Steel producers are under constant pressure to improve efficiency, protect product quality, and respond faster to changing market demands. Many already operate at a high technical level, with experienced engineers and internally developed analytics supporting day-to-day decisions. The challenge isn’t a lack of data or expertise - it’s finding ways to scale that expertise without adding operational burden.
Recently, we worked with a global producer of premium steel products that took a disciplined, focused approach to evaluating Fero Labs’ AI software. Instead of launching a broad transformation initiative or testing across multiple use cases, they defined a single, high-impact problem and used it as a proof-of-concept to answer one question: can AI reliably take over work that consumes significant engineering time today, without sacrificing accuracy?
Start with one problem that everyone recognizes
The team chose a challenge that will sound familiar to most steel producers: maintaining internally built machine learning models.
Each year, their process engineers spent several weeks retraining and recalibrating models used to monitor and optimize process behavior. Across multiple mills, this added up to hundreds of engineering hours dedicated to model maintenance - time that couldn’t be spent on deeper process analysis, troubleshooting, or improvement work.
The goal of the pilot was intentionally narrow. They weren’t trying to solve every optimization problem at once. They wanted to confirm three things:
- That AI could match and improve the accuracy of the models over time that their engineers already trusted
- That retraining could happen automatically, in real time
- That doing so would meaningfully free engineering bandwidth to focus on strategic initiatives
If those conditions weren’t met, there would be no next step.
Prove accuracy first, then remove the manual work
During the pilot, Fero’s explainable AI was applied only to this use case. The results were clear. The models matched the accuracy of the existing internally maintained models, while eliminating the need for manual retraining and recalibration.
Even without expanding into any additional applications, this alone demonstrated measurable ROI. A recurring workload that had absorbed weeks of engineering time each year could now be handled continuously, without ongoing intervention. Just as importantly, real-time training ensured that model accuracy did not degrade over time - a common issue with manually maintained models.
This was the moment that mattered. The pilot didn’t rely on future promises or hypothetical value. It showed, in production-relevant conditions, that AI could take ownership of a well-defined task and do it reliably.
Let the license phase unlock the real value
Only after that proof was established did the company move into a broader license agreement. At that point, the conversation shifted naturally from “does this work?” to “where else could this help?”
With the model maintenance burden removed, engineers could now apply Fero Labs software to problems they had long wanted to explore more deeply but hadn’t had the time to prioritize. These included minimizing alloy usage while maintaining required mechanical properties, identifying upstream contributors to physical defects, predicting properties such as elongation, tensile strength, and yield in multiple mill types, and understanding how chemistry, dimensions, and rolling conditions interact to influence final product performance.
What made this expansion successful was that it didn’t require a new system or a new workflow. The same explainable AI foundation used in the pilot was extended to additional questions, allowing engineers to move faster without changing how they worked day to day.
Fit into the plant, not the other way around
A key requirement from the start was that insights had to reach operators and engineers without disrupting existing workflows. In this case, that meant working directly within tools like Excel rather than introducing a separate interface.
By integrating AI outputs into familiar environments, the teams were able to start using insights immediately. Adoption wasn’t a change-management exercise, it was a natural extension of how decisions were already being made on the plant floor.
A repeatable model for proving AI value in steel
What stands out about this experience is how little was overcomplicated. The pilot wasn’t broad. It wasn’t abstract. It focused on one expensive, time-consuming task and answered a simple question with real data.
For other steel producers and consulting partners evaluating AI, the lesson is straightforward: start with a problem that engineers already feel, prove value quickly, and let the results justify broader adoption. When done this way, AI doesn’t compete with engineering expertise - it gives that expertise room to scale.
It also solved a problem that we incur often - which is that process engineers hardly have bandwidth to focus on new areas of improvement unless you can reduce their time spent on current activities. This proof of concept approach did exactly that. Relieving hundreds of engineering hours per year that are able to be reassigned to more valuable initiatives.
As the steel industry continues to evolve, producers who can adapt quickly, protect product quality, and operate efficiently will set the pace. The path to getting there doesn’t start with transformation programs. It starts with one well-chosen pilot that proves its worth.
If you're ready to run a proof of concept pilot - get started here!
