Transforming Manufacturing Data Into Actions with Agentic AI

Yousef Mohassab, CEO of Facilis.ai and former analytics executive at Koch Industries, spent five years running a major analytics department before concluding that the traditional model, corporate data science teams waiting for problems to come to them, simply doesn't work for manufacturing. His solution? Agentic AI that puts advanced analytics directly in the hands of process engineers and operators who understand what's actually happening on the floor.

‍

The Two Constants Every Manufacturer Shares

‍

After years leading analytics teams and now working with manufacturers across industries, Yusuf identifies two universal truths that never change:

Nobody wants solutions that take days or weeks to deliver insights. When you have a problem right now that's impacting production, waiting for a data scientist to pull data, run models, and report back is too slow. By the time you get the analysis, you've already made the decision—usually based on experience or instinct rather than data.

Nobody wants solutions that require armies of specialized resources. Building and maintaining teams of data scientists, data engineers, and MLOps specialists is expensive, difficult, and ultimately doesn't scale. Even mid-sized manufacturers would need to spend over $3 million annually just on the FTEs—and that doesn't include the tools, infrastructure, or time to value.

These constraints aren't going away. If anything, they're getting more severe as the skills gap widens and competitive pressure increases. The traditional approach of centralizing analytics expertise fundamentally can't meet the speed and scale requirements of modern manufacturing.

‍

What Agentic AI Actually Means for Manufacturing

‍

The term "agentic AI" gets thrown around a lot, but here's a practical definition for manufacturing: human-level automation where complex problems get divided into manageable pieces, specialized software agents tackle each piece, and an orchestrator ensures they work together toward the goal.

This isn't just theoretical. Consider a real deployment at one of Facilis's customers for quality monitoring:

Traditional approach: A quality metric drops below threshold. The system alerts the operator that something went wrong. The operator now faces a multi-dimensional problem—figuring out which of hundreds of upstream process variables changed and caused the quality drop. They manually pull data from different systems, run analysis in Excel or Minitab, and spend hours or days hunting for root cause.

Agentic AI approach: The same quality drop triggers a coordinated response from multiple specialized agents:

• A monitoring agent detects the change in real time and automatically establishes new baselines (online learning, not batch models that drift)
• A domain SME agent understands the specific chemistry or physics of the process
• Machine learning agents identify correlations and optimal ranges across hundreds of variables
• An optimization agent finds the specific set points to return to target
• A critique agent validates that recommendations don't contradict known physics or SOPs

The result? Within minutes, the operator gets a notification that not only says "quality dropped" but provides specific recommended actions: "Variables X, Y, and Z have changed from their optimal ranges. Recommend adjusting set points to [specific values] based on when performance was good."

The accuracy might be 60-80% depending on the scenario, but that still gets operators dramatically closer to actionable insights in minutes rather than days of troubleshooting.

‍

Why Traditional Machine Learning Models Keep Failing You

‍

Here's an uncomfortable truth about most deployed ML models in manufacturing: they drift and become unreliable because they're trained on historical data assuming the future will look like the past. But manufacturing processes change constantly—equipment upgrades, different operating conditions, new product formulations.

A classical batch-trained model becomes less accurate over time in these dynamic environments. You need data scientists to retrain models regularly, which brings you back to the resource bottleneck.

The agentic approach addresses this through self-adaptive online learning. The monitoring agent doesn't just learn from historical patterns—it continuously learns and establishes new baselines automatically as operational conditions evolve. No intervention required.

But the real breakthrough is combining multiple types of intelligence rather than relying solely on correlational machine learning.

‍

The Critical Innovation: Marrying Physics with Statistics

‍

One of the biggest problems with pure ML approaches is spurious correlations. You can easily find statistical relationships that violate physical laws—like finding that NOx emissions decrease as combustion temperature increases, which contradicts basic chemistry where the rate equation depends exponentially on temperature.

Present that kind of correlation to an SME and you'll lose all credibility. They'll question everything else your models produce.

