GIE ADVISORY GROUP | INSIGHTS
The most important steps in your AI strategy may have nothing to do with AI
Artificial intelligence is becoming part of almost every technology conversation.
I hear versions of the same questions from executives, operators, engineers and technology teams: How can we use AI in our facilities? Can it improve maintenance? Can it optimize labor? Can it make our WMS smarter? Can it predict equipment failures?
Those are reasonable questions. But based on some recent conversations, I think many organizations are starting the discussion a level or two too high.
Before asking what AI can do for an operation, I would ask a simpler question:
Is the operation ready for AI?
If equipment failures are not consistently categorized, what exactly are we asking AI to predict? If two facilities calculate downtime differently, what should it learn from? If the WMS says one thing and the floor is doing another, which version of the operation should it trust? If critical equipment is not connected, where does the condition data come from?
And if nobody consistently responds when a KPI moves outside its target, what value does predicting that movement really provide?
AI does not remove the need for operational discipline. In many ways, it makes that discipline more important because decisions that once relied on tribal knowledge increasingly depend on whether that knowledge has been translated into reliable data, rules and processes.
1. AI Can't Fix an Unstable Foundation
Most industrial organizations do not need more data simply for the sake of having more data. I have often found myself looking at too much data—unorganized, inconsistently defined and unclear in what it was actually telling us. Sometimes the answer is not more information. It is less information viewed through a sharper lens.
That starts with some fairly basic questions. Are the processes repeatable? Are losses categorized consistently? Do people agree on how the KPIs are calculated? Is there clear ownership when performance moves outside the target? Are maintenance events captured in a way that can actually be analyzed later?
OEE is a good example. Reporting an OEE of 62 percent tells you that performance is below where you want it to be. It does not tell you why.
Was the loss availability, performance or quality? Was the equipment down because of a failure, changeover, starvation or blockage? Was the line running slowly because of micro-stops, an upstream constraint or product mix?
A KPI tells you that performance changed. The operating foundation underneath it helps tell you why.
Before introducing a more sophisticated layer of analytics, the organization needs enough process discipline and standardization to trust the inputs.
2. Start by Connecting the Operation
The next step is making the operation visible.
In many facilities, useful information already exists—but it lives in separate places. Equipment has PLC data. Controls systems have alarms and states. SCADA may have runtime and process information. The CMMS has maintenance history. The WMS has inventory and transaction data. ERP contains another view of the business.
The problem is often not the absence of information. It is that the information is disconnected, inconsistently defined or difficult to put into operating context.
Connecting the operation does not mean every organization needs to launch a massive digital transformation. Start with the assets, processes and systems tied to the problems you are actually trying to solve.
At a previous company, we built much of this visibility through a relatively simple Ignition architecture. As we selected new equipment, we were intentional about controls, communication and the information we wanted available upstream. For existing or simpler assets, sometimes that meant finding practical ways to bring basic signals—run status, faults, temperatures, pressures or other conditions—into the same environment.
The result wasn't technology for technology's sake. Automation, utilities, HVAC and other critical systems could increasingly be viewed through a common operating layer. It was easier to see what was happening, easier to control what needed to be controlled, and it gave the organization a foundation it could continue building on.
You don't need every asset to be sophisticated. You need a practical way to get the information that matters.
If downtime is the problem, start with the equipment and failure information that drives downtime. If inventory accuracy is the problem, start with the transactions and process discipline behind inventory. If throughput is the problem, connect the information necessary to understand where flow is being lost.
I'm also not a fan of layering software simply for the sake of complexity. Sometimes a relatively simple platform sitting above several systems is faster and more useful than trying to force multiple enterprise transformations at once.
Connectivity gets the data out. Context makes it useful.
3. Create Value Before You Ever Get to AI
This is where I think the AI conversation can become unnecessarily complicated.
The first useful application of connected data does not need to be a machine-learning model. It might be a dashboard. It might be an automated alert. It might even be conditional formatting in a spreadsheet.
Take a simple maintenance example. Assume an organization wants to move away from purely calendar-based preventive maintenance and instead service an asset based on actual runtime as one step toward more condition-based and eventually predictive maintenance.
The first version might be extremely simple:
· Green: the asset is comfortably below its maintenance interval.
· Yellow: runtime is approaching the threshold.
· Red: the asset has reached the point where action is required.
There is nothing artificial about that intelligence. But think about what had to happen to make the simple red-yellow-green indicator trustworthy.
The organization needed accurate runtime data. The asset needed to be identified correctly. A maintenance standard had to exist. Someone had to define the thresholds. Ownership had to be clear. And there had to be an agreed response when the indicator turned red.
That same foundation can support progressively more sophisticated capability:
Conditional formatting → automated alert → trend analysis → condition monitoring → predictive model → AI-supported recommendation
The value does not suddenly appear at the AI step.
