The future of analytics isn't moving data — it's moving meaning

The modern data stack solved the wrong problem — and nobody noticed, because it solved it so well.

For the better part of two decades, the best minds in data engineering focused on one thing: getting data from where it lives to where it needs to be, reliably, at scale, and without breaking everything downstream. They largely succeeded. The infrastructure we built — the pipelines, the warehouses, the transformation layers — is genuinely impressive. But AI has exposed a gap we were always too busy to notice.

Clean data is not the same as understandable data.

That distinction is the defining challenge of the next era in analytics. And if you're building a data strategy today without grappling with it, you're optimizing the wrong layer.

What the modern data stack was actually built to do

Let's be honest about what the classic stack optimized for. ETL pipelines existed to move data. Cloud warehouses like Snowflake and BigQuery existed to store it cheaply and query it fast. dbt existed to transform it reliably. Orchestration tools like Airflow and Dagster existed to make sure none of those steps fell over at 3 am.

Success was measured in pipeline reliability. Query speed. Cost per terabyte. Freshness windows. The whole ecosystem was engineered around a single assumption, so baked-in that nobody thought to question it: a human would be on the other end, interpreting the results.

Dashboards were the presentation layer for human cognition. A chart doesn't need to understand what "monthly recurring revenue" means — it just needs to render a number. The human looking at it brings the context. They know which revenue lines to include, which to exclude, whether to count trials, how to handle upgrades and downgrades. The chart just draws the picture.

That assumption held for a long time. It was a reasonable design choice. And it produced an era of genuine innovation that compressed what used to take months of engineering into days of configuration.

But it left a gap. A big one.

ETL isn't dead — it's just no longer sufficient

I want to be careful here, because the nuance matters.

Data pipelines are not the problem. Ingestion, cleaning, transformation, normalization — all of that still needs to happen. The modern data stack didn't fail. It succeeded at exactly what it set out to do. The problem is that "clean data" was never the finish line. It was always the prerequisite.

You can have perfectly orchestrated pipelines, a meticulously modelled data warehouse, and dbt tests passing green across the board — and still have three different people in a Monday morning meeting quoting three different revenue numbers. You can have beautiful infrastructure and total semantic chaos.

I've seen this at companies that would describe themselves as data-mature. The pipelines work. The dashboards load. And yet nobody quite agrees on what "active customer" means. Is it someone who logged in this month? Placed an order? Has a paid subscription? The data is clean. The definition is contested.

This is the gap the modern data stack was never designed to close. It moved data. It didn't encode meaning.

AI changes who — and what — consumes analytics

Here's where the stakes get much higher.

For most of the dashboard era, the consumer of analytics was a human. A business leader squinting at a chart. An analyst building a pivot table. A founder checking their metrics before a board meeting. Humans are remarkably good at filling in context gaps. We bring domain knowledge, institutional memory, and the ability to ask a colleague, "Wait, does this number include refunds?"

AI systems cannot do that. Not reliably. Not safely.

The "user" of analytics is no longer just a person. It's a business leader, yes — but it's also an embedded copilot, an autonomous agent, a workflow automation, another AI system calling an API. These consumers don't bring context. They need context to be encoded in the data itself.

When an AI system encounters ambiguity — two metrics with the same name and different definitions, a field with no description, a KPI calculated differently in two different tools — it doesn't pause and ask for clarification. It guesses. It picks one interpretation and runs with it. In a business setting, that's not a minor inconvenience. That's a confident, fluent, completely wrong answer delivered to an executive at speed.

The problem isn't that AI is bad at analytics. The problem is that we're feeding AI systems data infrastructure that was designed for human interpretation. We're asking machines to fill in the meaning gaps that humans used to fill themselves.

Level up data-driven decision making

Make metric analysis easy for everyone.

Get Started Now

The evolution of the stack, honestly told

If you zoom out, the history of business analytics is a story of solving one layer at a time.

Early BI solved data access. Before data warehouses and ETL pipelines, getting data out of operational systems was a project in itself. The first wave of the stack solved that. Data became queryable.

The modern data stack solved transformation at scale. ELT, dbt, cloud warehouses — this wave made it possible for companies with small data teams to do what only large enterprises with dedicated infrastructure could do before. Transformation became reliable, version-controlled, testable.

Operational analytics solved activation. Reverse ETL tools pushed insights back into the operational systems where work actually happens — your CRM, your marketing automation, your customer success platform. Data stopped being a read-only archive and became something you could act on.

Each wave solved a real problem. Each wave also deferred a harder one.

The AI analytics era needs to solve understanding. Not just moving data, not just transforming it, not just activating it — but encoding what it means, consistently, in a way that both humans and machines can rely on.

