THE LAST MILLWRIGHT
Industrial AI Brief · Issue 10

September 28, 2026

Maintenance · Reliability · CMMS · Shutdowns · Practical Automation

Ryan McIntosh · Red Seal Journeyman Millwright

Executive read

This week's strongest signal is not another predictive-maintenance model. It is the movement from prediction to workflow to execution, and a shift toward proving value across the whole maintenance system rather than optimizing one dashboard.

1. Maintenance AI finally has concrete operating evidence

What changed. On September 23, McKinsey published "Maintenance meets AI: A proven approach for asset-heavy industries," with results from client work. Reliability agents that search for common failures and bad actors raised one reliability engineer's productivity more than 20-fold in some cases. A US oil and gas company used a gen-AI workflow to update asset hierarchies, build FMEAs from work history and draft PM plans, turning a six-month effort per site into six weeks. A packaging manufacturer raised technician productivity 85 percent with an agentic scheduling tool. At an open-pit mine, an agentic troubleshooting workflow tripled the average time between failures. And a turnaround team risk-ranked more than 500 discretionary items and found close to 35 percent were low-risk candidates for a later turnaround or further review.

Why it matters. McKinsey also warns that aiming AI at one step of the maintenance process often just moves the bottleneck somewhere else. The prize isn't AI that diagnoses bearings. It's removing friction across identifying work, planning, staging, scheduling, execution, spares and contractors.

What to do with it. Stronger evidence than a vendor launch, but not independent research: these are McKinsey client engagements and most clients are unnamed. Read the magnitudes as direction, not benchmarks.

2. Shutdown and turnaround AI gets an Aramco-backed product

What changed. Wipro, Aramco and SAP launched STO360 on September 17. It runs on SAP Business Technology Platform and targets shutdown, turnaround and outage management across oil and gas, chemicals, mining, manufacturing and utilities. Aramco says it intends to use it in its own turnaround and inspection operations.

Why it matters. Shutdown planning has been neglected next to predictive maintenance, yet it is near-perfect territory for agents: scope, dependencies, material readiness, contractor loading, break-in work and schedule recovery. McKinsey puts turnarounds at 20 to 40 percent of overall fleet or plant downtime in its experience.

What to do with it. Still a vendor and partner claim, with no before-and-after results public. The question for any AI shutdown planner is what it knows about the real world: missing material, access conflicts, permit delays, crane conflicts, contractor density, and duration estimates that were fiction before the shutdown began.

3. The industrial-agent chain is becoming recognizable

What changed. Rockwell Automation and Augury announced an agent-to-agent workflow: Augury's Reliability Agent analyzes machine condition and recommends corrective action, and Rockwell's Fiix turns that into prioritized work orders, with FactoryTalk Optix adding production context. The first joint product was expected as soon as September. Separately, Treon launched Treon IQ in early access on September 15: instead of a maintenance manager spending the morning in dashboards, it correlates conditions, diagnoses issues, prioritizes them and recommends actions.

Why it matters. Condition monitoring is becoming an input, not the product. The emerging product is a chain: evidence, diagnosis, priority, planned work, execution, verification.

What to do with it. Both are vendor claims with no public customer outcomes yet. What still wrecks the chain is ordinary: bad asset IDs, poor job plans, parts not on the shelf, false priorities, and recommendations that ignore production reality.

4. Security becomes part of agent design

What changed. Google Cloud published a manufacturing blueprint for secure agentic AI that treats agents as non-human identities with least-privilege access, API limits, monitoring, fallback triggers and isolated pilots, and keeps deterministic physical control separate from cloud intelligence. On September 22, Okta, AWS, Google Cloud, ServiceNow, CrowdStrike, Databricks, Docker, Salesforce, Wiz, Zscaler and others formed the Blueprint Alliance around a shared architecture for agent identity, scoped access, runtime monitoring and containment.

Why it matters. An AI that reads a work order is software. An AI that creates work, orders parts, changes schedules or touches OT interfaces is an actor. Who authorized it, what can it touch, and can it be stopped become maintenance-system questions, not only IT questions.

What to do with it. Think of it as an authority ladder: read, recommend, draft, create a transaction, take a reversible action, affect the physical process. Governance should tighten at every rung.

5. IMTS showed the market getting practical

What changed. The first IMTS Industrial AI Conference, on September 16, was built around choosing high-value problems, data readiness, existing infrastructure, edge versus cloud, and small models versus LLMs. Siemens' Eigen Engineering Agent, working inside TIA Portal on PLC code, documentation and routine engineering tasks, has been piloted with more than 100 companies in 19 countries; Siemens reports two to five times faster execution and up to 50 percent efficiency gains.

What to do with it. The Siemens figures are Siemens-reported. The useful lesson is the design constraint: not an industrial chatbot, but a narrow agent attached to one expensive workflow, grounded in the plant's own information, with a before-and-after number.

Signal vs. Noise

Signal. AI is moving downstream from detection into maintenance execution: shutdowns, scheduling, FMEAs, spares and contractor management.

Signal. Credible architectures combine domain knowledge, operational data, deterministic constraints and human approval.

Noise. Giant adoption percentages and six-month ROI claims with no method behind them, and every analytics feature renamed an "agent."

Noise. Autonomous-maintenance demos that leave out material, permits, access, production constraints and verification. Diagnosing the bearing doesn't put the bearing on the shelf or the millwright at the machine.

What I'm Watching Next

time without producing junk work orders? - STO360 at Aramco. Scope stability, schedule adherence, break-in work, contractor productivity, material readiness and actual turnaround duration. - Agent authority. Do vendors stay at recommendations and work orders, or start asking to take bigger actions? - Evidence-backed troubleshooting. The winning assistant shows why it believes something, points to the telemetry, history or manual, admits uncertainty, and lets the technician say no.

Bottom line

A year or two ago the question was whether AI could detect or explain an industrial problem. Now it is whether AI can carry the evidence all the way through a messy organization until useful work actually happens. That is much harder. The AI industry is discovering that diagnosis is one link in the chain. Maintenance people have known that for decades.

Sources

Industrial AI Brief is published free, with no signup and no paywall. Written by Ryan McIntosh, Red Seal millwright, thirty years on the tools.

Want this looked at in your own plant?

I do efficiency audits and capital project design reviews: where your operation is actually losing money, and the maintenance architecture looked at before the concrete is poured.

ryanmcintosh.ca → All papers