Facilis's hybrid architecture intentionally combines:

Classical statistical methods and machine learning for pattern detection and correlation analysis

Physics-based agents that understand thermodynamics, combustion chemistry, fluid dynamics, and other domain-specific principles

Explicit domain knowledge from SOPs, engineering manuals, and tribal knowledge embedded in the system

Generative AI where it adds value, with explicit guardrails—but never for critical tasks like writing SQL queries where 95% accuracy means 5% of your operators get wrong data

When a machine learning agent finds a correlation, it gets validated against physics and domain knowledge. If the recommendation contradicts known principles, the system flags it. The result is insights that combine statistical power with engineering validity—the way an experienced engineer thinks through problems by considering multiple angles.

As Yusuf explains: "We marry physics and engineering with statistics and machine learning. That's very intentional. We do not want correlations that are not causation happening in our actionable insights. If we're true to ourselves and want to give people actionable insights, we have to make sure they're close to human-level insight."

‍

The Architecture That Enables Speed: 23 Agents, 15 Industries, Out-of-the-Box

‍

Most AI platforms are either playgrounds for data scientists (drag-and-drop model building) or rigid single-purpose tools (one platform for energy efficiency, another for predictive maintenance, another for optimization). Neither approach scales.

Facilis built a multi-agent framework with 23 specialized agents that understand 15 different industries out of the box. These agents cover:

• Domain SMEs for different processes (combustion, chemical processing, pulp and paper, discrete assembly)
• Classical machine learning specialists
• Optimization agents using deterministic algorithms
• Physics-based simulation agents
• Real-time monitoring with online learning
• Critique and orchestration agents

The platform automatically routes to the right agents based on the problem you're trying to solve. If you're troubleshooting a combustion issue, it engages the combustion specialist agent. If you need optimization, it picks the right optimization algorithm without you specifying which one.

You're not building models—you're describing business outcomes you want to achieve. The agentic system figures out which combination of techniques and domain knowledge to apply.

And crucially, customers can customize and build their own multi-agent workflows as needs evolve. It's flexible and expandable, not a fixed set of use cases.

‍

Data Security and Integration Reality

‍

For data leaders, two questions immediately come up: Where does the data go, and how do you integrate with existing systems?

Air-gapped deployment: Facilis runs entirely on-premises or in private cloud within customer firewalls. No external API calls. No data movement outside customer boundaries. The system can run completely disconnected from the internet.

This matters for two reasons. First, it addresses security and compliance concerns for sensitive manufacturing data. Second, it enables deployment on process networks where you might push optimized set points directly to control systems (Aspen Tech, Rockwell, DCS systems).

Standard connectivity: The platform supports all standard industrial connectors (OPC-UA, OPC-DA, etc.) out of the box. For customers with centralized historians, data starts flowing in days. For those without historians or with legacy systems, Facilis's integration partners can configure custom connectors—though this obviously takes longer.

Quanta Pipeline architecture: The underlying data infrastructure is built specifically for real-time AI workloads, not just trend visualization. Unlike traditional historians that can choke on heavy analytics, the Quanta pipeline scales up and down efficiently to handle machine learning computations on streaming data without stressing the system.

This architectural approach means the platform becomes another data consumer and producer in your existing infrastructure rather than a separate analytics silo.

‍

Conclusion

The traditional approach to manufacturing analytics—centralized data science teams building models that take weeks or months to deploy—fundamentally can't meet the speed requirements of modern operations. By the time you get the analysis, the decision has already been made based on instinct.

Agentic AI changes the equation by distributing advanced analytical capability directly to the people who understand the processes. Not by making them data scientists, but by creating specialized agents that combine machine learning, physics-based reasoning, and domain knowledge to deliver actionable insights in minutes.

The technology is ready. Manufacturers are seeing results in weeks with platforms that deploy air-gapped on their infrastructure and work with their existing data systems. The success rate for AI projects could potentially jump from 15% to 60-70% as the barriers of complexity and resource requirements come down.

For data and analytics leaders, the strategic question isn't whether agentic AI will reshape manufacturing analytics—it's whether you'll be among the early adopters who gain competitive advantage or playing catch-up in two years.

The $3 million data science team will always have a place for complex R&D and advanced innovation. But for the hundreds of daily operational decisions that drive quality, throughput, and efficiency? The future is specialized agents working at the speed of the process, not meetings scheduled for next week.

‍