Six months ago, I started a series called “Talking to Data.” The argument was straightforward but, at the time, felt early: open protocols like MCP would shift AI integration from vendor lock-in to governed orchestration. Semantic contracts would separate meaning from mechanics, making it possible for AI to navigate enterprise data without guessing what “revenue” means. The conversational model would move analytical initiative from IT to the business. And the human side of the equation, the skill of actually thinking with data through an AI, would matter more than anyone was acknowledging.
The series ran five posts. The response was strong, the pushback was useful, and the ideas found traction in conversations I did not expect.
Then the rest of the market showed up.
What Changed While We Were Talking
In March 2026, Gartner’s Data & Analytics Summit made a prediction that would have been surprising a year ago but now reads as obvious: by 2028, 60% of agentic analytics projects relying solely on MCP will fail due to the absence of a consistent semantic layer. They went further. Universal semantic layers, they said, should be treated as critical infrastructure on the same level as data platforms and cybersecurity. Organizations that prioritize semantics in AI-ready data will increase agentic AI accuracy by up to 80% and reduce costs by up to 60%.
It should be mentioned that Juan Sequeda notes that 44% of data and analytics leaders have already implemented semantic layers, with an additional 48% planning to by 2027.
In January 2026, the Open Semantic Interchange specification was finalized as a vendor-neutral standard for sharing business context between semantic layers and AI consumers. Snowflake, Salesforce, dbt Labs, Atlan, Alation, Mistral AI, and ThoughtSpot signed on as partners. dbt Labs open-sourced MetricFlow under Apache 2.0, explicitly framing it as critical infrastructure for AI accuracy. The semantic layer stopped being a feature inside BI tools and became an architectural concern in its own right.
MCP itself went from niche protocol to industry backbone. By April 2026, it was implemented on more than 10,000 enterprise servers with over 97 million SDK downloads. Anthropic donated it to the Linux Foundation’s Agentic AI Foundation, with OpenAI and Block as co-founders. Google, Microsoft, and AWS adopted it. The protocol that felt experimental when I wrote my first post is now the standard connector between AI agents and enterprise data.
The term “agentic analytics” entered the mainstream. Gartner published a Market Guide for it. Vendors rushed to position their products around it. LinkedIn filled with posts about semantic layers, context engineering, and AI-ready data architecture. The infrastructure argument that my series spent five posts building is now, effectively, consensus.
That is good news. It is also the beginning of a problem.
Where the Consensus Stops
Read through the current discourse carefully, the analyst reports, the vendor announcements, the LinkedIn thought leadership, and you will notice something. The conversation is almost entirely about plumbing.
Build the semantic layer. Deploy MCP. Govern access. Connect the sources. Standardize the definitions. This is important work. I spent three posts of this series on exactly these points. But the current conversation treats these as the finish line rather than the starting conditions.
Nobody is asking what changes about how people actually work with their data once the plumbing is in place.
The assumption, rarely stated but everywhere implied, is that once you connect the semantic layer to the AI agent through the right protocol with the right governance, the value follows automatically. The infrastructure enables. People consume. The intelligence layer does its job.
It is my opinion that assumption is wrong, and it is wrong in a way that the Gartner failure predictions should have made visible but did not. When they say 60% of agentic analytics projects will fail without a semantic layer, they identify the most obvious failure mode: AI without meaning. But there is a second failure mode that the current discourse has not named: Meaning without Inquiry.
You can build a flawless semantic layer. You can connect to every source through governed MCP servers. You can define every business term with machine-readable precision. And the whole thing can still produce mediocre insight if the people using it do not know how to think with it.
What "Talking to Data" Got Right and What It Simplified
The phrase “Talking to Data” served a purpose. It made a new interaction model accessible. It gave people a mental image they could grasp to sit down, ask a question, get an answer from your actual data instead of a pre-built report. The simplicity was the point. It opened a door.
But I have spent enough time on the other side of that door now to know the phrase undersells what actually happens.
You do not talk to data. Data does not listen, interpret, or respond. What you engage with is an intelligence construct, an AI that has been given access to your data through governed connections and grounded in semantic definitions that tell it what your business terms mean. The AI is not a transparent window between you and your data. It is a reasoning participant. It carries context from one exchange to the next. It makes interpretive choices about which sources to query and how to structure the response. It synthesizes across domains in ways that feel seamless but involve real analytical judgment, sometimes sound, sometimes not.
