AI in industrial automation is producing real productivity gains and real risk at the same time. This is an honest account of where it helps and where it will confidently hand your team a wrong answer.
Subscribe for industrial automation and manufacturing perspective from active project work.
Learn more at Joltek:
- Edge Computing, AI and Manufacturing Data:
https://www.joltek.com/blog/edge-computing-ai-value-manufacturing-data- Root Causes of Downtime in Industrial Automation:
https://www.joltek.com/blog/root-causes-downtime-industrial-automation- Workforce Development and Education:
https://www.joltek.com/services/service-details-workforce-development-education- Operations and Performance Improvement:
https://www.joltek.com/services/service-details-operations-performance-improvementStart with the economics, because it is the first question a manager asks. A basic subscription runs $20 per month and runs out of capacity within a few hours a day under heavy use. The higher tier is $200 per month. Against a loaded engineering rate that is not a meaningful line item. The real cost is the learning curve, paid up front by the people least able to spare the time. Anyone expecting a gain in week one will stop before getting one.
Where the gains are genuine. Problem definition comes first and is underrated. A systems architect is usually handed a vague requirement already attached to a proposed solution. Putting that in front of a model and asking where the approach fails produces counterpoints worth raising before anybody commits. Second, carrying context across platforms. Experience spread over Rockwell Automation, Siemens, Omron, Mitsubishi Electric, Phoenix Contact PLCnext and Opto 22 cannot be held in working memory, and returning to one after months away is where scaffolding a known function saves hours.
Third, and this is the strongest case, reading code somebody else wrote. A conveyor that will not start, logic written by a firm nobody can reach, is a normal Tuesday. Screenshotting routines into a model and describing the symptom narrows the search dramatically. Equipment arriving with German or Italian rung comments used to cost a day and now costs an hour.
Now the failures, because these determine what you permit. Ask for a whole application in one prompt and you get something unusable. The palletizer example makes the point: infeed handling, staging and orientation, end of arm tooling, and recipe management are separate problems, and the architect decomposes them before prompting. The tool does not do that for you, and it is precisely the skill your senior people have and your junior people do not.
Part number families are where confident wrong answers live. Give it a PowerFlex fault code without naming the exact drive family and it answers for a different one, with no hedging. Firmware revisions compound it, as do instruction sets deprecated years ago that it will still recommend. Structured text is handled noticeably better than ladder logic, which follows from how little advanced ladder material exists publicly.
The organizational conclusion. These tools accelerate a competent engineer and do not replace judgment. On live equipment the person at the keyboard is the last line of defence, because a wrong rung is not a bad draft, it is a hazard to equipment and to people. With 94 percent of executives acknowledging the manufacturing skills gap, and 59 percent of frontline skilled workers over 55 planning to retire within five years according to a 2024 Schneider Electric survey, the pressure to make less experienced engineers productive faster is not going away. These tools help with that, and they also make it easier for someone without the experience to produce confident nonsense. Your review standards have to account for both.
Timestamps
0:00 What this video is and is not
1:10 Which model, and why software engineers converge on one
2:10 Two workflows: a desktop context store and the phone in the field
4:20 What the subscription actually costs
5:20 Where it helps first: defining a vague problem
9:00 The palletizer example: decompose before you prompt
11:50 A real request: an hourly productivity screen
12:50 Carrying context across Rockwell, Siemens, Omron and Opto 22
15:10 Where it breaks: less common platforms and firmware revisions
17:15 Why structured text works better than ladder logic
18:40 Troubleshooting: turning a vague complaint into a direction
22:40 The PowerFlex problem: fault codes differ across drive families
24:40 Reading code you have never seen, in a language you do not speak
27:00 Why you are still the last line of defence
28:30 The learning curve nobody wants to pay
Connect:
LinkedIn: https://www.linkedin.com/in/vladromanov
Joltek:
https://www.joltek.comSolisPLC: https://www.solisplc.com