By Jared Pfeiffer, Principal Client Partner, Augury
Key takeaways:
- Monthly plans can’t keep pace with a floor that changes daily, so scheduling is where agentic AI has the most to offer.
- Predictive maintenance is the lowest-risk place to start. A false alarm costs minutes, and a missed failure costs a shift.
- Set a red light, green light rule. Agents run reversible scheduling changes within set limits, and a person signs off on labels, allergen statements, and regulatory documents.
Before I moved into my current role, I spent years on the plant floor planning and scheduling, managing maintenance and reliability engineering, and eventually running plant management. One of the biggest things that time taught me was that most operations leaders already know their planning is reactive. In fact, according to a recent report, nearly half (44%) describe their operations as reactive. But with the current, rapid expansion of AI, a reactive posture won’t work. However, most leaders still lack the language to distinguish what agentic AI can truly change from what’s just another dashboard.
CPG manufacturers can no longer rely on static planning in an environment defined by constant volatility. The next generation of operations will be built around AI that can continuously sense changes, coordinate decisions, and take action in real time, doing tasks like resequencing and scheduling. But there is a caveat: AI cannot and should not fully sign off on anything with compliance exposure.
The problem with plans past their prime
Static planning in CPG runs on batch cycles, quarterly forecasts, and manual handoffs between planning, procurement, and the plant floor. The plan isn’t wrong the day it’s written, but it is wrong three weeks later when no one has gone back to touch it. It isn’t the tools that cause problems, but the cadence of change. The plan updates monthly, but the floor is different every day.
And the floor is changing remarkably faster than it used to. Demand volatility, ingredient and material shortages, co-packer complexity, SKU proliferation and faster launch timelines aren’t background players in the planning problem. These factors are the actual reasons that static planning is failing now. A forecast built weeks earlier doesn’t account for a breakdown, a priority schedule change or a material issue that shows up the following Tuesday. This means the plant either scrambles or ships late.
That planning gap has a visible cost as well as one that is harder to spot. Automation, or lack thereof, is the cost you can see. Teams report wasting 7.5 days per month on average on tasks they say could be automated. That capacity should be going toward decisions rather than coordination. The less visible cost is that about half of products with labeling, packaging, or documentation errors can be traced back to version confusion or miscommunication.
That aligns with what I’ve seen. Manual version control tends to fail quietly until something wrong ships. It’s rarely one person’s mistake. The failure stems from a handoff that had no system checking it.
Where AI has earned trust
Predictive maintenance is the clearest place where industrial AI is proving itself. Models flag equipment degradation before failure, and AI-assisted root cause analysis is cutting diagnostic time on unplanned downtime.
This is particularly impactful because 40% of organizations cite unplanned production downtime as their biggest manufacturing challenge. Since the implementation of AI for predictive maintenance and downtime reduction, over half of respondents said they can point to the measurable business results because of it, and that’s no accident. A false alarm on predictive maintenance costs someone five minutes to check, while a missed failure costs a full shift. That is why this kind of AI is usually the first a plant trusts, before anything with higher stakes. It’s also the clearest example of the technology sensing changes, which is the safest, most proven ground to build the rest of the industrial AI argument on.
The same pattern shows up in throughput, such as when computer vision and dynamic scheduling adjust for changeovers and material availability, as well as in supply chain resilience, where demand-sensing models react to live signals instead of static buffers. But these are still just advisory point solutions. A human has to see the insight and act on it, which is not a limitation to route around. AI’s accuracy isn’t the real constraint. The real challenge is that plant teams haven’t learned to trust something they can’t fully see inside. Keeping people in the loop is how that trust gets built, but it isn’t a permanent ceiling.
From surfacing insights to acting on them
Agentic AI, however, is in a different category. Predictive AI tells you something. Agentic AI initiates a workflow, coordinates across systems, and closes the loop without a person patching each step together manually. Agentic AI isn’t just a higher rung on the ladder that predictive maintenance already climbed. It is a genuinely new capability for most plants today.
But while an agent detecting a supply disruption should be able to re-sequence production and draft updated documentation, it shouldn’t be the one deciding that documentation is correct and sending it out the door. Food and beverage is a regulated, safety-critical environment. Most CPG operations are not ready to let AI act unilaterally on labeling, allergen statements, or regulatory paperwork, and I don’t think they should be yet. “Acting in real time” shouldn’t include automated approval of anything that has a compliance piece. It requires people to review those actions, approve them, and then the agents can learn from those decisions.
Governance and guardrails
The conversation around AI guardrails shouldn’t be an abstract one. There must be a distinct line between what it can do and what it can’t. Anything with regulatory or compliance risk, like allergen statements or label claims, needs a person to sign off with a clear audit trail of what happened. Things that are lower risk and reversible, like scheduling adjustments within set limits, can run on their own. That red light, green light split is what governance actually needs to say.
This governance mandate also changes what jobs look like. I already made a version of this shift myself, moving from the floor to a more analytical position, and the AI-influenced evolution of work isn’t as simple as current staff adding supervision to their existing jobs. Part of it is a genuinely new role: someone whose whole job is checking and tuning what the AI is doing, while the system learns alongside them.
There is no efficiency without resiliency
Volatility isn’t a temporary condition to plan around. It’s the permanent operating environment, which means the models have to be built for it, not against it, but that’s not a new idea. Most plants already think this way about equipment downtime risk. This is the same mindset, just applied to planning and supply decisions instead of just machines.
The manufacturers who make the shift from static planning to continuous, agent-driven operations with clear boundaries around what acts autonomously and what requires a human will be the ones who set the pace for this next era of the industry. We’ve already watched that gap open up between AI leaders and laggards in predictive maintenance, and I expect to see the same pattern for agents. By this time in two years, the dividing line won’t be analysis. It will be action.
Jared Pfeiffer is Principal Client Partner at Augury, where he helps manufacturers turn machine health insights into prioritized action and measurable results. He brings more than a decade of plant-floor experience, from maintenance planning to leading reliability teams at Basic American Foods and Advanced Technology Services (ATS) and serving as a production site lead at Bayer. Jared is a Certified Maintenance & Reliability Professional (CMRP).
