Picture a line that misses its output on Tuesday. On Wednesday morning someone opens a report, sees the gap, and starts asking what happened. By then the shift has changed, the product has changed, and the people who saw the problem have moved on. The data is correct. It is just late, and it lives in more than one place.
This is the problem I want to look at today, because it limits continuous improvement more than most teams admit.
Three systems, three partial truths
The m365.fm episode Connect ERP, MES, and IoT Without Replacing Systems starts with a bottleneck machine that stops. Each system holds a correct but partial view of that event.
The ERP knows the commitment: the order, the material, the quantity and the promised date.
The MES knows the execution: the active operation, who is working on it, what has been reported and whether there is a quality hold.
The IoT and PLC layer knows the physical event: the alarm, the stopped state, the counters.
None of them is wrong. The gap is that nothing connects their meaning. The same press can be called “press line 2” in one system, “PRS02” in another, and a tag name in a third. Without an agreed asset identity, a simple question such as “which order was affected?” becomes a manual job.
Why this hurts improvement work, not just reporting
A Gemba Walk, a kaizen event or a PDCA cycle all rely on a check step: did the countermeasure work, where, and under what conditions? The episode Why Continuous Improvement Needs Better Data argues that kaizen often shows quick progress but the same losses come back weeks later on another shift, because the team cannot prove what actually changed. The evidence sits in separate systems and in operators’ heads.
It gives a changeover example. After a fix, the changeover gets faster, but the loss moves to restart time and quality checks. If you only watch the one visible step, you declare success too early.
The related article on manufacturing data models adds a point I find easy to forget: the context also needs versioning. If recipes, routings or target rates change and you judge old production data by today’s rules, you can draw the wrong conclusion about whether an earlier countermeasure worked.
The tempting fix, and why it breaks
The quick answer is to connect two systems directly. For one use case this works. But as more teams want the same data, the mappings multiply. The article on why point-to-point manufacturing integrations fail describes the worst outcome: the feed stays up, but values are quietly wrong, such as swapped asset IDs or incorrect units. Those errors can distort OEE and quality figures for weeks before anyone notices.
The episode’s answer is not a bigger system. It asks for a few agreed things:
an asset identity map that links local names to one governed identity;
a named owner for each fact, with the principle to read broadly and write narrowly;
data contracts that define meaning, units, timestamps and how late data is handled;
both source time and arrival time, so late data is visible as late.
It also describes a unified namespace as a governed event space, typically over MQTT, where an event is published once and approved consumers subscribe. The episode is clear that this does not replace ERP transactions or historical storage, and that control stays with the PLC. ERP and MES stay where they are.
Where Microsoft tools fit
The sources are careful here, so I will be as well. The episode treats Microsoft products as building blocks, not as a replacement strategy. Microsoft Fabric is described as the shared analysis space where telemetry, MES events and selected ERP records meet, and Power BI as the decision-facing view on top of governed models. Power Platform, meaning Power Apps and Power Automate, is suggested for bounded tasks such as capturing downtime reasons and approvals, not for uncontrolled writes back into production systems. Dynamics 365 is mentioned as one example of an ERP.
Excel gets a fair treatment. It is often the hidden integration layer, holding relationships no formal system captured. The point is to move those relationships into governed shared models, not to ban spreadsheets. In my view it becomes a risk when it is one person’s private reconciliation process.
What to do on Monday
You do not need a platform project to start. Pick one recurring loss and write down, for one cell, which system says what about it. List where the names and time boundaries disagree. Name an owner for each definition: downtime, scrap, good count. Then take that list to the line and check it with the people who run it.
That last step is the Gemba part. The data model tells you where to look, and the people on the floor tell you what it means. Both are needed, and the first is what makes the second happen sooner.
Listen / Read more
Episode: Connect ERP, MES, and IoT Without Replacing Systems on M365.FM.
Article: Why Manufacturing Data Models Are Essential for Continuous Improvement on m365.fm.

