Every Case Your Machine Vision Can't Solve Is a Customer Someone Else Will Take
A vision integrator hit the ceiling of rule-based inspection. How it closed the AI gap on food tray contamination without building the model lifecycle itself.
Your customers grade a new crop, a new animal, a new bake every week, and expect the machine to keep up. Add that capability without building the infrastructure yourself. You keep the result. Governed for the life of the model, delivered with our team.
That expectation does not stay on the production floor. It reaches the integrator's desk first. A vision inspection system integrator serving food and beverage, with a large installed base of traditional machines, found that a growing share of what customers asked for sat beyond what its proprietary system could solve.
Why is classical machine vision hitting a ceiling with food customers?
Because the product stopped holding still. Rule-based systems were built for predictable use cases, and for a long time that covered most inspection work. Food and beverage customers now ask for subtle, inconsistent, context-dependent defects, across products that keep changing, without re-engineering the system for every new SKU.
An integrator with deep expertise in optics, lighting, and hardware could once walk into most quality challenges and come out with a solution that performed. The straightforward cases still work. The hard ones do not, and the hard ones are where the market is moving. Many integrators respond by looking at building AI in house. It looks logical, because the hardware, the optics, and the integration are already theirs. Deep learning is a different discipline, though. Development cycles are long, the specialisation is different, and a model built today risks aging before it ships.
What does it cost when your system can't take the job?
Twice over. The visible cost is the project itself: you assess it honestly, conclude your offering cannot reliably solve it, and an integrator with AI in its stack takes the work. The invisible cost is the relationship, because that customer now knows exactly where your capability ends.
The installed base you built over years of solid traditional work then turns into a liability, because the customers in it keep needing something you cannot provide. The in-house route rarely closes the gap in time. A consulting firm speeds up the learning, but delivery stays slow. What often results is a partial solution that covers a narrow set of cases and needs maintenance the team is not structured to provide. Meanwhile customers keep accumulating unsolved use cases and competitors keep adding AI, by partnership or acquisition.
| Dimension | Rule-based vision | Governed AI vision (Indurion Suite) |
|---|---|---|
| Threshold on variable residue | One fixed cutoff for every tray type | Learns what clean looks like across tray types and cleaning batches |
| New SKU or tray type | Re-engineering | Retrains on new examples |
| Harmless variation, such as water droplets | Often flagged as contamination | Learned as acceptable |
| Ongoing maintenance | Falls on the integrator's own team | Versioned and governed in the Suite |
Why does food residue defeat a rule-based inspection system?
Because neither the tray nor the residue is fixed. After a cleaning cycle, food residue shows up in different forms across tray types and cleaning batches. No single threshold separates clean from contaminated when the product and the contamination both vary, and harmless water droplets can look like the real thing.
That was the use case the integrator took to an end customer in the food industry: reusable food trays of different types, inspected for contamination after cleaning, across several production lines. The Indurion Suite runs a contamination classification model that sorts each tray as good or bad. It learned from a small set of example images, clean trays and contaminated trays from different sources, including trays with water droplets that it learned to accept. It runs inside the integrator's existing machine architecture, so the hardware stays as it was.
Where does the AI layer sit, and who owns what?
Underneath the integrator's own brand. The Indurion Suite supplies the AI decisioning and runs under the integrator's name, while the integrator stays the customer-facing expert. Robovision owns the intelligence and the record of every decision. The integrator owns the line-level PLC and the physical routing of each tray.
The commercial effect reached past the first use case. Work that had been declined or deferred, because the technology to solve it did not exist in the integrator's offering, became winnable. The installed base turned into a re-engagement opportunity: machines sold on traditional vision can be upgraded, retrofitted, and extended with AI inspection without replacing the underlying hardware. The integrator's leadership frames it as mastering one thing and doing it well, and partnering for the rest.
What does this mean for integrators weighing the AI build?
That the part worth sourcing is the model lifecycle. Your optics, lighting, hardware, and integration skills are the product. Governing, retraining, versioning, and maintaining models across every customer and every SKU is a separate discipline, and it is the one that quietly consumes teams who try to carry it themselves.
Sourcing it does not shrink the integrator's role. The integrator still designs, installs, and supports every line. The integrators winning in AI-enabled inspection recognised that early, chose where to source the lifecycle, and reached the market before the gap in their offering became a gap in their customer relationships. The question is where you are on that timeline, and how many unsolved cases your customers have already carried elsewhere. You keep the result. We run the lifecycle for you.
See how governed vision AI runs in food production
Codified grading, continuous retraining and an audit trail, on the line you already run.
