Semantics Without Silos: AI Finally Solves BI’s Oldest Problem

Dave:
Hi everyone, and welcome to another edition of the Data-Driven Podcast. Today I have a very special guest, someone I’ve known for a long time who’s been instrumental in data and analytics in this industry — it’s Donald Farmer, principal of Treehive Strategy. Welcome to the podcast, Donald.

Donald:
Thank you very much, Dave. It’s great to be here.

Dave:
In terms of your background, you’ve been a pioneer in data and analytics for a long time. Could you let listeners know some of the companies you’ve worked at and what you’ve done? You have an amazing background.

Donald:
My entire career has been spent in data analysis in one form or another. I worked at Microsoft for ten years, developing the ETL products — Integration Services — and led the user experience team for what was known as PowerPivot, which later became Power BI. I was also on the SQL Server data mining team, leading program management for predictive analytics. From there I went to Qlik, the Swedish company that pioneered self-service business intelligence, and led development of their second-generation product, Qlik Sense. Before Microsoft, I was involved in a couple of award-winning startups in Scotland. Now I advise investors, startups, and enterprises on data and analytics strategy, I’m a research fellow for TDWI, and I do a lot of writing. Busy, but always in the data and analytics space.

Dave:
And you were there at the beginning of business intelligence itself — Power BI is the number one BI tool today, and you were there for its origins, and for Qlik’s. So with AI in the picture now, what’s changed? Has AI modified BI’s trajectory, or not really?

Donald:
It’s a good cure for making prognostications. One thing we love to do in business — because we’re storytellers, marketers — is declare that a new technology makes the old one obsolete. When I first came across Qlik, before I joined, their marketing team used to march up and down outside SAP and data warehousing conferences with signs saying “no more data warehouses, no more cubes.” The data warehouse is dead, they said, you do it all in Qlik. And we’ve been through this before. Of course, OLAP isn’t dead, data warehouses aren’t dead — Snowflake’s IPO was the largest tech IPO in history at the time, and it’s a data warehouse. That technology didn’t die; the opposite happened. But there was still a real point being made: a new technology had entered the market.

We see this again and again. Sigma’s big release this year was pixel-perfect enterprise reporting — who’s still doing that in the 2020s? People still need it. So what happens is new technology offers new capabilities, but the old technology hangs around. I think the future of business intelligence is much like the future of enterprise reporting: it’s not going to be where the competitive edge lives, but it’s still essential for keeping the business running smoothly and repeatably.

Then there’s the other part of business intelligence — the exploratory, discovery side. How do I find something new that I don’t know about my business? That’s where AI is doing exciting work, surfacing insights we couldn’t get to before because of scale or complexity. So the future of BI as we know it is these now-slightly-outmoded but still-essential repeatable dashboards, while the cutting edge becomes AI-enabled discovery. Does that make sense? Long answer to a short question.

Dave:
It does. Some of my customers put it this way: dashboards have a place, because do you want five thousand employees asking an agent the same question every morning? No — a dashboard is a more efficient way to serve that up. But when it comes to exploratory analysis, we’ve always talked about different levels of analytics, and I don’t think BI ever did a great job on that exploratory side.

Donald:
It didn’t, because it took a genuinely skilled analyst — someone with real experience in these tools — to discover new insights. With AI, that researcher role gets automated. AI handles a lot of the pointing, clicking, and dragging, and the analysis that happens in between, to do a better job finding the needle in the haystack.

Dave:
That’s exactly where I’ve seen it play out. I used to demo AtScale by opening Tableau, dragging in revenue, looking at revenue by product, then by month — “oh, this product dropped off this month versus last.” Then the question becomes why, and I’d have to go hunting through dozens of metrics and attributes to figure out where to look. Now I do that same demo with Claude. I just say, “show me all products that had a drop in revenue and tell me why,” and it runs the relevant queries, analyzes the results, and goes deeper on its own. That’s a game changer two ways: it’s much faster, and anyone can do it. We’ve talked about democratizing data and analytics for our entire careers — a conversational interface really is the ultimate version of that, isn’t it?

Donald:
It is, in the sense that you only need to be able to formulate the question — and that’s genuinely important. A lot of business leaders can formulate great questions but struggle to navigate to the answer. AI solves that navigation problem.

But there’s something else BI was never good at. I used to talk to design teams about a five-step analytic mindset, borrowed from library science — originally the field that tackled the first “big data” problem: how do you navigate a library? A researcher named Marcia Bates worked out a model of browsing. It starts with orientation — what am I looking at, how do I navigate. Then comes glimpsing — how do I even notice something worth looking at? The right visualization helps you glimpse things; the wrong one doesn’t.

Dave:
I like that.