The goal should not be to get to AI as quickly as possible. The goal should be to create value at every step toward it.
4. Maintenance, Operations and WMS Follow the Same Pattern
Maintenance is an easy example, but the principle applies across the operation.
Maintenance
Before predicting failures, consistently capture what failed, when it failed, what the equipment was doing, what conditions were present and what corrected the problem.
At a previous company, I implemented a CMMS from scratch in the middle of a live operation. For anyone who has done that, you know how difficult it can be. For those who haven't, there is usually a reason an organization suddenly decides it desperately needs a CMMS—and it isn't because everything is running perfectly.
We were effectively operating at roughly 20% preventive and 80% reactive maintenance. Within several months, we had reversed that relationship to approximately 80% planned/preventive and 20% reactive.
There was no AI involved.
We established the asset hierarchy. Built the PM structure. Improved work-order discipline. Created ownership. Started collecting better history. And changed the way the organization approached maintenance.
That foundation is what eventually makes predictive maintenance possible.
Moving from reactive to predictive isn't primarily a technology journey. It's an operating maturity journey.
Operations
The same principle applies to broader operations. Before asking AI to optimize OEE or throughput, first make sure you trust the OEE calculation and understand the losses underneath it. Before asking AI to identify a throughput constraint, make sure you can reliably measure flow, downtime, cycle time, starvation, blockage and process variation.
Start by getting the information into a place where the operation can actually use it. That may be Power BI, a SCADA interface or another reporting layer. Understand the data. Establish the operating response. Make incremental changes against the underlying losses.
I've always thought about AI in operations somewhat like simulation modeling. The model can be extremely sophisticated, but the quality of the answer still depends on how the problem is structured, the assumptions underneath it and the objective you're asking it to optimize.
First make the operation visible. Then make the response consistent. Then start asking what can be predicted or optimized.
WMS and Inventory
The same logic applies to warehouse systems. AI cannot compensate for poor transaction discipline, inaccurate inventory, inconsistent location practices or a gap between the system process and what actually happens on the floor.
There is another issue as well: AI still needs to know what you're asking it to optimize.
Consider a WMS. The 'best' operating decision can look very different depending on the business objective. Are we optimizing for order fulfillment speed and service level? Labor utilization and cost? Travel distance? Inventory accuracy? Order consolidation? Dock utilization? Customer priority?
Those objectives can conflict with one another.
Reducing travel distance may improve labor productivity while creating a different inventory or replenishment constraint. Prioritizing service for one customer or order profile may reduce the efficiency of the overall wave. Maximizing storage density may negatively affect accessibility and throughput.
AI can become extraordinarily useful in evaluating those tradeoffs. But it doesn't eliminate the need to decide which tradeoffs the business is willing to make.
And if you simply ask AI to optimize all of those targets at once, you may find that you don't meaningfully satisfy any of them. Like a simulation model, optimization still requires priorities, constraints and a clear definition of what success looks like.
Strategy still comes first. The operation still needs to be stable, repeatable and measurable. Leadership still needs to define what 'better' actually means.
Different use cases. Same foundational problem.
5. Build Toward AI, Not Around It
I think the path is simpler than many technology roadmaps make it appear:
Stabilize → Connect → Understand → Predict
Stabilize the process. Establish standards, ownership and operating discipline.
Connect the equipment and systems that generate the information needed to manage the process.
Understand what the data is telling you. Turn it into trustworthy KPIs, operating signals and defined responses.
Then predict. Once the organization understands the operation and has reliable history underneath it, advanced analytics and AI have something meaningful to work with.
This also changes how I think organizations should approach capital projects. New equipment should not be evaluated only for mechanical performance, throughput and cost. Controls architecture, data access, communication protocols, integration requirements and ownership of the data increasingly belong in the equipment specification as well.
Even if the broader AI roadmap is five or ten years out, decisions being made today—from facility infrastructure to equipment selection—can either create the foundation for that future or make it significantly harder to achieve.
It is much easier to establish those expectations during the RFP and procurement process than to discover after commissioning that a critical piece of equipment is effectively a data island.
The Foundation Is the Strategy
AI will undoubtedly change how industrial and supply chain operations are managed. But it does not eliminate the need for good processes, stable operations, connected equipment, trustworthy metrics and clear ownership.
Start with the process. Stabilize the operation. Connect the equipment. Define the metrics. Make the data trustworthy. Use simple tools to make better decisions today.
Then keep building.
The first step toward predictive maintenance might eventually lead to a sophisticated machine-learning model. Or it might begin with conditional formatting in a spreadsheet that finally gets the organization responding consistently to the right information.
Both are progress.
Because the question isn't how quickly you can get to AI.
It's whether you've built an operation worth applying AI to.
