Agentic Analytics Needs a Control Plane, Not Another Chatbot
AI is changing business intelligence faster than any interface shift since the dashboard. In the last few months the category has gone from "can I ask a question in natural language?" to agents that inspect dashboards, create assets, enrich rows, trigger workflows, and carry context into Slack, documents, applications, and AI assistants.
That shift comes with a governance problem.
When analytics lived mostly inside dashboards, the boundaries were visible. Users opened a report, applied filters, shared a link, exported a CSV, or asked an analyst for help. Modern AI workflows blur those boundaries. A user may start with a dashboard, continue in an AI client, ask an agent to investigate an anomaly, share the result in a team channel, and expect the same business definitions, permissions, filters, and data context to survive the handoff.
Teams will judge the next generation of BI by whether it gives them a trustworthy control plane for analytics context. The number of AI features will matter much less.
From AI features to AI operating models
Vendors across the analytics ecosystem are racing to make BI assets programmable and agent-accessible through MCP, at stages that run from preview to general availability. Sigma bills itself as "the AI runtime for business," with analytics, apps, and agents built on live warehouse data. ThoughtSpot offers a generally available MCP server that connects Claude, ChatGPT, and Gemini to its semantic layer. Tableau has an official MCP server for Tableau Cloud and has announced integrations for Claude, ChatGPT, and Codex. Looker's managed MCP server, in preview, gives agents access to its LookML semantic model. Metabase includes an MCP server for searching, querying, and creating content. Databricks offers Genie MCP servers that enforce Unity Catalog permissions. Omni and Lightdash expose their semantic layers to AI tools over MCP. On October 27, Hex introduced Hex 4.0, now in private beta, which backs core logic and visualizations with code files and treats the notebook as a collaboration surface for people and agents.
Traditional BI treated the platform as the destination: analysts built dashboards, admins governed them, and users went there to look. AI analytics treats analytics as a service other tools can call. People now ask questions in chat, investigate anomalies in an AI assistant, work inside an IDE, and receive scheduled reports in Slack or email. The dashboard still matters, but it is one surface among several.
That raises a new buyer question. Beyond whether a platform has AI, buyers need to know what happens when AI becomes another analytics client. The answer depends on whether the platform can preserve trust outside its original interface.
Context is the product
AI analytics only works when it understands the business context behind a question. That context includes metric definitions, permissions, row-level security, dashboard filters, chart scope, data lineage, query history, and the plain-language vocabulary teams use every day. An agent without it produces confident guesses.
A larger model context window does not fix this. It cannot tell an assistant which revenue definition finance approved, which filters are active on a dashboard, or which rows a user may see.
The semantic layer conversation has become loud for this reason, and every vendor wants to own the place where business meaning lives. A semantic layer defines the vocabulary. Business meaning also lives in dashboards, saved queries, filters, alerts, reports, embedded experiences, and the way teams collaborate around a specific slice of data. Definitions, permissions, state, and durable artifacts together make up the full working language. A vendor that owns the vocabulary but loses the rest still leaves agents guessing.
Preset's approach starts from a practical foundation: Apache Superset, one of the most widely adopted open-source BI platforms. Superset gives teams a transparent, inspectable analytics layer for SQL, charts, dashboards, datasets, and permissions. Preset builds on that foundation with managed reliability, enterprise governance, embedded analytics, and AI-ready interfaces that help teams use their existing analytics context safely.
The conversation is scaffolding, the dashboard is the deliverable
Conversational analytics lowers the barrier to entry and the cost of exploration. A user can ask a follow-up question or investigate an unexpected result without starting a new development cycle. Conversations make poor systems of record, though. Important decisions still need artifacts that teams can review, share, monitor, and revisit.
Dashboards have always played that role. What has changed is where the work happens. A dashboard used to be both workshop and showroom. Now analysis often starts in a chat thread, an IDE, or an agent task, and the dashboard is where the result lands. Coding agents offer a useful comparison: the conversation is scaffolding and the pull request is the deliverable. In analytics, the governed dashboard plays the part of the pull request.