Semantic layers and metrics layers are becoming infrastructure

A few years ago, semantic layers and metrics layers were considered sophisticated additions to a mature data stack. Nice to have. Worth building eventually. The kind of thing you'd get to once the pipelines were stable and the warehouse was clean.

That framing is obsolete.

Semantic layers — tools like the dbt semantic layer, Cube, and the metrics layers being built into analytics platforms — encode something the rest of the stack has always left implicit: shared definitions, business logic, dimensional relationships, governance rules, lineage, time context. They answer the question: what does this number actually mean, and how was it calculated?

That used to be a question for the analyst. Now it's a question for the AI.

When an AI system queries your data — whether through a natural language interface, an agentic workflow, or an API call — it needs those definitions to be explicit, not assumed. It needs to know that "revenue" in your business means recognized revenue, net of refunds, in the currency of the customer's contract, calculated on a monthly basis. It needs to know that "churn rate" is calculated differently for your enterprise segment than your self-serve segment, and why.

Without that, every AI interaction starts with the same implicit negotiation: "Please explain your business logic again." And every time you do, you're one misunderstanding away from a decision made on bad data.

Semantic consistency isn't a modeling preference. It's the foundation that makes AI trustworthy in a business context.

Analytics systems are becoming machine-consumable

The implications of this shift go deeper than tooling choices.

Agents will query trusted metrics directly — not raw tables. The future isn't an AI system that writes SQL against your warehouse and hopes for the best. It's an AI system that calls a governed metric, with a known definition, a certified calculation, and a clear lineage. The metric is the API. The semantic layer is the contract.

Workflows will trigger from semantic conditions. Not "when this column exceeds this threshold," but "when customer health score, as defined by our success team, drops below 70 for a paying account." The condition carries meaning, not just a number.

Copilots will reason over business metrics. The best AI assistants won't be the ones with the biggest models. They'll be the ones with the best business context — the ones that know your metrics, understand your segments, and can answer "why did ARR drop last quarter" without hallucinating a cause.

This is what MCP-style systems are beginning to make possible. Exposing contextual capability, not just raw functionality. Trusted metrics, definitions, governance rules, dimensional hierarchies — packaged in a way that AI systems can consume reliably. The analytics platform becomes a context platform.

That changes how these tools need to be designed from the ground up. A dashboard builder that happens to have an AI chat widget bolted on is not an AI-ready analytics platform. The architecture has to be different. The data model has to be different. The way meaning is encoded has to be intentional, not incidental.

Where this goes from here

I don't have all the answers. But I have a clear enough view of the direction to make some predictions I'm willing to stand behind.

Metrics become APIs for AI. The governed metric — with its definition, its calculation logic, its certification status, its dimensional context — becomes the primary unit of exchange between analytics systems and AI consumers. Raw tables become an implementation detail. Meaning becomes the interface.

Semantic layers become standard infrastructure. Right now, a well-built semantic layer is a competitive advantage. Within five years, it will be table stakes. Every serious analytics stack will have one, the same way every serious analytics stack has a transformation layer today. The question won't be whether to build it — it'll be which approach fits your scale and your team.

Dashboards become one interface among many. They won't disappear. Humans still need to see things. But dashboards will stop being the primary way organizations consume analytics. Chat interfaces, embedded copilots, automated alerts, agentic workflows — these will carry as much or more of the analytical load. The dashboard becomes a validation surface, not the destination.

Business context becomes a competitive advantage. Companies that invest in well-structured semantic layers — clear metric definitions, governed metrics, documented lineage — will adopt AI faster and more safely than companies that don't. The quality of your AI-driven decisions will be directly proportional to the quality of your business context. This is a moat that compounds over time.

Analytics platforms evolve into context platforms. The winners in the next generation of analytics won't be the tools with the prettiest charts or the fastest queries. They'll be the tools that best encode organizational meaning — and make that meaning available everywhere decisions get made. The question shifts from "can you visualize this?" to "can you be trusted?"

The modern data stack gave us clean data at scale. That was a genuine achievement, and the teams who built it deserve real credit.

But clean data was always a means to an end. The end was always understanding — confident decisions made by people who trust what they're looking at. AI is forcing us to close the gap between those two things. Not by moving more data, faster. By encoding what the data means, once, in a place that both humans and machines can rely on.

That's the work. It's harder than building pipelines. It requires organizational alignment, not just engineering. It requires you to have the uncomfortable conversations about what your metrics actually mean — and to write the answers down in a place that doesn't forget them.

But it's the work that makes everything else trustworthy.