Every decade, business gets sold a new layer of technology, and every decade most of the promised value fails to show up. We did it with ERP. We did it with business intelligence. Now we’re doing it with AI, and the boardroom slides look suspiciously like the last two times.
I’ve spent a little over twenty years watching this pattern repeat. I helped build the tools companies used to try to make BI pay off, then spent more than a decade implementing them for clients. AI is stalling for a lot of reasons right now, and I won’t pretend one article can cover them all. But one of those reasons is a rerun – one I watched play out in the BI era.
Where BI bogged down
BI didn’t underdeliver because dashboards were a bad idea. It struggled because the meaning underneath the dashboards was almost never captured in a reusable way.
Think about what a word like “revenue” actually requires. Someone has to decide which systems of record are authoritative, which transactions count, which refunds get subtracted, and how the fiscal calendar lines up. In most shops, that logic got rebuilt from scratch for every new dashboard and every new query, because the tools gave you nowhere durable to store these definitions in a reusable, machine-readable format. The definitions lived in a few people’s heads, and every request routed back through those same heads.
That was the reimplementation tax, and it was mostly invisible. You’d ask for a report, wait days or weeks while someone hand-coded the same definitions again, and not question the delay because you’d never seen it work any other way. That delay was often synonymous with failure, however, because “days or weeks” didn’t keep up with business realities. You quickly learned not to ask, and instead reverted to the same guesswork and spreadsheets that BI was supposed to replace.
The fix already existed, and it had a name: the semantic layer – the place where you centrally define what your business data means, so it gets reused everywhere instead of rebuilt every time. Some BI vendors took this seriously; Power BI built the most developed version of it, while others came to the idea late or never really embraced it. But a capable tool was only ever half the battle. Owning one didn’t mean you had a good semantic layer, because the discipline was subtle and widely misunderstood, and a striking share of deployments were just raw data dumps with dashboards bolted on top.
A slow tax for BI, a wall for AI
Here’s why this matters more now than it ever did, at least for one particularly exciting kind of AI: the agents you want reasoning over your own business data. A missing semantic layer was survivable in the BI era because the thing waiting for an answer was a human, and humans can wait.
An agent asking a real business question can’t. If every question it asks has to funnel back through the one analyst who knows what “gross margin” really excludes, you don’t have automation. You have a very expensive bottleneck with a chatbot on the front.
So people try the obvious shortcut: point the LLM at the raw databases and let it write its own queries. This fails three ways at once. It’s slow, because even a fast model reinventing your business logic still makes a person wait; it’s expensive, because that reasoning burns a startling amount of compute on every question; and it’s wrong, confidently and inconsistently wrong, because the model hits dozens of coin-flip judgment calls about your business and guesses differently every time you run it.
The BI era let humans quietly absorb that ambiguity. The AI era can’t afford to.
The fix is a layer, and for once, the entire industry agrees
The answer isn’t a smarter model. It’s the same layer BI needed all along, except now it’s non-negotiable.
You don’t have to take my word for it, because the whole industry arrived at this conclusion at once. In September 2025 a coalition of companies including Snowflake and Salesforce launched the Open Semantic Interchange, an open standard for defining semantic layers. Even Tableau – long a champion of connecting straight to your data rather than modeling it first – published a post arguing that “the agentic future demands an open semantic layer”. When a vendor known for the opposite instinct changes its tune, the shift is worth noticing.
By early 2026 the roster had grown to include AWS, Databricks, and most of the BI field. When competitors who agree on nothing suddenly agree on one thing, it usually means the thing became impossible to avoid. The semantic layer was never a new idea. It just became the thing this kind of AI can’t run without.
Turning the tables: AI as BI’s savior
Now the part that cuts against my own headline, and it’s delightfully valuable. BI had a second failure the semantic layer alone never fixed. But AI backed by a semantic layer can eliminate this second bottleneck and finally deliver on the full original promise of BI.
Even in the shops that did the semantic layer right, the last mile stayed brutal. When a business user had a real question, they had to guess which dashboard might answer it, discover that none did, file a request, and wait. If a dashboard did exist, they had to learn its filters and its particular idea of how to slice the world, then chain five sub-questions together by hand to get where they were going. Most people gave up and decided on gut.
A natural-language AI front end sitting on a real semantic layer collapses that last mile. You ask in plain English, and the agent picks the right measures, applies the right filters, and follows its own sub-questions to an answer you can actually trust, because the meaning underneath is certified rather than improvised. That is the value BI always promised and rarely delivered.
So a single investment pays off twice. Skip the semantic layer and this kind of AI fails loudly and expensively. Build it, and AI finally cashes a check BI wrote fifteen years ago.
The people who can fix this already work for you
Which brings me to the most overlooked assets in your building. The fix here doesn’t start with a new hire, a Chief AI Officer, or a governance committee. Instead, it starts with people you already employ.
They’re hiding amongst your spreadsheet power users, your BI analysts, the people who know which of two systems to trust when they disagree about who a customer is. They’ve been authoring your semantic definitions for years, in spreadsheets and queries and their own memory, without anyone calling it that.
Look around the office and you can name them. It’s the finance manager who knows which expense categories don’t really count against gross profit, the ops lead who can explain why the customer ID in your CRM doesn’t match the one in billing, the person with the only working definition of “active customer.” Right now that knowledge lives in their heads, and their heads are not available to answer an agent’s question at 2 a.m. on a Tuesday.
The actual assignment
None of this is the whole of your AI strategy; plenty else has to go right. But of the reasons your best data use cases will stall, this is the one we’ve seen before, and the one you’re strangely well-equipped to fix. It doesn’t take the most models or the biggest budget. It takes capturing your own meaning in a layer, and involving the people who have quietly understood it all along.
That’s less thrilling than the demos, but a lot more durable. AI didn’t create either of the BI-era problems I’ve talked about here, but it’s going to be the reason both get fixed.
