ИИ-агенты в производстве: от диагностики к автономным действиям с человеческим контролем

Kostiantyn Gitko is the CEO of Devox Software, specializing in AI-native transformation and scaling complex systems for global enterprises. A year ago, one of my company’s manufacturing clients wanted better predictive analytics but nothing to do with AI agents. They jokingly called them “ninjas,” software that might quietly go off and do something you never asked it to do. Their concern was model reliability. A bad answer in a chatbot is one thing, but it can stop equipment or create a safety risk on a production line. Eventually, the potential savings were too large to ignore, and they agreed to a narrow test: help operators understand failures faster and cut unplanned downtime, nothing more. When the IoT systems flag a problem—a drive drawing too much current, a press drifting out of spec—the operator and the AI see it at the same moment, and within seconds the AI reviews the machine’s history and context to explain what happened, why and what to do next. Technically, that was already a small form of agentic AI, but the same idea can now go much further—and it got there fast. Those shifts turned fault explanation into an agentic workflow that can close the loop from diagnosis to action. The way I look at it, those changes gave the system five jobs it really couldn’t do before: archivist, Sherlock, partner, dispatcher and twin. Archivist: Remembering Everything Until recently, the system simply couldn’t keep enough of a machine’s history in view at once. Now, it can. Persistent asset context lets the agent see a fault against the machine’s history instead of treating it as an isolated event. The next problem is identity. Different systems often refer to the same machine differently, so an industrial semantic layer can give the agent one consistent asset to work with. The idea predates AI: Manufacturing has been building shared asset models for years through standards such as ISA-95 and OPC Unified Architecture. Now, the agent can use that model to follow the same asset across systems. Sherlock: When Anomalies Become Evidence An anomaly can now trigger a root-cause investigation. The agent gathers the relevant evidence, and the LLM turns it into a likely explanation. Many Models, One Picture That investigation no longer has to rely on one model. Cheaper inference has made specialist AI stacks practical. At GM, the LLM turns the output of those systems into one operational picture. And that’s where factories get messy: The data rarely lines up cleanly. The LLM can connect a mechanic’s description to the corresponding system record and recover what actually happened. The payoff is in-line inspection: At GM’s Factory ZERO, GM uses a proprietary technology to score every weld on every body in real time. And when GM suspected a missed weld at Fort Wayne, the data helped narrow thousands of potentially affected vehicles down to fewer than 10. From Signal To Hypothesis A fault code tells you something went wrong. AI can now go after the harder question—why—and test its explanation against the evidence. Once it can answer why, the next question is what it can safely do about it. Partner: The Right To Take Action The biggest change in 2026 is bounded autonomy: The agent can carry a fault into an approved workflow, while the API remains the execution layer. One way my company has approached this problem in our own work is with an AI harness that keeps the agent inside an operating envelope. What made that possible is a longer horizon: AI models can now hold a task across many steps instead of resolving one prompt and stopping. On the line, a bad weld can divert the body-in-white to rework before it moves downstream. In this case, the human stays in the loop, but moves from doing the work to supervising it. From Sensors To Vision The same thing is happening on the physical side of the line. Poka-yoke used to live largely in hardware, so a model changeover often meant retooling. Computer vision moves some of that logic into software, making the changeover much lighter. At BMW Spartanburg, humanoid robots work in the body shop, feeding sheet-metal parts into the welding process, contributing to the production of more than 30,000 X3s with more than 1,250 hours of operation and 90,000 parts handled. Meanwhile, Agility Robotics noted in an SEC filing that one of their humanoid robots has logged over 65,000 hours in client environments. Dispatcher: Knowing What To Ignore Once context becomes abundant, the bottleneck shifts to attention allocation: deciding what deserves action and what should be ignored. A system that can hold everything in view has to learn what to look away from. This failure mode has a name in production planning: schedule nervousness. In a just-in-time body shop, a crew may already be staging parts to the build sequence when the system suddenly issues a new “optimal” plan. Agents sometimes do the same thing: They pull in more data, second-guess the plan, reach for one more tool and intervene where standing still was the better call. One fix is retrieval-based tool routing: Give the agent only the tools relevant to the current task. One framework, RAG-MCP, lifted tool-selection accuracy from 13.62% to 43.13% simply by retrieving a relevant subset of tools instead of dumping all of them into the prompt.