Many teams assume they’re doing agentic analytics because they support text-to-SQL or chat-based BI. And while conversational BI does change how questions are asked, it doesn’t go so far as to change how decisions are made.
It’s like transitioning from GPS to autopilot. With GPS, you ask for directions and decide whether to follow the route provided. That’s conversational BI.
With autopilot, the system determines the route and adjusts speed, changes lanes, and executes turns in real time. That’s agentic analytics.
Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024. This signals a fundamental shift in what we’re asking data systems to do.
Conversational BI: Faster Questions, Same Responsibility
Conversational BI is still fundamentally reactive. A human remains in the loop to interpret results, validate correctness, and decide what action (if any) to take. Errors are usually caught before they propagate.
This works because answers are read, not executed. If there are inconsistent metrics, a business analyst can spot when “revenue” doesn’t match their expectations. They can reconcile discrepancies before making decisions.
As with GPS, the human remains the control plane.
What Actually Changes with Agentic Analytics
Agentic analytics introduces three structural changes that fundamentally alter the risk profile:
- From answers to actions: Agents explain results and trigger downstream steps. Whether a budget adjustment or a pricing change, the output is operational impact.
- From one-off queries to workflows: Agents execute multi-step reasoning across tools, systems, and time. They chain decisions together, with each step building on the last. Errors compound.
- From advisory output to operational impact: When an agent updates a forecast or reallocates resources, the consequences are immediate and systemic.
The cost of being wrong moves from confusion to consequences.
Why Naive Conversational Architectures Break at the Agent Layer
Even as adoption increases, trust remains selective. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of next year due to escalating costs, unclear value, or inadequate risk controls.
When projects move forward, it’s with caution. PwC’s AI Agent survey showed that business leaders are confident in delegating tasks such as data analytics (38%), performance improvement (35%), and day-to-day collaboration with human colleagues (31%). Trust dropped sharply for high-stakes use cases, including financial transactions (20%) and autonomous employee interactions (22%). The report cites a “growing need for role-specific governance and transparency to guide when and how AI agents are introduced into sensitive workflows across the enterprise.”
What AI agents need is a shared semantic foundation. Without it, agents have to infer metric definitions, time windows, business logic, and access rules. Those inferences are implicit, non-deterministic, and inconsistent across agents and sessions.
The result is that two agents can reach different conclusions from the same data. Errors compound as workflows chain together. And when something goes wrong, there’s no clear way to trace why the agent made the decision it made.
Why Agents Require a Semantic Layer (Not Just Better Prompts)
Trying to prompt your way to business logic is futile because the data and the questions are always changing. Prompting doesn’t provide central meaning or lineage.
For agents, semantics are the control infrastructure. A semantic layer provides the foundation agents need to operate reliably:
- Deterministic definitions. Metrics mean the same thing every time, for every agent. “Revenue” doesn’t shift based on prompt phrasing or model behavior.
- Governed access. Agents only see what they’re authorized to act on. Data security and business logic are enforced before the agent sees the data, not after.
- Reusable business logic. Logic lives outside the prompt and outside the LLM. It’s versioned, auditable, and consistent across systems.
- Auditability. Decisions can be traced back to certified definitions. When an agent adjusts a budget, you can explain exactly what data it used and why.
You wouldn’t run autopilot without calibrated sensors. That’s what a semantic layer provides for AI agents: calibrated, certified inputs that the system can trust to make automated decisions.
AtScale as the Foundation for Agentic Analytics
At AtScale, we’ve spent 13 years building the semantic control plane that agents require. Our universal semantic layer standardizes metrics once and reuses them everywhere, across BI tools, AI agents, and data applications. Our open semantic modeling language (SML) is consumable by agents via standards such as the Model Context Protocol (MCP), which injects governed business context directly into AI workflows.
This is a tool-agnostic architecture designed from the ground up to support both humans and agents operating from the same source of truth. So when an agent takes an action, you can explain why and truly scale governance.
What Enterprises Need to Rethink Before Deploying Agents
Before rolling out agentic analytics, ask these four questions:
- Where do our metric definitions actually live? If they’re scattered across BI tools and documentation, agents will infer their own versions.
- Are those definitions executable, or just documented? Agents need machine-readable logic to generate live queries.
- Can we guarantee consistency across tools, agents, and time? If not, automation is a liability.
- Can we explain why an agent took an action six months later? Without semantic lineage, troubleshooting and auditing become impossible.
If the answer to any of these questions is no, you need to rethink your data foundation.
Agentic Analytics Is an Architecture Decision
Conversational BI is an interface upgrade. Agentic analytics is an operating model change.
Going back to the GPS analogy: you wouldn’t trust autopilot in a car without lane sensors, speed governors, and collision detection. Those are the control systems that make automation safe. The same applies to agentic analytics. Once AI acts on data, semantics become mandatory infrastructure.
While nearly half of agentic AI projects fail due to inadequate controls, the teams that invest in semantic governance first will actually realize returns from automation. They’ll be able to scale agents across high-stakes workflows, such as financial transactions, resource allocation, and autonomous operations, while competitors remain stuck delegating only low-risk tasks.
Closing the gap between pilot and production requires a data foundation that can support the decisions you’re willing to automate.
SHARE
2026 State of the Semantic Layer