HOUSTON – The demonstration floor at the AspenTech Optimize conference in early May was bustling with customers, analysts and journalists watching live demos of new technology releases in action. More than 1,500 people, including 600 process-industry customers, attended the conference in Houston.

AspenTech had just announced several new releases, including the introduction of its AI-powered adviser, AVA AI, for the company’s process technology offerings.

Like many other industrial software companies, AspenTech was stirring excitement around the potential of AI-enabled process technologies. To those new to the process industry, AI tools like AVA look like a genuine leap forward, automating tasks that once required an experienced engineer's judgment.

But some experts who work closely with process technologies say the advantages of AVA and similar AI platforms are more incremental and less visible than many vendors are suggesting. For the cautiously optimistic, AI is still limited in the value it can provide.

A critical challenge is that AI advisers often can't explain how they arrived at their conclusions. Operators and engineers will abandon them when they can't get that explanation, said Brian Ashcraft, a central engineering specialist at Dow.

“Everybody who is experienced has experienced that,” Ashcraft told Chemical Processing during Optimize. “If you see something and you can’t explain it, the operator is obligated to turn it off because it’s his job to keep the plant safe and running.”

AspenTech first introduced AVA in 2022 for use within its DMC3 advanced process control (APC) platform. The latest release of AspenTech AVA is an AI platform for the AspenTech portfolio, with advisers available now for Aspen DMC3, Aspen GDOT and Aspen Unified PIMS. The company also introduced AspenTech.ai, a web-based interactive experience where users can engage directly with AVA and ask questions.   

Dow has used AVA on a test basis for a couple DMC3 applications to assess its value with mixed results Ashcraft said.

“The SMEs for DMC3 in the Dow businesses were trying to ask some questions of the live application from AVA and didn’t feel it gave much extra insight as we would see from looking at the web viewer and the results from AspenWatch,” said Ashcraft in a follow-up email. “We haven’t turned it loose for operations or the local support folks (less experience) since the cost is fairly prohibitive versus the value it currently delivers.”

AI tools may be most valuable for operators and newer engineers who want to understand why a unit is behaving a certain way, Ashcraft said. Experienced process engineers can often reach the same conclusions faster without relying on an AI agent, he added.

Existing performance monitoring tools already offer the basic insights a veteran operator needs to make informed decisions, Ashcraft explained. Right now, the advantage of these AI tools lies in their ability to provide answers to process-related questions posed by less-experienced operators, he said.

It’s possible that the value in AI advisers may grow as more experienced engineers leave the chemical sector. Deloitte expects roughly 20% of the current chemical industry workforce to retire by 2030, according to a report by ICIS from Deloitte’s presentation at the American Chemistry Council's annual meeting in June. Ashcraft's point about who needs the help most, newer engineers rather than veterans, suggests AI's biggest test may come once that more experienced cohort is gone.

Why ‘Just Ask AI’ Isn’t Good Enough

Conventional large language models (LLM) are risky in process operations because they are prone to “hallucinations.”  For instance, a probabilistic LLM couldn't reliably explain a controller's behavior, said Heiko Claussen, chief technologist, AI, AspenTech. It would likely generate a plausible-sounding answer, which could lead to serious safety risks in a chemical manufacturing environment.

“You couldn’t just say, ‘Here’s the controller. Please, large language model, explain what it’s doing.’ It’s just making up something you wouldn’t be trusting in this environment,” Claussen explained.

AspenTech designed AVA AI to make recommendations based on first-principles simulation rather than statistical pattern-matching. When a user poses a question, the system pulls information from process simulation environments and returns answers based on established engineering principles rather than a statistical likelihood or probabilistic model.

The system runs what-if scenarios across multiple variables and presents the operator with a ranked set of options, each traceable back to the simulation that produced it, Claussen said.

For example, a demo on AspenTech’s website shows a user asking AVA about a crude unit running below its target feed rate of 25,000 barrels per day. Within AspenTech’s DMC3 web viewer, users can examine the process model and query AVA through a search function in the right-hand corner. Traditionally, an operator would manually scan the model to identify which variables are constraining the feed rate, offering limited visibility into the relationships affecting the target.

AVA captures a snapshot of the controller's current state, runs predictive capabilities on demand and then conducts simulations for this scenario. It iterates on each variable to get as close as possible to 25,000 bpd. AVA ranks the resulting scenarios by how closely each approaches the target.

Out With the Old in With the Old?

Some people say, though, that AI platforms are just small tweaks to well-established process technologies. Vendors are trying to keep their products relevant as more AI-native products enter the market.

“The companies that have been around for 30, 40 years are now seeing pressure,” said Mark Luciw, senior director of strategic services at ARC Advisory Group. “They’re seeing their cheese being eaten by some of these smaller companies that are coming into their space.”