Dashboard state belongs to the context too. A dashboard filtered to a region, date range, customer segment, or product line represents a question. When the state is lost as the link is shared, the recipient has to reconstruct the question by hand, and an agent that sees only the dashboard title may answer a different question than the one the user meant.
Preset now supports dashboard links that encode filter values in the URL, so recipients open a pre-filtered view. Through Preset MCP, supported clients can also update dashboard layout, styling, filters, refresh frequency, and tags. In an agentic workflow, the exact slice of data is part of the prompt. In a Slack thread or customer-facing application, it separates "look at this dashboard" from "look at the exact business situation I am talking about."
MCP needs a front door
The Model Context Protocol is becoming a common way for AI clients to connect with business systems, including BI platforms. That helps openness, because analytics teams can make trusted data and dashboard context available to a growing ecosystem of AI tools.
An endpoint alone does not make a strategy, though. As more AI clients reach analytics, enterprises need to decide which clients are approved, what they can access, and how those connections fit existing governance models. Without that, MCP becomes another unmanaged integration surface that works in a demo and carries risk in production.
Preset's per-workspace MCP connection allowlisting lets admins control which AI clients can connect to Preset data. It is available through your customer success manager. Governance works best when it starts at the connection point, before an agent has access.
Permissions also need to follow the user across clients. The same identity and platform rules that govern the web UI apply to AI actions through Preset MCP, and write tools stay hidden from users who lack write permission. Rate-limited API responses include standard X-RateLimit headers, so agents and scripts can see their remaining capacity, and MCP tool calls recover when a pooled database connection drops between requests.
Open foundations reduce AI lock-in
Vendors are answering AI complexity with a "one runtime" story. A single product can reduce integration work and make governance easier inside its own walls, and some teams will find that appealing. Agentic analytics is still developing, though. Model providers, interfaces, protocols, and user habits will keep changing, and the more a platform absorbs metrics, prompts, actions, app logic, and institutional knowledge, the more expensive lock-in becomes.
Preset takes a different path. Built on Apache Superset, it lets organizations take part in the agentic analytics era without surrendering every asset to a closed runtime. Teams can use governed dashboards, SQL-native workflows, embedded analytics, and approved AI-client access while keeping their analytics foundation open and inspectable. Managed Private Cloud runs on AWS, Azure, and Google Cloud, so organizations can align analytics with their own cloud and security boundaries.
An open control plane centralizes governed context while the interfaces around it keep changing. When agents can act on analytics, teams need to know what those agents see, where the definitions live, and how to change the system without being trapped by one vendor's runtime.
Four questions for evaluating agentic analytics
As teams evaluate the next wave of BI and AI capabilities, four questions matter more than whether a product has a chat box:
- Can every interface use the same governed context? Metrics, datasets, filters, and business definitions should not fragment across assistants.
- Do permissions follow the user? Agent access should inherit real identity and authorization rules without privilege escalation.
- Does agent work become a durable asset? Useful outcomes should be inspectable, shareable, and maintainable after the conversation ends.
- Can the architecture adapt? Teams should be able to change AI clients, models, clouds, and workflows without rebuilding their analytics foundation.
Agents change where and how analytics work happens, and the need for governance, shared meaning, and operational reliability stays. Preset is built around that need, with an open Superset foundation, managed enterprise reliability, governed permissions, MCP client allowlisting, and dashboard state that travels with the question.
Agentic analytics is already here, and the open question is whether your analytics platform can govern it.
Further reading
- Preset MCP: AI That Doesn't Just Answer, It Builds
- Preset MCP: Analytics Your AI Agent Can Build, Not Just Read
- Beyond the Demo: Seven Controls for Production-Ready AI Analytics
- The Semantic Layer Is Back. Here's What We're Doing About It.
- Open Source BI Governance & Security Without Vendor Lock-In
