Earning The Right To Close The Loop in Industrial AI
Real Industrial AI doesn't replace your control systems or operators — it earns the right to participate and respects decades of hard-won safety culture.
(Our topic. Our tone. Sponsored by TwinThread *)
There’s a lot of discussion about using AI in industrial environments.
On one side: proponents claiming this technology changes everything. You need to embrace it fast or you risk being out-innovated.
On the other side: the dangers seem so large that even thinking about using it in production means certain doom.
So why not leave it to IT to play with their new tools??
This probably sounds familiar to many. We’ve had exactly the same debates with cloud computing, wireless IIoT sensors, machine learning and before that even Advanced Process Control (APC).
Every new technology comes with unknowns.
And if there’s one thing engineers don’t like, it’s handing control to something they don’t fully understand. (We’re convinced the decision to give SkyNet the launch codes was made by a politician, not a control engineer ;))
Can We Learn From History?
Why is it that these tech savvy engineers aren’t jumping on every new technology?
Is it because they’re just conservative old-timers resistant to change?
We’re not denying that there is a little bit of that with some of them, but at the core the issue comes from the difference between IT and OT, the virtual and the physical.
In IT when your prediction goes sideways, the impact is a wrong product suggestion or some other form of revenue loss. In a plant you’re talking about equipment damage, regulatory issues, injury and even death.
That means that whatever we do in an operational context needs to be explainable. To operators, for them to trust the system, but equally important to regulators (especially after an incident has happened)
For APC, this journey took almost 30 years to move from “interesting idea” to “trusted, closed-loop optimiser.”
Let’s trace that evolution and see what it teaches us.
The Trust Ladder (with APC as an example)
A story about trust…
Phase 1: Read-Only (1970s) Model-based control started when engineers in petrochemical plants ran simulations on mainframes to determine setpoints. These simulations took hours, the idea of real time control was outlandish at the time. Instead, they served as additional information for engineers and operators. The algorithm watched the process. Operators watched the algorithm. There was no interaction with the real world that didn’t involve a human operator. This gave everyone freedom to experiment and validate. What works? What doesn’t?
Phase 2: Recommendation (mid-1980s–early 1990s) Faster computers enabled near-real-time APC running alongside the DCS. The logical next step: APC giving live proposals. “Increase temperature by 5°C. Lower reflux ratio by 2%.” Initially, these “black boxes” weren’t trusted. How could they replace hundreds of years of combined control room experience? Most operators simply overrode the recommendations. But over time, trust grew. The models were actually performing.
Phase 3: Limited Autonomy (1990s) A huge step. Instead of humans making every decision, APC was allowed to directly modify system parameters, but only within narrow bounds and on a limited scope! Soon we discovered that over time models tend to drift off and model maintenance came to the forefront. Engineers learned from this failure that closed-loop requires continuous model care.
Phase 4: Supervised Autonomy (2000s–today) APC now runs 24/7 on thousands of units worldwide. Operators trust it but still have override authority. Some plants run “lights-out” optimisation overnight, with operators reviewing in the morning.
Let’s use this analogy and apply it to where Industrial AI stands today:
Most industrial AI is still in Phase 1—observing, predicting, learning patterns. We’re seeing the first steps into Phase 2, where AI actively recommends actions based on your systems.
Moving to Phase 3 and 4 requires earning trust at each step. Just like APC, AI will need to prove itself before it’s given more autonomy.
This is where Industrial AI platforms have to prove themselves.
In a market prone to overpromising, the platforms that actually fulfill that promise do so by embedding where real work happens on the frontline. The vendors who succeed there and help teams move through the progression of autonomy, offloading unnecessary work and finding new opportunities for improvement, ultimately earn their trust.
TwinThread was founded with that reality in mind. The platform was created by people with decades of experience in the same roles as the teams it serves. They understand that operators and engineers only trust tools that fit into existing systems, follow a responsible acceptance progression, solve real problems, respect operational guardrails, and make daily work easier without adding new risk.
That shows up across the TwinThread platform. Predictive analytics allow teams to observe, anticipate, and understand performance issues earlier. Prescriptive recommendations guide better decisions. And where trust has been earned through adoption, evidence, and guardrails, closed-loop adjustment helps teams standardize execution and continuously improve performance.
These capabilities give teams a legitimate path to scaling this new technology: start with a focused use case, prove value quickly, and expand into deeper integrations as trust grows.
The Example (Hill’s Pet Food)
Hill’s Pet Food is a good example of this in action.
Within Hill’s, TwinThread deployed quality optimization models through edge agents connected to their existing AVEVA (Wonderware) Historian and AVEVA MES. The models predicted end-of-line product quality and recommended startup changes. Rapidly, Hills targeted and hit ideal conditions for each quality parameter in each run.
As adoption grew, TwinThread also captured and digitized onsite operator expertise. They embedded best practices back into both startup and in-run quality work processes. That reinforced trust in the system and eventually helped Hill’s reach closed-loop operations.
The project’s impact was substantial. Hill’s scaled the quality optimization framework across their entire enterprise, improving CpK across several critical quality measures—while significantly reducing material loss.
This happened even as average operator tenure was cut in half. Despite that, with a digitally mature, closed-loop solution, less experienced operators still delivered consistent, on-target performance. It’s a testament to how Industrial AI, when done right, can preserve tribal knowledge, orchestrate optimal production, and drive measurable value at scale.
The question isn’t whether AI will close the loop in industrial operations. It’s how fast it can earn the right.
The Trust Ladder isn’t new. It’s how every control technology has proven itself—from the first PID controller to modern APC. AI is on the same journey.
The only question is: where does your application sit on that ladder?
About TwinThread
TwinThread is the world’s first complete Industrial AI Platform built to amplify human expertise and optimize legacy systems. Through Predictive, Prescriptive, Generative, and Agentic AI, we enable operations teams to anticipate issues, accelerate problem-solving, and innovate on a single line or across global operations. With over 1 million digital twins, nearly 50 million sensors, and 2 million AI models deployed, TwinThread is proven to deliver scalable AI, maximize ROI, minimize downtime and provide continuous value.
Stay Tuned for More!
🙋 Join the ITOT.Academy (new cohort in September) →
📘 Pre-order the IT/OT Handbook (+ claim the bonuses!) →
Subscribe to our podcast and blog to stay updated on the latest trends in Industrial Data, AI, and IT/OT convergence.
(*) At the IT/OT Insider we do value our independence and transparency. So as we look for ways to pay the bills we were looking for ways to work with sponsors without giving up on those principles. This is where the idea of sponsors comes from. Together with a few selected sponsors we’ll explore some topics that we both find interesting in the same way we write our normal articles. Feel free to contact us if you are interested in a partnership as well.
Image to share:
Good point that "this journey took almost 30 years" for APC, but the analogy risks glossing over new architectural constraints. The Problem is AI models demand continuous data plumbing, drift detection, and governance layered into the DCS and edge agents, not just vendor promises of closed-loop nirvana. Treat failures as structural, not human error, and you stop chasing silver bullets.
I think our tech-savvy engineers will continue to carefully manage expectations within the trust ladder to balance operational risks, leadership pressure to advance innovation agenda and new and young engineers who could be naturally more likely to advance to the last ladder. Knowledge repositories driven by AI could also help this journey to thrive when these tech-savvy engineers retire