A different kind of Artificial Intelligence

“Many people associate AI with chatbots of large language models (LLMs) that appear to give different answers every time. That kind of behavior may be fine for writing text, but it is not acceptable for applications like we use in the food industry, such as machine learning," says Sindri Ólafsson.

"Our focus is not on LLMs, but on task specific Artificial Intelligence that allows us to predict and measure performance. We don’t develop such large AI models, but we deploy smaller models that are an integrated part of the software chain that we deploy in the machine. 

Once our AI model is installed, there are no surprises because we know exactly how it behaves, as it has been thoroughly tested for its intended application. This approach helps us deliver higher quality solutions faster and more reliably to our customers.”

Not a guessing game

Sindri Ólafsson continues, “In our world of high productivity, reproducibility is key. If the same product passes the same sensor under the same conditions, the system must give the same answer every time for the same data points. Right or wrong, the outcome must be predictable.

Unlike many LLM applications, which are designed to adapt their responses to the user, our AI model is built for consistency. Our business does not benefit from results that feel like a guessing game. It depends on stable, measurable behavior that customers can trust.” 

AI without surprises

“From the customer’s perspective, AI must not improvise, and our AI model doesn’t. It doesn’t change its mind. It supports the machine by interpreting sensor data in a consistent and measurable way. We develop and provide the intelligence, while the customer benefits from its consistent behavior in daily operations.

We train our AI models to perform specific tasks with clear requirements, using tools that allow us to do this at scale. We carefully track which data is used, how it is labeled, and how the models perform across different test sets. This enables us to report accurately and confidently, without any surprises for our customers, just as we would for any other software component.”

Going into the theoretics, JBT Marel uses primarily convolutional neural networks (CNNs), combined with vision transformers, and also MLP (MultiLayer Perception) neural networks.

Teaching machines 

Traditionally, quality decisions depend heavily on human experience. One operator may judge a situation slightly differently from another. AI helps remove that variation.

Sindri Ólafsson continues, “During the development of a new piece of equipment, we train AI models using tens of thousands of images depicting all sorts of variations. Each image is carefully labeled by experts, so this is still a human task. Bit by bit, the model learns what to look for, just like a new human inspector would learn the profession by studying many examples.

Once trained, the model is frozen and tested extensively. We measure accuracy and reliability and can clearly report results such as finding 99 percent of specific features under defined conditions. Only when performance is proven, we deploy the model in a machine.”

Interpreting images

For example, in the VC-i machine, the images of the vision sensor are processed by a tracking algorithm, which correlates each image to a specific shackle ID. Once linked to the corresponding shackle, the image is sent to an AI model that identifies important features, relevant areas, and determines conditions such as:

• Is a chicken present in the shackle?
• Where is the cloaca?
• Where is the bad area of the chicken?
• How well is it attached in the shackle?
• Is it attached with two legs or one leg?

The AI models themselves do not make decisions; their only purpose is to interpret the image based on pre-trained datasets and provide structured information. During the machine’s development, the AI model is translating the (visual) sensor input into something that the machine software can understand. This information, together with the original image, is then passed to a post-processing system and stored in a database. 

“In addition, we use an extensive system to evaluate the expected quality and accuracy of the AI models. For example, by using a validated dataset of 50,000 images, we can get close to 100% correct cloaca detection in the VC-i,” adds Sindri Ólafsson.

Frozen intelligence means predictable performance

JBT Marel’s AI model is static or ‘frozen’ for the whole lifetime of that model inside the machine. This means it doesn’t learn anything, evolve, or change behavior once it’s installed in the machine and is put into production. That may sound limiting, but it is actually a strength. Frozen models guarantee stable and predictable behavior over time. The machine will perform tomorrow exactly as it did today.

“Before the equipment is installed, JBT Marel and the customer engage in detailed discussions to confirm that the AI model meets all requirements of the customer. All equipment will show expected behavior only. In no case where we deploy the machines does the AI model change behavior.” 

Using AI where it adds real value

JBT Marel does not use AI everywhere, and that is intentional. Some tasks are better solved with physics, calibration, or classical algorithms. For example in the SensorX, X-ray based fat or bone detection relies on physical properties that do not need AI.

When you’re only looking for simple anomalies, a basic algorithm can be enough—for example, thresholding a red line and identifying its highest point. If that’s all you need, an AI model isn’t necessary.

However, if you want to detect more complex or unusual patterns, AI enters the field. Where AI truly shines is in interpretation. It is very good at reading sensor data, recognizing patterns, spotting anomalies, and understanding complex visual information. In those cases, convolutional neural networks are particularly effective, as they can analyze patterns within series of data and extract more meaningful insights.

Less dependent

Artificial intelligence helps machines “see” and interpret visual information better, so they don’t rely as much on human judgment. This means:

• Less need for constant manual supervision
• More stable quality across shifts and sites
• Faster detection of issues before they become problems

Systems like VC-i benefit from AI-based vision technology, and that improved quality trickles down downstream. Even machines that do not contain AI technology directly still profit from the detailed and more consistent input they receive, such as the Nuova-i, that can also perform better now.

Our AI technology increases the security and the quality of the product moving between those machines, so the whole evisceration line is influenced by the AI that the VC-i is providing. Will AI be integrated in every new machine that we develop? Definitely not. Will AI influence all of them? Definitely yes.” 

Built for scale and consistency

One of JBT Marel’s AI strengths is scalability. The same AI platform supports multiple machines and applications across the poultry, meat, fish, and further processing industries. This ensures consistent quality standards and faster innovation across the portfolio. “We can solve more complex tasks faster with better accuracy. And we can do it cheaper because we don’t need to spend so many engineering hours as before.

AI is not just an “us too” initiative for JBT Marel. It is a carefully controlled capability, applied only where it delivers measurable benefits.”

Looking ahead

In the future, not every machine will rely on AI, but more production lines will benefit from it over time. By improving machine performance through AI, JBT Marel helps customers achieve higher quality, greater stability, and less reliance on individual expertise.

In the end, AI is not about replacing people. It is about giving machines better senses, so people can focus on running their operations with confidence.