← All articles

Mikhail Polyakov

AI in manufacturing: how to give automation more independence step by step and where to keep a person in charge

Level: BeginnerReading time: 9 min

Topics: AI, Industry, Automation

An industrial pump in a dim plant and a control panel with a rotary knob, around which a narrow permitted range is lit, next to a red stop button

Short answer: on October 8, 2026, MIT Technology Review released a podcast with Arti Garg, chief technologist at AVEVA, a company that makes software for industry. A caveat up front: this is sponsored content. It was produced by the Insights unit, not by the magazine's editorial staff, in partnership with AVEVA itself. So everything below is the view of a vendor's representative, not independent research. Garg's main point: unlike AI in an office, industrial AI acts on real equipment, and an unexpected decision can put people's safety and the reliability of operations at risk. The path she describes: first, AI collects and correlates data and helps a person find the cause of a fault; then it recommends a setting that a person applies; and only then does it adjust the setting itself. According to her, successful pilots of such adjustment have already taken place, and as limits she suggests a narrow permitted range of values or letting the automation work only on part of the equipment. The person, meanwhile, moves from operator to supervisor of the system. For a small manufacturer, I have turned this into a four-step ladder: each step has who does it and how to check it. Below is what Garg says, the ladder and a calculation.

What has changed in industrial AI

According to Garg, industrial AI is not new: AVEVA has been working on it for more than 20 years, including, she says, on its own model for detecting anomalies in equipment operation. What has changed is the type of AI. General-purpose large AI models have appeared that far more people can use, and in the last couple of years so-called physical AI, which controls robots and autonomous machines, and AI agents, which carry out steps in software by themselves, have developed rapidly.

Garg explains how this differs from the office as follows. Industrial AI interacts with real physical systems, often in hazardous environments, and the equipment itself can harm people. In the past, AVEVA chose a model for a task whose behaviour was well understood. The newer models, she says, are by design hard to explain, and their behaviour can change over time as they are further trained. So the main question for her is where a company is willing to entrust such a system with more automation.

Where it all starts: data

For Garg, the first step is collecting and correlating data from different systems, not automation. Her example: readings from the sensors of a pump or mixer, the maintenance log showing when that pump was last repaired, and the original design and maintenance documentation. If all of this is linked, when a fault occurs an operator with a tablet can ask AI to quickly gather the information needed and help understand what is happening. Beyond that, Garg asks us to imagine a robot that walks through a hazardous area by itself and gathers data so that a person does not have to go there. This is a scenario, not a description of a specific deployment.

As a customer example, Garg cites the Thai petrochemical company SCG Chemicals. It first put its data in order, combining operational data with engineering data, and then added anomaly detection on top, in order to spot risks earlier and turn unplanned downtime into planned downtime. According to Garg, the company is aiming for plant reliability of about 99%, and in the early pilot the return on investment (ROI) was almost ninefold. These are figures from a vendor's representative in sponsored content, and it contains no independent verification.

How companies move from advice to automation

The most practical part of the conversation is about setpoints. A setpoint is the target value that the automation maintains: temperature, pressure, speed. Garg describes a solution AVEVA is developing with several customers: the system takes data on how the plant operates and its model and recommends, for example, switching to a different setpoint because conditions in the plant have changed. There is a lot of interest in having the automation change the setpoint by itself, and, according to her, they have had successful pilots in real-world conditions.

But changing setpoints on industrial equipment, Garg admits, makes people very nervous. So, according to her, limits can be kept in place: for example, the automation cannot go outside a certain band of values or only works on part of the plant. She compares this to a manager who gives employees freedom in some decisions and firm boundaries in others. In the industry, she says, people talk about the human moving from operator to supervisor of systems, and how exactly to arrange this, she frankly admits, is still being worked out. AI is not a human, and the limits for it look different.

Two more of Garg's ideas that are useful for a small manufacturer:

  • Test where it is safe. The industry, she says, wants to try new things, but in environments that can be made more isolated and safe for testing.
  • A model that fits the task. A narrower model chosen for the task is, in her view, less likely to behave unpredictably, if its architecture is chosen correctly, and it also uses fewer resources.

And Garg's personal example: she was building a toy robot, and AI helped a lot in making sense of the documentation for its software, but sometimes gave bad advice on where to start troubleshooting. A person who understands where AI is strong and where it is not was still needed.

