THE LAST MILLWRIGHT
Industrial AI Brief · Issue 1

July 27, 2026

Maintenance · Reliability · CMMS · Shutdowns · Practical Automation

Ryan McIntosh · Red Seal Journeyman Millwright

Executive read

The consequential shift is not better predictive maintenance. It is the movement from isolated prediction tools toward evidence-grounded agents that assemble work, retrieve plant history, estimate resources, and propose actions inside the maintenance workflow.

The useful opportunity is not another vibration dashboard. It is the connective tissue between ugly plant data and actual decisions:

symptom → evidence → failure hypothesis → job scope → resources → reviewed work order

Everything below is a variation on that sentence.

1. Maintenance AI is moving from chat assistant to governed decision layer

What changed. IBM researchers published a deployed maintenance framework called the Condition Insight Agent, combining free-text work orders, operating behaviour and structured failure knowledge. The important part is not the model. It is that the system creates evidence deterministically and runs a verification loop intended to suppress unsupported conclusions. The reported deployments were built on incomplete, heterogeneous CMMS data rather than clean laboratory datasets, which is the only test that means anything.

Why it matters. Industrial AI cannot safely operate on eloquence. A confidently written hallucination about a pump bearing is still a hallucination wearing safety glasses.

The architecture that is emerging looks like this:

  1. Retrieve asset-specific evidence.
  2. Organize it into a constrained structure.
  3. Compare it with known failure mechanisms.
  4. Generate a recommendation.
  5. Verify that every conclusion is supported.
  6. Leave approval with a human.

That sequence answers the largest single obstacle to adoption. Plants do not trust an untraceable answer, particularly one that can change maintenance scope or production risk.

Who it affects. Reliability engineers, planners, CMMS owners, technical authors, maintenance supervisors, and anyone building an industrial copilot.

What to do with it. If you are evaluating a maintenance AI product, stop asking how good its answers are and start asking what it does when the evidence is thin. Ask the vendor to show you an output where the system declined to conclude. If it has never produced one, it does not have a verification loop, it has a writing style.

2. CMMS agents are beginning to create and modify real maintenance work

What changed. Oracle now documents maintenance agents that create and update work orders, retrieve similar historical jobs and repair procedures, estimate labour, material and equipment costs, and run what-if cost scenarios before execution. IBM's June 25 Maximo AI Service 9.2 release added meta-agent support, pointing toward multiple coordinated agents rather than one assistant bolted onto Maximo.

Why it matters. The agent is crossing from reading the CMMS to writing into it, and that is the real threshold. Summarizing a work order is a convenience. Creating scope, choosing resources, estimating cost and updating records changes planning capacity, and changes risk.

It also exposes a brutal truth. A plant with poor job-plan standards is now able to automate the production of beautifully formatted garbage.

Who it affects. Planners first, then supervisors, storerooms, procurement, schedulers, and whoever owns master-data governance.

What to do with it. Before allowing any agent to write, put a quality gate in front of approval. A work order should not advance unless it has a clear failure symptom, asset and location, isolation requirements, required trades, parts and tooling, access constraints, verification criteria, and something that distinguishes corrective work from exploratory work.

The gate does not decide what the job is. It flags what is missing and generates questions for the requester. That is less dangerous, easier to integrate, and worth more to an ordinary plant than autonomous diagnosis.

3. Shutdown AI is finally targeting scope, not just schedules

What changed. Turnaround vendors are applying generative systems to the front end of planning rather than to schedule visualization. IAMTech's iPlanAI claims it can turn work descriptions, drawings, photographs and engineering documents into proposed steps, trades, crew sizes, labour estimates, durations, materials, costs and confidence scores. Prometheus describes AI-guided event setup drawing on historical turnarounds and ERP revision data. ALICE focuses on generating and stress-testing alternative schedules against time, cost and resource constraints.

These are vendor descriptions, not independent validation. The direction is still consequential.

Why it matters. Shutdown planning has two separate AI opportunities and they are routinely confused. Scope construction asks what work belongs in the event and what each package requires. Schedule optimization asks, given the work and its dependencies, what sequence performs best.

Most shutdown failures happen before optimization is relevant. They begin with incomplete scope, false durations, missing materials, undiscovered access conflicts, and work packages that restate the notification instead of planning the job.

AI schedule optimization cannot rescue fictional input.

Who it affects. Turnaround managers, planners, schedulers, contractors, materials coordinators, operations and engineering.

