from matters more than the formula itself.
Saved labour time rarely means removing a position. More often the inspector moves to tasks the system will not do: handling REVIEW results, analysing defect causes, working at other stations. Count it as time recovered, not automatically as cost removed.
The most common ROI traps
- Taking detection performance from a demo or someone else's project instead of measuring it on your own data.
- Counting only labour cost and leaving out the cost of escaped defects, which is often the largest item.
- Assuming the system will eliminate both kinds of error at once — in practice you set a balance between them.
- Leaving out maintenance: the product, material and lighting change, and the model needs new cases from time to time.
- Calculating the whole line at once instead of one station where the assumptions can be tested quickly.
Why a pilot is the cheapest way to test your assumptions
Every ROI calculation rests on one uncertain number: how many defects the system will actually find on your product, and how many good parts it will reject along the way. That number cannot be reliably estimated up front — it has to be measured. A pilot covers one Trainer workstation, one inspection station and up to three defect classes for 30 days, with a free extension available, and delivers exactly that data: recall, precision and the false-accept and false-reject rates on an independent test set.
With those results the calculation stops being a forecast. If the numbers do not hold up, an honest NO-GO after a pilot costs far less than a full station that does not meet expectations.