Updated: October 15, 2025 (October 15, 2025)
BlogCIO Talk: The Enterprise Semantic Layer: The Missing Link for AI (and Microsoft)?
This blog post is a little different. I’m not going to tell you about the pros and cons of a Microsoft technology – rather, I’m going to talk about a missing Microsoft piece of the puzzle.
Maybe it’s even the Redmond giant’s Achilles’ heel.
As AI transforms practically everything about how we work, it’s become clear that a well-managed and well-governed data estate is critical for its success in the enterprise (see my video). But even with data that’s appropriately secured, and whose regulatory compliance is constantly monitored with Purview and other tools, the question remains: how can AI put all this data work?
Sisyphus and the Enterprise Data Model
We in IT have struggled for decades with so-called “islands of data,” that is, incompatible clumps of information squirreled away in ERP, CRM, Support, Legal, IoT, e-Commerce, and other systems. One application’s “customer” is another’s “contact” or yet another’s “client,” or “lead,” or “account.” Some keep two address lines for the individual, some four; some have a 5-digit ZIP code (in the US), some nine; and keeping them all in sync is, in a word, a nightmare that can cost enterprises millions.
To simplify this, information architects and data managers have struggled for years to build an “enterprise data model:” common definitions of key subject areas (“customer,” “product,” etc.). With such common definitions – a blueprint, if you will – all departments and applications can have a shared understanding of the subject area, can better achieve compliance goals, can improve communications, and can streamline application integration efforts.
And there have been many attempts to create industry-standard models, and some have been useful. But there’s a key missing point here, which is that data models encapsulate a company’s competitive differentiation. (Think about it. It’s true.) Standardize the data model, eliminate your differentiation.
Because of that, data models that affect competitive differentiation – how a company thinks of customers, products, pricing, discounts, and so on – necessarily change rapidly, much faster than an EDM could track.
Yet…the dream remains…wouldn’t it be nice if there were one single data model for each subject to which all applications aligned?
Well, yes. But it’s been an uphill battle (thus Sisyphus). Why? Creating and deploying an EDM is excruciatingly expensive, painful, and difficult, both in terms of the human cost of negotiating common models as well as the technical costs. While potentially streamlining integration (no more “T” in ETL!), simplifying process modeling, and other cost-saving benefits, the cost and risk of updating applications schemas is so high that it’s never been seen as a worthy effort.
Until now, maybe?
Maybe, Just Maybe, the Enterprise Data Model is Possible
But imagine – just for a moment – if you could describe, even at a very high level, your core business entities. (A quick note on terminology. When we (and vendors) say “subject area,” “business object,” and “business entity” we mean pretty much the same thing: the nouns that run your business, as in a customer buys a product. See how hard it is to agree on terminology?)
Say business folks and IT staff can come together and decide, “here’s the authoritative definition of a retail customer” for our company.
Then (keep imagining) you push a button and automagically, using automated connectors, software reaches out to applications and maps your definition to real data in the real world. And maybe it (perhaps with human help) creates relationships (a retail customer “is a” kind of customer, “has an” account manager, and so on).
Companies like Celonis and Palantir propose to do just this. On top of the oceans of data collected in Snowflake, Databricks, or Fabric, then, these models form the foundation of what my friends George Gilbert and Peter O’Kelly call a “semantic layer,” imposing order on chaos.
The Enterprise Digital Twin: Stir in AI, and Magic Happens
The vision for both companies: to create an AI-powered “digital twin” of the entire enterprise. Think of it as the organization’s automated pan-intelligent COO: you can issue it commands, you can ask questions of it, and expect it to know everything that’s going on, everywhere in your business.
Greatly Simplified View of Enterprise Ontology
Celonis uses sophisticated process mining to discover both how data is used in processes, and the often myriad variations of processes hidden in the data. By shining a light on enterprise processes — remember, a company is nothing less, nothing more than a collection of processes – you can optimize them. Where are the bottlenecks? Why, for example, when a customer orders a widget does it sometimes take weeks to fulfill?
Whereas Palantir focuses on prediction, simulation, and coordination based on holistically linked real (and real-time) data. So if CEO wants to know “how will raising the price of a widget by 5% in Japan affect revenue and profitability?” – in theory with such a digital twin those sorts of questions could be answered.
Color Me Skeptical
Now at this point I have to inject some personal experience and opinion. I’ve been involved with efforts to build an EDM in large enterprises; it’s unspeakably hard, for all the reasons noted above. It’s very labor-intensive and detail-oriented: at one company I recall seeing a model for “customer” printed on a plotter that stretched out across an entire conference room table (in 8-point font!). Fundamentally: useless.
I’m willing to keep an open mind – but for any such endeavor like this I would caution all: show ROI quickly. Do pilots. Look for quick wins. Focus on department processes, not the full enterprise to start. Demonstrate the value as quickly as possible!
Where is Microsoft?
Now I’m sure you’re saying: Barry, this is the Directions on Microsoft blog, and you’ve barely mentioned Microsoft!
True! Microsoft, of course, has made massive investments in data management, most recently with Fabric, its answer to Snowflake and Databricks (and others). And let’s be clear, what’s different about today is the presence of these massive data lakes, lakehouses, and so on, that consolidate data in ways we were never able to before.
But, to date, Microsoft has done little to nothing about a semantic model. Purview does have a Business Glossary of common business terms – but it’s not linked to real data in any way. Mary Jo reminds me that they bought a process mining company called Minit back in 2022 – it, however, seems to have been folded into Power Automate. And (just to confuse matters) there’s Flow Builder, which industry watcher Jukka Niiranen speculates could replace Power Automate. But Power Automate – like the Power Platform generally – isn’t targeted at enterprise-scale, enterprise-class scenarios. And it doesn’t pretend to create a “semantic layer.”
And yet this semantic layer may be the thing that truly unlocks the power of AI in the enterprise.
I’ve heard rumors of a stealth mega-project underway at Microsoft to build some sort of “semantic layer” that competes with Palantir and Celonis. Is it real? Is it good? Is it Minit reborn? (If you know, drop me a line). We’ll see.
For now, it seems like a gap. And it may be a big one.
Do I have gaps in my thinking? Drop me a line at bbriggs@directionsonmicrosoft.com. And if you haven’t read The Laws of Business Process, well, you really must.