What to do with it. Before buying an optimizer, audit the scope you would feed it. Pull twenty work packages from the last event and check each for vague descriptions, missing inspection history, absent prerequisites, unconfirmed materials or tooling, likely multi-trade conflicts, crane and scaffold contention, and jobs with no completion test. The proportion that fails is the ceiling on what any scheduling tool can do for you.

4. The stack is going multi-agent, but context is the bottleneck

What changed. A 2026 industrial asset-operations study tested a supervisor-and-specialist architecture using reusable structured artifacts, replanning and parallel tool calls. It reported better task completion and much faster later interactions, because the system reused earlier work instead of rebuilding context every time.

Separate research on persistent agents found reliability degrading across repeated sessions through memory compression, interference, revisions and routine maintenance of the agent system itself. An agent can keep behaving plausibly while factual precision quietly decays.

Why it matters. Industrial work is longitudinal. The useful question is almost never what the manual says. It is: given the last six interventions, the temporary repair made in March, the changed motor, the operator report and the vibration trend, what should we verify next?

That needs durable, revision-aware memory. A chatbot conversation is not an asset history.

Who it affects. Anyone building plant copilots, reliability knowledge systems, digital work instructions, or an AI layer over CMMS data.

What to do with it. Treat every AI output as a controlled artifact rather than chat vapour. Record the evidence snapshot, the assumptions, source identifiers, version date, unresolved questions, the human decision, and what actually happened afterwards. That last field is the one everybody skips, and it is the only one that closes the loop.

5. Security is becoming part of the purchasing decision

What changed. On July 14 the U.S. announced a coordination group connecting major AI developers with providers of essential services to share vulnerabilities and coordinate response, with energy and other critical sectors central to it. IBM has published multiple 2026 Maximo AI Service security notices, including vulnerabilities in dependencies used by the AI component.

Why it matters. Connecting an agent to work orders, manuals, asset hierarchies, inventory and plant systems opens a new path through the operational information stack.

The dramatic risk is an attacker commanding a machine. The ordinary risks are probably worse: poisoned maintenance history, unauthorized work-order creation, retrieval of obsolete procedures, exposure of sensitive plant documents, recommendations changed by a compromised knowledge source, and agents holding more write authority than the humans they serve.

Who it affects. CMMS administrators, IT and OT security, reliability leaders, software suppliers, and any consultant proposing industrial automation.

What to do with it. Make authority boundaries part of the workflow design rather than a later hardening exercise. Read-only first. Source citations on every output. No automatic closure of work. Explicit approval before any CMMS write. Role-based access. Revision-controlled procedures. An audit log of prompts, evidence and actions. And keep operational recommendations separate from control-system execution.

Plenty of vendors sell autonomy as though fewer brakes make a better truck.

Practical tool watch

Oracle Maintenance Advisor and Work Order Builder. Worth examining even if you never touch Oracle. The feature boundaries are a useful map of what enterprise CMMS agents are starting to do inside real workflows.

IAMTech iPlanAI. Study the proposed work-package outputs. The question is not whether the demo impresses. It is what source evidence accompanies each generated labour, duration and tooling estimate.

Prometheus STO AI. Connects historical turnaround information to scope and package generation, rather than treating every shutdown as a fresh spreadsheet festival.

IBM Maximo AI 9.2. Meta-agent support matters strategically. The near-term question is whether agent coordination becomes transparent and controllable, or just another black box above Maximo's existing complexity.

The buildable idea this week

Not predictive maintenance. A maintenance readiness review.

Something that reads a work-order backlog or a shutdown scope and identifies the jobs that are not ready to execute. It does not diagnose equipment. It exposes operational uncertainty: missing evidence, unclear scope, unverified parts, absent tooling, hidden coordination, likely discovery work, resource collisions, and jobs with no completion criteria.

It needs no sensor infrastructure, no training dataset, and no control-system access. It can run on a CSV export in week one.

The wedge is narrow. The underlying market is not.

Signal vs. Noise

Durable signal. Evidence-constrained reasoning. CMMS write-capable agents. AI-assisted work-package generation. Auditable human approval.

Probable hype pocket. Claims that a system can turn thin historical data into accurate failure predictions, or into complete shutdown plans, without serious data preparation and planner intervention.

Core mechanism. AI is lowering the cost of producing a proposed answer. It is not lowering the cost of determining whether that answer deserves to be trusted.

What I'm Watching Next

feature separates a decision layer from a writing tool. - Whether shutdown scope auditing appears as a product category on its own, or stays buried inside planning suites where it competes with the scheduling features that sell. - Whether authority boundaries get specified per action rather than per system. "The agent has write access" is not a policy. - What happens the first time a plant investigates an incident and finds the decision trail runs into a chat window.

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