An independence ladder for a small manufacturer

This is my own application of Garg's ideas to a small manufacturer, not her recommendations. It is worth moving to the next step only after the previous one has run long enough for you to see where AI makes mistakes.

  1. AI collects data and answers questions. Sensor readings, the repair log and equipment manuals are gathered in one place, and AI helps the operator find what is needed. AI has no effect on the equipment at all.
    • Who does it: the chief engineer or process engineer together with whoever keeps the logs.
    • How to check: take five past faults whose cause is known and see whether the operator, with AI's help, finds the information needed and whether AI leads them off track.
  2. AI advises, a person decides. AI proposes a setpoint and explains why. The proposal is first checked by a responsible specialist, and the setpoint on running equipment is changed by an authorized employee according to the procedure approved at your company.
    • Who does it: the operator, with the chief engineer leading the review.
    • How to check: over a month, a record is kept of how many recommendations were accepted, how many were rejected and why. If they are rejected often, it is too early for the next step.
  3. AI changes the setpoint by itself, but within a narrow band. The limits of the band are enforced by the equipment's control system, not by the model itself: if AI requests a value outside them, the system will not accept it. The equipment's emergency protection does not depend on AI and works as before.
    • Who does it: the control system technician; the chief engineer approves the width of the band.
    • How to check: on an isolated test rig or in a computer simulation of the control system that is not connected to the equipment, send the system a value outside the band on behalf of AI and make sure it is not accepted, and that the emergency protection in the simulation triggers the same way as without AI. A simulation only checks the logic reproduced in it, so testing the real emergency protection and any checks on real equipment are left to responsible specialists following the procedure approved at your company.
  4. The band is widened only based on data and after a separate check. A history of operation within the previous band is necessary, but it does not by itself prove that a wider band or other equipment is safe. So the new limits are first assessed by the chief engineer, and the extended mode is checked separately on an isolated test rig or in a simulation before it is applied.
    • Who does it: the chief engineer; the owner approves the decision.
    • How to check: there is a log of all of AI's setpoint changes over the period with the reason for each, none of the changes required emergency intervention by a person, and for the new limits there is a written assessment by the chief engineer and the result of a check on the test rig or in the simulation.

At every step, a person must be able to take back manual control immediately, and everyone knows this. The procedure for taking back control on a specific installation is defined and tested by specialists.

I covered how to set limits for AI agents in office work, with email, documents and accounts, in the article When an AI agent circumvents the rules: three cases from 2026 and six rules for a company.

How much the first step can save

An example with made-up numbers; substitute your own. A plant has 20 faults a month. Currently, for each one the operator spends 40 minutes gathering readings and finding records of past repairs and the right manual: about 13.3 hours a month. If an assistant cuts this to 10 minutes, it comes to about 3.3 hours, a saving of 10 hours a month. At 1,200 rubles an hour, that is roughly 12,000 rubles a month of freed-up time. Reducing search time does not by itself reduce costs: the benefit appears if that time goes to other work. From this sum, you need to subtract the cost of the assistant, connecting and supporting it, the time spent checking its answers, and the work of gathering the data in one place. If gathering the data takes months, calculate payback over a year, not a month.

An even bigger benefit at this step may lie in downtime: if the information needed is found faster, the equipment may be back in operation sooner. Calculate separately what an hour of downtime costs you.

When it is too early

If sensor readings are not stored anywhere and repairs are written in a notebook, it is too early to start with AI. First you will have to put the data in order, and this, according to Garg herself, is the main condition for benefit in industry. And the biggest obstacle, she says, is not even the technology or the data, but the willingness to change work processes themselves to fit what AI can and cannot do.

Summary

In a sponsored MIT Technology Review podcast of October 8, 2026, AVEVA's chief technologist Arti Garg describes the path of industrial AI towards independence: first data and helping a person find causes, then recommendations, then automatic adjustment of setpoints, for which she suggests limits: a narrow band of values or operation only on part of the equipment. In a plant, an AI error affects people and equipment, and the behaviour of newer models is harder to explain, so limits and testing in a safe environment are central for her. A small manufacturer should take the ladder from this: move up the steps only based on data, keep the limits in the control system rather than in the model, and do not make emergency protection depend on AI.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources