The Pitch Is Correct
Industrial operations AI can predict equipment failures before they happen. It can answer natural language questions from operational data. It can identify maintenance patterns that human analysts would take weeks to find. It can synthesize sensor data, maintenance history, and operational context into recommendations that improve how operations teams prioritize their time.
The capability is real. The operational results are real. The use cases — from predicting bearing failures to optimizing PM intervals to answering questions like “which of my compressors have the worst cost-per-operating-hour ratio over the past year” — are genuinely valuable.
The problem isn’t the pitch. The problem is the prerequisite.
What AI Actually Runs On
When an AI system answers a question about industrial operations, it reasons from the data it has access to. The quality of the answer is a direct function of the quality and completeness of the underlying data.
This seems obvious stated plainly. It becomes less obvious when you’re watching a demonstration of AI that seems to know everything about the demo environment, or when a vendor’s pitch focuses entirely on the AI capabilities rather than the data model those capabilities require.
AI for industrial operations doesn’t run on air. It runs on:
Asset identity records — who is this equipment, what is it supposed to do, when was it installed, what are its specifications. Without accurate identity records, AI can’t reason about age or replace-vs-repair.
Sensor telemetry history — the time-series record of what the equipment has been doing. Without comprehensive sensor history, the AI can’t identify baseline behavior, detect deviations, or analyze trends.
Maintenance records — every work order, every intervention, every finding, every part replaced. Without complete maintenance history, AI pattern analysis has gaps. It may conclude that an asset has never had a bearing failure when in fact the bearing failures were documented in a spreadsheet that was never integrated.
Alarm history — the record of every threshold exceedance, every alert condition, every time the system detected something worth noting. Without alarm history, the AI can’t identify recurrence patterns or correlate alarm conditions with maintenance events.
Current state — what the equipment is doing right now. Without live sensor connectivity, the AI can reason about history but not current condition.
The AI doesn’t know what it doesn’t know. It will answer questions from the data it has. If the data is incomplete, the answers are incomplete — and they will appear complete, because the AI is responding confidently from the data it can see.
The Confidence Problem
AI-generated answers have a characteristic that makes data quality problems especially dangerous: they look confident regardless of whether the underlying data is good.
A human analyst who doesn’t have complete maintenance history will say “I don’t have complete records, so take this with some uncertainty.” An AI system without complete maintenance history will say “The maintenance cost for this asset over the past 3 years has been $12,400” — because that’s what the maintenance records show, and the AI doesn’t know that 40% of the maintenance events weren’t captured.
The answer is technically accurate given the available data. It’s operationally wrong because the available data is incomplete.
This is particularly dangerous for maintenance pattern analysis. If an AI recommends extending PM intervals because “no failures in the normal service window,” and the reason is failures were documented in spreadsheets that weren’t integrated — the AI’s recommendation leads to extended intervals that create failure risk.
What “AI-Ready” Actually Means
AI-readiness for industrial operations is a data model question, not an AI question.
The AI capability — the model, the inference architecture, the natural language interface — is available. What varies by organization is the quality and completeness of the operational data model those capabilities run on.
Asset identity coverage: What percentage of physical assets in the operation have complete identity records in the system? Manufacturer, model, installation date, location, service specifications. An operation with 70% identity coverage will have AI capabilities that work well for 70% of assets and produce incomplete answers for the rest.
Sensor telemetry coverage: What percentage of operationally significant assets have active sensor connectivity? Coverage decisions involve trade-offs — not every asset justifies the cost of sensor instrumentation. But the AI’s condition assessment capability is only as comprehensive as the sensor coverage.
Maintenance history integration: Is maintenance history captured in a single system, or distributed across multiple systems (CMMS, spreadsheets, email)? Distributed maintenance history produces incomplete AI analysis. Consolidated maintenance history produces comprehensive analysis.
Data staleness: Are sensor readings arriving at appropriate intervals? A sensor reporting every 15 minutes provides different baseline and anomaly detection capability than one reporting daily. For condition monitoring applications, reporting frequency affects the AI’s ability to detect developing failures before threshold exceedance.
The Foundation-First Principle
The vendor pitch for industrial AI rarely leads with “first you need to build a complete operational data model.” It leads with the AI capabilities. The data model is a footnote, if it’s mentioned at all.
This creates a sequencing problem. Organizations buy AI capability without first building the data foundation. The AI is deployed on an incomplete operational model. The initial results are good enough to seem valuable — the AI can answer the questions where data is complete — but the most important analyses are unavailable or unreliable because the data isn’t there.
The correct sequencing is foundation-first:
-
Build the asset identity model. Create complete asset records for every asset that matters. Populate identity fields. Establish the organizational hierarchy — which assets belong to which lines, which lines belong to which facilities.
-
Connect sensor telemetry. Instrument the assets that warrant continuous monitoring. Get data flowing into the asset records. Let baseline statistics develop over 30+ days.
-
Consolidate maintenance history. Migrate existing maintenance records into the CMMS. Establish the workflow for future maintenance to be captured. Close the gap between “what happened” and “what’s recorded.”
-
Activate the alarm layer. Configure threshold and condition-based alerts. Begin accumulating alarm history.
-
Deploy AI on the complete model. Now the AI has access to identity, state, behavior, and context. Now the answers are comprehensive.
Organizations that try to shortcut this sequence — deploying AI on incomplete data — get AI that works partially. The partial results create a false impression that AI isn’t living up to its potential. The real issue is that the foundation wasn’t built first.
The Time Investment Is Real
Building the operational data model takes time. For a mid-size manufacturing operation with 500 assets, a 2–4 month foundation build produces fully operational AI. Building complete identity records, connecting sensor telemetry, and consolidating maintenance history is the required investment.
For organizations that have been running IoT monitoring with a device-centric platform — sensors connected, dashboards working, but no asset identity layer and no integrated maintenance history — the migration to an asset-centric model requires building the identity and maintenance layers that the monitoring platform didn’t have.
This is not an argument against building the foundation. It’s an argument for being realistic about what the timeline looks like and what the AI will deliver at different stages of foundation development.
The AI is only as good as the data model. The data model has to come first.
A Calibration Question
Before deploying industrial AI, the honest question to ask is: if the AI answered a question about our maintenance cost history, what percentage of actual maintenance costs would it have access to? If the AI assessed the condition of our asset fleet, what percentage of assets would it have complete telemetry for? If the AI identified failure patterns, what percentage of historical failures would be in its accessible records?
Below 80–90% coverage, the AI produces incomplete answers — confidently.
The goal of the data foundation work is to raise those percentages to the point where AI answers are comprehensive enough to be operationally reliable. Not perfect — perfect is an unrealistic target for any operational data model. Comprehensive enough that the gaps are edge cases rather than significant blind spots.
Build the foundation. Then deploy the AI. In that order.
Talk to our team about building an AI-ready operational data model on ArgusIQ.