Luciw likened many of the latest AI announcements by major vendors to a fresh coat of paint on an old car.  Much of what's being marketed is predictive analytics and old algorithms with some new terminology sprinkled in, he said.

“The hype around AI has been so intense and so overwhelming that it’s just AI this, AI that, and the truth is that it’s really not AI,” Luciw said.

But APC platforms were already performing many tasks that are now called AI, said Jonathan Alexander, manufacturing AI and advanced analytics manager within Albemarle’s global manufacturing excellence team. The hype around AI has pressured many process technology companies to rebrand their offerings as AI-enabled.

“And so, they have two options: They can refuse to brand their existing stuff as AI, and they lose out on all of the increased capital funding that's going to AI initiatives, or they just rebrand their stuff as AI,” Alexander said.

The chemical sector also faces several challenges related to AI adoption in process control. For one, the industry has many legacy systems and operational silos that create integration challenges, said Luciw. Another critical issue is the impact that AI has on operator proficiency and decision-making.

Among the benefits of an AI adviser is the ability to select an action from several options, said Dow’s Ashcraft.

“It’s not just telling you one answer, but it says to fix the problem, you can do this or this, and according to the costs that you’ve told the application, this is the most profitable way to fix it,” Ashcraft said.

But the same capability that makes AI advisers appealing to less-experienced operators concerns Ashcraft. Allowing an operator to execute a recommendation without understanding why the underlying condition exists in the first place is a risk that experienced engineers take seriously, he said. A heat exchanger restricting production, an instrument beginning to plug, a shift in feed quality — any of these could explain why a unit that has run reliably for months or years is suddenly underperforming, and each carries different implications for what happens downstream if a change is made without investigation.

"Don't just blindly make a change to improve the performance of a software application," Ashcraft said.

It’s a concern that Dave Strobhar, founder and chief human factors engineer at Beville Operator Performance Specialists and Chemical Processing’s Operator Training columnist, has raised in the past. Even without AI, APC came with some downsides for the workforce. Operators became rusty over time as the processes they manage moved to “autopilot,” he said. He noted that an operator once told him APC “makes good operators better and bad operators worse.”

AI systems, like conventional APC, can only act on what they can measure, Strobhar said. Long-time operators know their equipment better than anyone and have developed their own best practices over the years.

Strobhar recalled watching two operators respond very differently to equipment failures on the same type of unit. One, faced with a broken belt on a fin fan overhead cooler, recognized that the APC would respond to the rising temperature by cutting feed rates and manually adjusted the model to hold rates steady while repairs were made. This allowed the operator to restore normal operation within 20 minutes. However, in another scenario, an operator relying on the APC system couldn’t explain why she encountered a feed rate cut, only saying the APC had done it.

"If you're going to use APC or AI, operators need to be part of the crew," Strobhar said. "In many cases, they aren't, and that can present problems."

ExxonMobil's experience suggests that confidence in AI-enabled APC can coexist with that same caution, according to event coverage from Chemical Processing sister publication Control Global. The company is applying digitalization and AI across its Permian Basin operations, drawing on a 30-year partnership with AspenTech as it works toward net-zero operations by 2030-35, said Dylan Pugh, engineering VP at ExxonMobil. The company recently deployed a DMC3 unified architecture with adaptive AI functions across 20 PLCs to optimize production at 100 wells. But Pugh echoed Strobhar's concern about readiness, not capability. “With digitalization and AI, many things are happening fast, but that doesn’t mean they’ll necessarily deliver value,” he said in the Control Global article. “Our OT data must be in good shape if we're going to try and allow more autonomous operations.”

He added that the company may also need to retrain staff for the organizational changes that come with increased autonomy.

Albemarle's $150M Proof Point

The potential, however, is not in question at Albemarle. The specialty chemicals company has deployed thousands of principal component analysis models in real time across five continents, using them to predict production rate problems, equipment failures and quality issues. The company has recorded more than $150 million in annual savings from its AI and analytics efforts, said Alexander.

The underlying methodologies for machine learning have not changed dramatically but the tools for building and deploying models have, he said. AI coding tools, such as Claude Code, now allow engineers to write thousands of lines of code in minutes, compressing work that once took days and dramatically expanding what small teams can deploy.

The results have been transformational for the company, according to Alexander. Albemarle is seeing 5%-20% improvements on OEE and production through better predictive maintenance insights, he said. They’re also reducing variation by monitoring outliers, hidden patterns and correlating them to things like feedstock changes, instrumentation drift, valves sticking or scaling within pipes.

For all the skepticism around AI-enabled APC, Ashcraft sees promise in what AI could deliver for the industry. Specifically, Ashcraft is looking forward to expanded capabilities in the future that provide more in-depth information about why an AI adviser is recommending a specific action.

Right now, Ashcraft said he “can’t visualize the relationship.”

It’s an explainability gap that must be addressed to fully deliver on the promises of AI, said Ashcraft.