Executive read
This week's strongest signal is not a new predictive-maintenance algorithm. It is the continued migration of industrial AI from advice into action: agents inside engineering software, ERP workflows, SCADA/HMI environments and, increasingly, robots operating in physical plants and jobsites. The second signal is less comfortable: the same increase in capability is beginning to compress the specialist effort required to attack operational technology.
1. Caterpillar and FieldAI push physical AI toward real industrial inspection and autonomy
What changed. On September 2, Caterpillar announced a collaboration with FieldAI around physical AI, robotics, autonomy and digital twins for factories and industrial jobsites. The initial applications include autonomous inspections, facility visibility, situational awareness and operational optimization. FieldAI's pitch is robot-agnostic autonomy that can operate in changing, unstructured environments rather than requiring a tightly scripted route or fixed map.
Why it matters. Inspection work is one of the cleanest bridges between AI and maintenance because it converts a task already performed by humans — walk, look, listen, measure, document — into a stream of repeatable observations. If those observations also maintain a continuously updated digital twin, the robot stops being just a mobile camera and starts becoming part of the plant's evidence system. The important caveat is that the announcement is still largely a vendor collaboration statement — no published maintenance ROI numbers exist yet.
Who it affects. Reliability teams, inspection programs, mining, construction, large facilities, utilities, remote operations and any plant where routine rounds consume labor or expose people to hazards.
What to do with it. The value of industrial robots may arrive first through evidence collection, not through replacing skilled trades. A robot performing repeatable thermal, visual, acoustic and spatial rounds can strengthen diagnosis even while humans still make every maintenance decision.
2. Syspro launches Torque: agents that can take approved action inside manufacturing ERP
What changed. Syspro launched Torque on September 2 — a platform that detects operational problems, recommends actions and executes approved actions inside manufacturing and distribution workflows. Syspro emphasizes auditable reasoning, operator control, deterministic workflow steps and visibility into cost before an agent runs. Controlled availability began in August across food and beverage, industrial equipment and fabricated metals, with a public showcase planned for IMTS.
Why it matters. Most enterprise AI has lived in a side window: summarize this report, explain this exception, draft this message. Torque targets the next layer, where the system can actually change the state of the business system. The governance design is more interesting than the AI branding — showing the data used, the rule applied, the reason for the action, and requiring approval for selected decisions is much closer to what industrial AI needs than a generic autonomous-agent model. The performance claims remain vendor claims; controlled availability is not broad independent proof.
Who it affects. Manufacturers using ERP-driven planning, procurement, inventory, customer service and production workflows, and any CMMS/EAM vendor watching the pattern migrate toward maintenance execution.
What to do with it. The useful question is not "Can AI decide?" It is "What decisions can it make, with what authority, using what evidence, and with what rollback/audit path?" Any industrial AI purchase should be evaluated on that question, not on how fluent the model sounds.
3. ADISRA puts machine learning and an expert system directly into HMI/SCADA
What changed. ADISRA SmartView 5.0 became available at the beginning of September with machine learning, a rule-based expert system, MQTT Sparkplug, Unified Namespace support, PLC data-model integration and AI-assisted development inside the HMI/SCADA environment itself.
Why it matters. This is a different architecture from sending plant data to a separate AI application — the analytics sit inside the interface engineers and operators already use. ADISRA pairs pattern recognition with explicit engineering rules: pure anomaly detection tells you something looks strange; rules supply the machine-specific context for why it matters and how urgent it is. This is still primarily vendor material, and the published air-compressor example is illustrative rather than independently validated.
Who it affects. Controls engineers, small and mid-sized manufacturers, machine builders, integrators, and plants experimenting with condition monitoring without wanting a separate enterprise analytics stack.
What to do with it. Industrial AI may work best as a hybrid: statistical systems notice abnormal behavior, encoded engineering knowledge constrains interpretation, and humans decide what the evidence actually means. That is much closer to real troubleshooting than "AI predicts failure."
4. AI-assisted PLC exploit development moves from theory to live hardware
What changed. Forescout Vedere Labs used Anthropic's Claude to port a pre-authentication remote-code-execution exploit from one WAGO PLC model to another, ultimately executing attacker-supplied ARM shellcode on live hardware. The final RCE-development stage took roughly 8.5 hours and about $536 in API usage, required substantial expert steering, and a later AI-assisted attempt to extend the exploit into command-and-control behavior permanently bricked the test PLC.