Donald:
After glimpsing comes examining — digging in, drilling down, drilling across, understanding in detail. But the real fourth step — in the library world, they call it acquisition. “That’s the book I want, I’ll take it.” In the BI world, what’s the equivalent, once you’ve oriented, glimpsed, and examined something? It’s: what do I do about it? And business intelligence never helped with that directly. You’d find an interesting data point, print a nice visualization, share it — and that’s where BI left you. It opened the door and said, now it’s up to you to walk through it.

Dave:
That’s exactly it — BI took you right up to the door and left you on your own to close the deal. Donald, we’ve always talked about prescriptive analytics — that was always the piece I struggled to explain, because BI and OLAP alone couldn’t do it. You had to bring in machine learning and data scientists to translate the insight into action, and that created enormous friction. It never really crossed that chasm.

Donald:
I always laugh about prescriptive analytics, because a pharmacist friend of mine tells me how many people never fill or take their prescriptions. That’s exactly the problem with prescriptive analytics — you hand someone the prescription, and they don’t act on it. I think agentic AI is what actually enables us to build the workflows that take the action. That’s a huge change, and I’m genuinely excited about it.

Dave:
That brings us to accuracy and semantics — because if AI is going to take prescriptive action on data, with no human necessarily in the loop on that action, the underlying data foundation has to be accurate. You were there for the birth of the semantic layer at Microsoft. How does that translate into this new AI world? How do we drive trust, and what’s the right context layer to keep AI from going off the rails?

Donald:
This is genuinely important. Semantics can sound overly technical, or overly philosophical — “that’s just semantics” — but it’s really about the meaning of the data. We used to talk about building a single version of the truth: extract data from every system, impose one model of meaning on top. The problem is that’s a prescriptive model of semantics — we define the meaning first, then pour data into it — rather than a descriptive model, which starts by understanding what’s actually there.

Take a simple word like “revenue.” It legitimately means different things to different people. Marketing running a campaign may define revenue differently than a salesperson being compensated on it — if a sale falls through or a product gets returned, that might not count as revenue for comp purposes, but marketing already did their job getting someone to buy. Finance has its own definition tied to revenue recognition, especially for public companies. All of these are legitimate.

The traditional fix was to create separate definitions and attributes in the data warehouse. What I think matters more now is understanding not just the definition of revenue, but its relationship to every other measure and data element around it. Semantics stops being just a question of definition and becomes a question of relationships — which is richer and more flexible. We always wanted to build that web of context, but it was too time-consuming. Now the web of context you can build is flexible enough to be consistent and still adapt to different scenarios across the business.

Dave:
That makes sense. We’ve both been in BI forever, and every BI tool always had some internal semantic layer — we sell one too. And the hardest part for customers was always building it, because it requires a unicorn: someone who understands the business and someone who understands the data structure well enough to map one to the other. That’s what a semantic layer really is — a mapping of logical to physical. What I’m seeing now is AI isn’t just for consuming a semantic layer, it can actually produce one. What used to be the real barrier to a single source of truth — building that semantic map — AI can now do for you. Are you seeing the same thing?

Donald:
Absolutely. And building that map was always hard. One of the important insights behind founding AtScale — and I don’t say this to flatter you — was recognizing that semantic layers built inside individual BI or visualization platforms just create more silos. You end up with competing semantic layers instead of something that speaks across the enterprise. But then you have the problem of how you actually build it. It’s easy to build implicitly — every time you draw a chart or lay out a dashboard, you’re implicitly building a semantic layer — but the problem is you end up building hundreds of them. That’s a distraction. I think we’ve reached a point where a genuine semantic network is interesting again, because AI makes it achievable — and, importantly, achievable by people with business knowledge, not just technical skill.

Dave:
Right, the challenge before was that building it was a genuinely technical job.

Donald:
Yes.

Dave:
It took someone who was, essentially, your most sophisticated Tableau user — someone who understood the business well enough to define things correctly, and also knew how to traverse the schema and relate tables to get the right answer. It took a unicorn. What we’re seeing now is that because the business has AI backing them, they only need to understand the business — the semantic models themselves can be generated. And having a library of semantic objects matters a lot here, because you don’t want to end up in the same mess BI tools created: hundreds or thousands of competing semantic models. Even if you let the business use AI to create these models, you still need CI/CD-style control so you’re not creating duplicate, competing semantic objects when you federate that work out. It’s important to treat semantic models like code — with a similar lifecycle, approval processes, pull requests — while still letting the business create the objects, because they understand the domain.

Donald:
That’s really important, and I think people are starting to understand it. TDWI does regular research on AI and analytics readiness, and we’ve found that about two-thirds of respondents understand that semantics is critical to AI success. So the message is landing — the practical implementation is a bit behind where people want it to be, but the need is clearly recognized, which is encouraging.

