Before talking about technology, we must talk about what it relies on. Agentic continuous improvement (agentic CI) first lived on paper and whiteboards: manual rituals, sticky notes, and floor tours recorded by hand. Then came digitalization, which replaced the whiteboard with digital tools. Today, a new chapter is opening: one of continuous improvement boosted by AI, where agentic CI has established itself as the buzzword in industry and manufacturing. Every vendor now promises an agent capable of detecting, deciding, and acting without human intervention.
Behind this promise lies a simple definition: agentic CI is when AI no longer just recommends an action but triggers it itself: detecting a drift, escalating or launching an action, and then capitalizing on the resolution to prevent the same problem from recurring elsewhere. All of this within the framework of an existing operating model.
What exactly does this look like on the ground? An agent that detects a throughput gap on a line before it becomes a production delay and directly proposes the corrective action to the relevant team. An agent that, faced with a recurring failure across several sites, automatically links similar incidents and suggests the solution already applied elsewhere, without waiting for an operator to think of searching for it. Or even an agent that prepares the priority points of a tier meeting before the team even sits down, based on drifts detected overnight.
On paper, it is attractive. But on the ground, what does it really mean?
The Promise vs. Reality
Most AI agents deployed in factories fail: not because of the model, but because of what they are given to manage. An agent is installed to accelerate problem-solving, and six months later, it runs in general indifference, or worse, it generates more confusion than value. The problem is almost never the sophistication of the model. It lies elsewhere, in two specific areas that the market prefers not to look at too closely: standards and field adoption.
No Standards, No Agentic CI
An agent cannot invent an operating model. It can only execute it and deploy it across sites. This is a distinction that seems obvious but is systematically forgotten at the time of deployment.
Take a simple case: an agent tasked with detecting a quality drift on a production line. Without a clear standard defining who should be alerted, at what threshold, and according to what priority, two scenarios invariably repeat. Either the agent over-alerts: it reports every deviation, even minor ones, and the team learns in a few days to ignore its notifications. Or it under-alerts, through algorithmic over-caution, and nothing is reported before the problem becomes visible in a more costly way. In both cases, the agent solved nothing: it simply moved the noise or the silence without ever relying on a shared decision criterion.
Conversely, when a standard already exists (a defined tolerance threshold, a clear escalation rule, a priority hierarchy known to all), the agent becomes a legitimate accelerator. It does not replace the quality team’s judgment; it applies a rule the team itself set more quickly and systematically. The difference between the two scenarios does not depend on the power of the model. It depends on what existed before the model arrived.
This is an inversion that many industrial and manufacturing AI projects have not yet integrated: the agent’s autonomy does not precede standards; it extends them. Deploying an operating model across sites remains a preliminary task: the agent only makes it executable continuously, not designs it in their place.
Field Adoption, Not Model Sophistication
The second blind spot is even more concrete and often more costly: a brilliant agent that no one uses is a silent failure. The true measure of success for an agentic deployment is not the quality of the model’s reasoning, but its degree of integration into the team’s existing routines: tier meetings, escalations, and shift handovers.
Two ways of deploying roughly the same model illustrate the gap well. In the first, agentic CI automatically feeds the actions of a tier meeting before the team even arrives on the floor: points to be addressed are already prioritized, contextualized, and ready to be discussed in the ritual the team already practices every day. In the second, the same agentic CI exists in the form of a separate chatbot that operators must voluntarily consult outside of their habits. The first version integrates. The second version adds on. And everything that adds on, in a field already under time pressure, ends up being ignored.
Adoption is not a pleasant side effect of a good tool; it is the very condition for an agent to have something to observe, and therefore something to improve. An agent without real usage learns nothing because it has no field data to refine its future decisions.
The Market’s Blind Spot
These two pillars do not meet as often as one might think. Many industrial and manufacturing companies have mature OpEx standards: years of Lean work and well-oiled management rituals, yet they still rely on whiteboards, Excel spreadsheets, and manual routines to bring them to life daily. Others have invested in sophisticated digital tools without the underlying standards being sufficiently consolidated or consistent across sites for an AI agent to rely on them without introducing additional variability.
Few organizations truly have both at the same time. It is precisely in this gap that most agentic CI promises fall short: the agent’s intelligence is sold while ignoring the foundations that would allow it to be useful.
What This Changes for a Deployment
This perspective has a direct consequence on how to approach an agentic CI project. The first question to ask is not “what is the best model available?” but rather two more uncomfortable questions:
- Are our standards clear and consistent enough across sites for an agent to execute them without reinventing them at every deviation?
- Are we ready to integrate this agent into the rituals our teams already practice, rather than asking them to adopt a new one?
Answering these two questions honestly often reveals more about the chances of success for a deployment than any technological demonstration.
In Conclusion
Agentic CI is not a race for the most autonomous agent. It is a race for the foundations such as clear standards and real field adoption that give this agent something solid to execute. The organizations that win this race will not be those that chose the most sophisticated model, but those that first consolidated what the agent must rely on to act.