Dr. Marc Feldmann, senior principal, Data & AI, Alexander Thamm, explains Agentic AI and its potential for the chemical industry in this Chemical Industry Journal exclusive.
Remember Clippy, Microsoft Word’s uninvited paperclip assistant from the ‘90s?
While Clippy faded into history, the human dream of a helpful, proactive virtual assistant has come a long way – thanks to the last decades of progress in deep learning, natural language processing, and computing power.
Enter AI agents: autonomous digital co-workers that pull data from your ERP systems, give business recommendations, discover new drugs, and even collaborate to help you run entire plants. Sounds like science fiction, but real-world prototypes are already being tested – including in the chemical industry.
Agentic AI is increasingly seen as a paradigmatic shift in how firms organise work, approach automation, and make decisions. Unlike earlier, reactive and rigid AI, today’s agents are pro-active, adaptive, and capable of working independently – problem-solving, collaborating, and even controlling tools in the physical world.
AI agents’ growing traction is powered by a recently mainstreamed breakthrough: large language models (LLMs) like GPT, which provide the ‘brain’ for AI agents to ask questions, reason, and give instructions to tools, people, and other agents.
Unsurprisingly, interest is surging: in LangChain’s 2024 report, 78% of surveyed professionals plan to adopt agentic AI soon, and nearly 70% of tech leaders say they have budgeted at least $500,000 annually for it (Tray.io 2024 study).
For chemical leaders, now is the time to explore how this technology might reshape their organisations.
JO.AI—AI Agents on the Factory Floor
Chemical manufacturing demands relentless output, efficient planning, and minimal downtime –all under strict safety constraints. According to Forbes, at chemical giant Celanese, these pressures drove the creation of JO.AI, a platform of specialised AI agents designed to revolutionise plant operations.
Rather than one monolithic tool, JO.AI is a team of digital specialists now deployed across 50 plants.
One agent optimises shift checklists. Another flags anomalies – “Corrosion might be an issue in heat exchanger 3.” A third drafts work orders.
Operators, technicians, and managers interact with JO.AI via natural language chat, just like they would with a human colleague. People-centric design and data governance are key here, as Ibrahim Al Syed, director of digital manufacturing at Celanese, tells Forbes.
JO.AI runs on an industrial data platform integrating over 40 sources, including 2.5 trillion records from sensors, applications, and images. Knowledge graphs help JO.AI ‘understand’ relations between people, equipment, and concepts in the plants. Celanese ensures safety and accountability by keeping a human in the loop. Piloted in one plant, JO.AI now supports over 40 use cases – a quiet revolution in chemical plant management, and a vivid look at what agentic AI can do on the shopfloor.
CACTUS—The New Chemists in the Lab
In chemical R&D, developing new molecules or materials is critical – but often slow, expensive, and manual. Traditional labs rely on fragmented tools and human effort, delaying breakthroughs by months or even years.
That’s changing at Pacific Northwest National Laboratory (PNNL) near Seattle, WA, USA with CACTUS – short for Chemistry Agent Connecting Tool Usage to Science.
Released as an open-source prototype, PNNL data scientists describe CACTUS as something close to a virtual lab co-worker. It predicts molecular properties, prioritises experiments, and continuously updates its experimental setups based on results. It is specifically designed to control lab tools directly.
CACTUS uses large language models (one is Meta’s LLaMA3-8B) to reason in human language, respond via chat interfaces to researchers’ instructions. Scientists can soon communicate with CACTUS as they would with a human colleague – asking for simulations, adjusting variables, or requesting experiment plans.
PNNL Chief Data Scientist Kumar sees CACTUS as a “milestone in the field of cheminformatics”, as CACTUS demonstrates how AI agents reinvent chemical R&D – making labs faster, smarter, and more autonomous. This might signal an era of AI-enabled molecule discovery once AI agents like CACTUS diffuse within the industry.
Towards Agentic Production Planning
A third agentic AI frontier is production planning – a complex, high-pressure task in many chemical firms. With thousands of open orders, tight delivery timelines, and shared production lines, prioritising ‘what’ to produce ‘when’ is a staggering logistical puzzle.
Based on one current industry example, imagine this: A multi-million specialty pharma company pilots a multi-agent system across more than 15 plants.
Traditionally, planners relied on SAP reports, Excel macros, and late-night calls with plant managers. Now, the agentic plant doesn’t wait for human input – it works ahead of time.
This could look as follows: At 7:45 a.m., just as the planner logs in, a notification appears: “Potential delivery delay detected for customer HealthCo due to raw material shortfall for order 4715. Estimated risk of SLA breach: 78%. Recommended action: reprioritise order 4719, which uses shared capacity but available stock. Execute?”
Behind the scenes, an agent team had been working nightshift: one monitored supply chain disruptions, another checked capacity constraints, a third evaluated contractual deadlines, and a fourth simulated the production sequence under various scenarios, flagging the issue for a fifth to raise it to the human planner.
What such cases may foreshadow is a chemical industry where AI agents completely re-shape operational roles – turning human planners into strategic overseers of intelligent, always-on agentic collaborators, and firefighting into pre-emptive problem-solving.
The Road Ahead – Challenges and Potential
Despite its promise, agentic AI faces real hurdles in the chemical sector. Safety, compliance, and trust are paramount – especially in labs and production environments. Many firms also struggle with fragmented data systems, legacy infrastructure, and people’s reluctance to hand over control to software.
Still, the upside is compelling: AI agents could save organisations thousands of days of expert time, speed up decisions, and help transition from reactive to proactive operations. If deployed responsibly – with strong data governance and human oversight—they may redefine how innovation, production, and planning happen across the chemical value chain.
How To Start with Agentic AI
As a chemical leader, where to begin? Prime hunting ground for agents are data-rich, structured tasks that consume a lot of expert time – like shift scheduling, lab work, or order prioritisation. Complex, high-stakes judgement decisions without good data should be approached carefully or left for later stages.
It’s clear that AI agents are no silver bullet. Like we wouldn’t use a hammer to drive in screws, agents are only useful if matched to the right problem. Pioneer companies start with well-defined pilot use cases where quick wins are realistic, and a strong focus on human-machine collaboration. Build on existing systems. Let agents “speak” the language of chemists, plant operators, and planners. And keep humans in the loop—to maintain safety, build trust, and help agents learn.
Evidently, AI agents are not replacements, but complements – the new chemists that never sleep, never forget, and keep getting better – but that only thrive when working alongside human colleagues.
Key References
Case 1: JO.AI
https://www.forbes.com/sites/stevebanker/2025/02/21/celanese-leads-the-pack-when-it-comes-to-agentic-ai/
Case 2: CACTUS
https://pubs.acs.org/doi/10.1021/acsomega.4c08408
https://www.hpcwire.com/2025/01/16/new-ai-agent-connects-computer-reasoning-with-chemistry/
Agentic Production Planning – own professional practice
Reports/Surveys
LangChain report https://www.langchain.com/stateofaiagents
Tray.io Study: https://cloud-computing.tmcnet.com/breaking-news/artic-les/461539-state-ai-agent-development-strategies-the-enterprise.htm