Dave:
We’re seeing that too. A couple of years ago we reinvented our platform around SML — semantic modeling language — a YAML-based language for defining semantic models. You can build models visually through our tool, but it generates that code behind the scenes. What I didn’t predict was how good AI would turn out to be at writing code. That becomes the bridge for AI to write semantic models — because if your semantic models are based on code, AI can do a genuinely good job building them for business users.

Donald:
That’s exactly what we’re seeing. People often get confused about what AI is actually good at — it’s really good at language, and especially good at translation. Creating a semantic model in YAML is, in a real sense, translating a human description in natural language into another language. If you think of it as code generation, that sounds technical and failure-prone. If you think of it as translation, the kinds of errors AI produces become a lot more manageable — they’re natural, almost human, errors. You just need a different mindset for finding them than the one we use for traditional technical validation. I think the insight you get from an AI’s translation of a semantic model is often greater than what limited human capability could model directly for something this complex.

Dave:
Definitely seeing that. There’s also been an effort — because we’ve created all these silos and competing semantic layer platforms — around OSI, Open Semantic Interchange, an initiative to standardize the semantic language so tools can interoperate. What’s your take on that?

Donald:
It shows this is a shared industry problem, which matters. The adoption has been pretty remarkable — last I read, something like 60 to 80 significant vendors had signed on, in a very short period of time. That tells you there’s a real, industry-wide problem being addressed. There are still hard parts — metric definitions, composability of models, integration with data and analytics catalogs — but there are working groups tackling that. Simple metrics like counts and sums are easy to define under OSI; more complex ones are harder. Still, that’s tremendous progress in under a year.

It also speaks to something important about this new industry: monolithic platforms aren’t the way forward. Every enterprise needs a portfolio of tools that work together without creating more silos — that’s why the network effect of OSI matters. Even SAP, historically the ultimate walled garden for data, has opened up its ecosystem and started inviting in vendors. We’re firmly in a world where enterprises need a portfolio of vendors and systems, and OSI is what makes that manageable. I’m a real enthusiast for it.

Dave:
Great to hear. That was one of our whole theses when we started the company — preventing vendor lock-in. Freedom of choice across data platforms and BI tools matters, because you never know what’s coming next. Apache Iceberg has done something similar on the data platform side — companies are storing data as Iceberg specifically so they can use it interchangeably, internally or externally. Workday, Salesforce — they’re all making their data available in Iceberg format now, so you can get a federated view across data, analytics, and AI.

Donald:
Absolutely, and it’s an exciting prospect. People often forget how early we still are in this stage of AI development. There’s a long way to go in making AI as robust as we want — in understanding user experience more broadly, not just interface; in governance; in compliance. All of that is still evolving. So sure, you can be critical, but we’re at a stage where things are developing rapidly, and the future looks very positive.

Dave:
I’ve never been more excited in my career about what’s possible. Donald, you talk to a lot of companies. What’s your advice right now for how they should approach AI and take advantage of it?

Donald:
I’d split that into two types of companies — enterprises and software vendors — with different advice for each.

For enterprises: get started. You cannot afford not to. That means being precise about the projects you experiment with, building sandboxes, developing some AI literacy across your teams, and getting initial projects moving now. Not all of them will make it to production, but you’ll learn a great deal. Don’t wait for something better to come along next year — and remember, your employees are already using AI, so you don’t want to be behind that curve. There’s an old Scottish saying that fits here: go carefully, but go.

Dave:
I like that.

Donald:
Do it carefully, but do it. For software vendors, my advice is different: understand the practice of your users. If you’re an existing vendor, your users have built up a community of practice — a specific methodology for how they work, and your software fits into that methodology. Take Tableau and Qlik, going back to when I was at Qlik — we had genuinely different methodologies, to the point where we often weren’t even in the same rooms at the same events. So understand your users’ methodology, understand how AI fits into their actual daily work, and you’ll succeed. If you try to force users into your way of working, you lose them. You have to work with their way of working and show how AI makes it better.

Dave:
That’s great advice. Last question: where do you hope AI takes us in data and analytics? Not a prediction — what’s the potential you’d like to see realized?

Donald:
Human beings are wonderful, natural analysts. My 90-year-old mother can look at a scatter plot and spot the outlier — she couldn’t write the algorithm to find it, but she can circle it with a sharpie. We’re fantastic at making sense of the world. AI gives us a superhuman capability to do that in an even richer, more complete way. You no longer need deep technical specialization to bring your knowledge of the world to your business and still make decisions that are deeply grounded in how the business actually operates. That’s a fantastic thing for AI to give us.

Dave:
I’m very much on board with that. We’ve been talking about the democratization of data and analytics for our entire careers — I think with AI, we can actually deliver on that promise, finally.

Donald:
We absolutely can.

Dave:
Donald, this has been a great conversation, full of sage advice for our listeners. I really appreciate your time.

Donald:
It’s been delightful to speak with you, as always.

Dave:
Thanks everybody for listening to another edition of the Data-Driven Podcast.