When the interaction works well, it is not because the human asked a good question and the AI fetched a good answer. It is because something more collaborative happened. The person brought direction, business context, and the instinct for which threads matter. The AI brought navigational capability, access to multiple sources, and the ability to maintain an analytical frame across a sustained exchange. Together, they produced insight that neither could have reached alone.
That is not talking to data. That is thinking with an intelligence layer that happens to have access to data. The distinction matters because it changes what organizations need to invest in if they want this to work.
The Missing Layer
The current market conversation has three layers: the data layer, the semantic layer, and the protocol layer. Build all three and you have the infrastructure for AI-powered analytics.
What is missing is a fourth layer that sits above all of them: the inquiry layer. This is where humans meet intelligence. It is where questions get formed, where analytical threads get steered, where context gets built across exchanges, where results get interrogated rather than accepted, and where the distinction between productive and unproductive use of the entire stack gets determined.
The inquiry layer is not a technology. It is a capability. It lives in how people approach the interaction, what patterns of questioning they use, how they maintain and direct an analytical thread, whether they verify or simply consume, and whether they bring enough business context to the exchange for intelligence to do meaningful work.
I wrote about this in the fourth post of the original series, cataloging the patterns that distinguish productive data conversations from frustrating ones. What I have come to understand since then is that this is not a secondary concern to be addressed after the infrastructure is built. It is the layer that determines whether the infrastructure investment pays off.
Gartner’s prediction about semantic layers is correct: without them, AI agents get enterprise questions wrong systematically. But even with them, the quality of what comes back depends on the quality of what goes in. Not garbage-in-garbage-out in the traditional sense of dirty data, but something more subtle: shallow inquiry in, shallow insight out. The AI will respond to whatever you give it. If you give it a vague question, it will give you a vague answer. If you give it a sequence of well-directed exchanges that build context across sources, it will produce something you could not have gotten from any dashboard or report.
The infrastructure does not determine this. The human does.
Why This Is Not a Training Problem
The instinct, when someone says the human side matters, is to propose a training program. Teach people how to write better prompts. Run a workshop on conversational analytics. Create a playbook.
Those instinctive approaches underestimates what is actually required. This is not about learning a new interface. It is about developing a different relationship with analytical thinking itself.
For decades, business intelligence trained people to be consumers. Open the dashboard. Read the chart. Drill the pre-built path. Request a new report if you need something different. The entire interaction model assumed that someone else had done the thinking about what questions to ask and what data to assemble. The business user’s job was to interpret what was presented, not to direct the inquiry.
The conversational model inverts this. The person asking the question holds the initiative. They decide what to explore, where to go next, when to shift direction, and when to challenge what comes back. The AI navigates, but the human steers. That requires a kind of analytical fluency that most organizations have not cultivated because the previous generation of tools did not require it.
It is closer to the difference between reading a research paper and conducting the research yourself. Both involve engaging with information. But one follows a path someone else constructed, and the other requires you to construct the path as you go, making choices about direction at every step.
Organizations that treat this as a prompt-engineering problem will get prompt-engineering results: marginally better queries producing marginally better outputs. Organizations that recognize it as a cognitive capability, one that develops through practice and depends on business knowledge more than technical skill, will get something qualitatively different.
Where This Series Goes from Here
The original five posts built a foundation: architecture, interaction, governance, human capability, and organizational orientation. The market has since validated the infrastructure arguments in ways I did not expect to happen so quickly. The semantic layer is now consensus. MCP is now standard. The governance conversation is maturing rapidly.
What has not matured is the conversation about what happens above the infrastructure. How do people actually use these systems to think through problems? What does it mean when the intelligence layer is not just a consumer of data but a participant in how meaning gets constructed? What happens to the analytical frameworks we have relied on for decades when the tool navigating them can shift between perspectives fluidly in ways that pre-built reports never could?
These are the questions I intend to explore in the posts that follow. Not because the infrastructure does not matter, it does, profoundly, but because the infrastructure conversation is now in capable hands across the industry. The conversation that still needs development is the one about the nature of the interaction itself: what it demands of us, what it makes possible, and why the phrase “talking to data” was always just the beginning of a much larger shift.
The market arrived at the plumbing. The real work is what we build on top of it
This is the continuation of a series exploring how AI and conversational interfaces are reshaping data architecture and business intelligence. Previous posts are available at insightsindata.com.
© 2026 Paul Nevill / Insights in Data
