Why agentic AI is the biggest story in manufacturing
Posted: September 22, 2025
Tariffs, complex supply chains, ever-changing market preferences: it’s a challenging time to manufacture consumer products. Let’s be honest, though—when haven’t things been challenging?
When I started working in this sector over 40 years ago, the need to balance resilience, efficiency and responsiveness was just as important as it is today. Manufacturers have been doing this dance for decades.
What is new—and what has the industry buzzing with the kind of excitement I haven’t seen before—is the emergence of agentic AI.
A few years ago, knowledge workers had their world rocked by generative AI’s ability to explore, synthesize and reconfigure information. Now, agentic AI promises a similar moment for those in the workflow-dominated world of manufacturing.
The really exciting thing about agentic AI is that it won’t require manufacturers to completely reconfigure their operations. Agents promise something rare, especially when it comes to digital technology: transformation without disruption. That makes their deployment feel achievable—and pretty darn close.
Our Industrial Life
Get your bi-weekly newsletter sharing fresh perspectives on complicated issues, new technology, and open questions shaping our industrial world.
AI agents and manufacturing workflows
To state the obvious, manufacturing processes and operations are not open to interpretation. They are meticulously optimized, predefined and documented in the form of workflows.
Want to start a production run? There’s a workflow for that. Want to place an order with a supplier? There’s a workflow for that. Want to flag up a quality issue? You get the picture.
As manufacturing has digitized over the last half-century, workflows have been codified, and increasingly executed, in digital systems. That’s unlocked a steady stream of efficiency gains and automation, but those improvements have tended to be limited in scope—higher feedstock usage rates here, fewer packaging snafus there.
Now, with AI agents, there is the potential to not only knit disparate digital systems together, but to largely automate the countless workflows that straddle these systems.
To show you what I mean, let’s take a simple example: asset maintenance in a potato chip factory.
In the following scenario, there are three levels of AI. The first, which many factories will already be using extensively, is assistive in nature. The second is advisory; some factories are starting to get to this level. The third is agentic. Although we haven’t reached this level yet, I hope this example shows how close we are.
Special Agent Crispy in action
The production run begins at 6:00 a.m., with potatoes moving from storage units onto the production line. Right from the beginning, the factory’s assistive AI is running. It’s a predictive analytics tool, a tried-and-true machine learning application that is drawing on historical data to find hidden signals in the noise of real-time data.
At 7:15 a.m., it detects an abnormal vibration pattern in a conveyor gearbox. Historically, this pattern has indicated a bearing on the brink of failure.
The predictive analytics tool generates a risk warning, which calls the AI advisor into action. To identify when a repair could be undertaken without causing any disruption to output, the AI advisor looks at the production schedule in the factory’s MES.
At 9:00 a.m., the production supervisor, Helena, starts her shift. Waiting in her inbox is an email containing the findings and recommendations of the advisor:
“Vibration data (see attached chart) suggests that Conveyor 3 will need a bearing replacement within 48 hours. Recommend scheduling repair work during tonight’s planned downtime.”
Helena strolls down to Conveyor 3. She doesn’t expect to be able to hear anything wrong herself. She just wants to make sure that nothing else—like a small rock wedged in the edge of the conveyor—might be causing the unusual vibrations. Everything looks fine, so she replies to the email from the agent with, “Approved. Proceed.”
An unusual vibration is detected on the conveyor.
Agent Crispy arranges for maintenance...
...while Helena works on her proposal.
After the run is over, maintenance is carried out.
Her command prompts the AI agent (let’s call it Crispy) into action. Crispy initiates a PO for the replacement bearing. It also schedules the maintenance job with a technician. The time Helena would have spent in her office, navigating the MES and ERP, is now spent working on her business improvement proposal, which she believes could save the factory over $1 million a year if implemented.
At the end of her shift, the production supervisor hands over to her colleague, James. She forgets to mention the planned maintenance—but that doesn’t matter, because Crispy has already sent James a summary of its activity.
The production run concludes without issue. A technician arrives at the factory with a detailed work order already in hand. James takes the technician to the gearbox in question.
The technician carries out the suggested bearing replacement. With the gearbox open, he can see it could do with a minor service. He relubricates the gearbox and changes the oil. James emails Crispy to say the planned work, plus the service, has been completed. Crispy updates the enterprise asset management log accordingly.
Overnight, James and his team prepare for the next production run, which is due to start at 6:00 a.m. that morning.
How close are we to AI agents?
It’s easy to imagine how AI agents could be deployed in more acute crises—how they might liaise with suppliers during a major supply chain disruption, or work with retailers and logistics during a recall incident.
The example above is not so dramatic, but that’s precisely what makes it exciting. It’s also plausible: Crispy didn’t completely change how the factory was operating. Instead, the agent was just another layer of technology draped lightly over an existing stack.
Now, it’s important to recognize that almost every factory will have a different tech stack, and for too many factories, that stack still consists solely of Microsoft Excel. AI agents will be of little use to them.
It’s also important to state that, even for digitally mature factories, poorly implemented or technically deficient agents pose risks. Before anything is deployed, serious thought needs to go into things like data quality, transparency, trust, governance, compliance and scalability.
However, for consumer product factories with the right foundations in place—i.e., good data and the right tech stack—AI agents like Crispy are probably only two or three years away.
Personally, I can’t wait. My time working in consumer products manufacturing has convinced me of two things: digitizing production can unlock massive efficiency gains, and the most powerful digital tools of all may be those which immediately enhance, rather than reconfigure, existing operations.
Agentic AI ticks both boxes. Bring it on.
Recommended reading
McKinsey: The real value of a digital and AI transformation in CPG
Hugging Face: Fully autonomous AI agents should not be developed