Why it matters. The headline is not "AI can now hack PLCs by itself" — it cannot, on this evidence. It still needed a skilled researcher, firmware, a working related exploit, reverse-engineering tools, live hardware and repeated correction. The more important point is that AI compressed part of that workflow, and this is exactly the kind of capability where improvement compounds. OT security is not separable from reliability: an unpatched controller with unnecessary exposed services is now both a cyber risk and an availability risk. The research also shows the danger of autonomous technical action without physical-state awareness — the process destroyed the test controller.
Who it affects. OT engineers, controls teams, maintenance managers responsible for legacy PLC fleets, utilities and any plant with weak IT/OT segmentation.
What to do with it. A plant with poor asset records, unknown firmware versions and undocumented network paths is poorly positioned both to deploy AI safely and to defend against AI-assisted attacks. The readiness problem is symmetrical — fixing one side strengthens the other.
Practical tool worth watching: Siemens Eigen Engineering Agent
Siemens' Eigen Engineering Agent works directly inside TIA Portal and can generate and explain PLC code, create HMI content, configure devices, document projects and work against real project context rather than a pasted snippet. Siemens says it has been piloted or deployed across more than 100 companies in 19 countries and claims 2–5x faster task execution and up to 50% engineering-efficiency gains — vendor claims, not independent benchmarks.
What makes it worth tracking is architectural: the agent sees the actual automation project, not merely a prompt. Project access gives it context, but verification, standards, plant knowledge and authority still determine whether its output is useful or dangerous.
Signal vs. Noise
Signal. Industrial AI is moving into systems that can change operational state, not merely generate text — Syspro Torque, Siemens Eigen and Caterpillar/FieldAI all point the same direction from different layers.
Signal. Context is becoming the real product. Project structure, ERP business rules, process data and continuously updated physical-world models give AI something generic assistants lack: bounded operational reality.
Signal. Hybrid systems are winning conceptually — ML plus expert rules, AI plus deterministic workflow, autonomy plus human approval. The market is quietly moving away from the idea that one giant model should reason about everything by itself.
Noise. Efficiency percentages remain mostly vendor-reported. Treat "50% faster" and similar claims as hypotheses until methodology is visible.
Noise. "Autonomous" covers an enormous range, from generating a PLC block for review to physically moving through a mine. The word is nearly useless unless authority and boundaries are specified.
What I'm Watching Next
- IMTS, September 14–19 — a large batch of industrial-AI announcements is likely; the filter is whether vendors show closed-loop, deployed workflows rather than demos.
- Industrial AI Summit, September 9–10 — agentic AI, physical AI, edge AI, digital twins and AI-connected OT security.
- Whether the industry converges on a real framework for "who is allowed to act" — the bottleneck is shifting from intelligence to permission, verification and accountability.
Bottom line
Industrial AI is leaving the dashboard. The valuable systems are beginning to observe real equipment, understand project or process context, recommend actions, and in bounded cases execute them. That makes the technology more useful — and makes verification, authority design, operational memory and cybersecurity substantially more important.
Industrial AI becomes consequential when it gains the ability to change the state of the system. At that point, intelligence is no longer the main design problem. Control is.
Sources
- Caterpillar and FieldAI, September 2, 2026: https://www.caterpillar.com/en/news/corporate-press-releases/h/caterpillar-and-fieldai-advance-ai-powered-industrial-innovation.html
- Syspro launches Torque: https://www.prnewswire.com/news-releases/syspro-launches-torque-the-industrial-ai-platform-that-transforms-manufacturing-decisions-into-auditable-automatic-action-302867052.html
- ADISRA SmartView 5.0: https://www.automation.com/article/adisra-adisra-smartview-5.0-machine-learning-built-hmi/scada
- Forescout, Can AI Create PLC Attacks?: https://www.forescout.com/blog/can-ai-create-plc-attacks-yes-but-it%E2%80%99s-not-that-easy-yet/
- Siemens Eigen Engineering Agent: https://press.siemens.com/global/en/pressrelease/siemens-launches-eigen-engineering-agent-bringing-purpose-built-ai-industrial
Industrial AI Brief is published free, with no signup and no paywall. Written by Ryan McIntosh, Red Seal millwright, thirty years on the tools.