AI-powered analytics: Building data visualizations with natural language vs. SQL

Richard Tomlinson, who leads product marketing for Databricks' business intelligence and analytics products, has spent his career championing technology that tames data to help people make better decisions and fuel innovation. In this conversation, he discusses how natural-language querying for visualizations can finally offer more than SQL, how to make sure AI-powered analytics are trustworthy and how business context is key.

Why natural-language querying for visualizations has underdelivered

Natural-language querying for visualizations has been around for a decade. Why hasn't it lived up to its promise?

Richard Tomlinson: The promise has been around since the early days of search-driven analytics, but it consistently underdelivered because the systems could parse the question but couldn't understand the business.

Earlier generations of NL-to-SQL treated the problem as purely a language-to-schema translation — map words to columns, generate a query and return a chart. That works for simple lookups, but it falls apart in the face of enterprise data complexity. In real life, you may have competing definitions of the same metric across teams, ambiguous joins between tables, business rules that live in people's heads rather than in metadata and no way to decide which source is authoritative when two numbers conflict.

The LLMs weren't the bottleneck; context was. Even as models improved dramatically, they were still guessing at meaning from raw table names and column headers, which is why the same question asked twice could produce two different answers.

What changed isn't just better models; it's the recognition that you need a governed semantic layer underneath the language model that tells the agent what "revenue" means for this team, which table is the certified source, how fiscal years are defined and which filters should always apply. Without that, natural-language querying was a demo that impressed in a conference room but not with real enterprise data.

Generating trustworthy answers for every user

Natural language has opened AI-powered analytics up to everyone. But how do you make sure non-specialists get answers they can trust?

Richard Tomlinson: Enabling more people to ask questions is the easy part. What's hard is making the answers trustworthy enough to act on. Three things have to be true underneath.

First, the data itself has to be governed. You need clean, well-organized data built around how the business actually works, where categories are consistent across the business, each record is clearly defined and there are no duplicate customers or products. It all needs to be secured through fine-grained access controls so every user sees only what they're authorized to see.

Second, the meaning of that data has to be explicit and machine-readable, not buried in tribal knowledge or someone's spreadsheet. You need governed metric definitions, documented business terms, clearly defined primary and foreign key relationships and certified sources, all registered in a system like Unity Catalog where an AI agent can consume them at query time.

Third, there has to be a trust and verification layer, with the ability to trace every AI-generated answer back to the definitions, sources and permissions that produced it. This enables a business user to inspect why the system said what it said, not just what it said. That's the foundation of Genie One: governed data, explicit semantics and inspectable provenance. Without it, you're just giving everyone a faster way to get the wrong answer.

How semantic grounding changes the game

What is semantic grounding and how does it contribute to effective natural language-based data visualizations?

Richard Tomlinson: Semantic grounding is the process of anchoring an AI agent's interpretation of a question to the organization's actual business definitions, relationships and rules, rather than letting the model infer meaning from column names and hope for the best.

In practice, this means that before a single line of SQL is generated, the agent already knows that "net revenue" uses a specific approved calculation, that it should be pulled from a certified finance table rather than a marketing summary, that fiscal quarters start in February and that free-trial users are excluded from retention metrics, for example.

At Databricks, Genie Ontology provides a unified context layer that combines explicitly modeled knowledge — such as metric views, governed pages, domains and certified sources — with context automatically learned from how the organization actually works. That context is extracted directly from dashboards, queries, notebooks and other enterprise assets.

Each piece of learned context, called an Ontology Snippet, carries an authority score based on its source, usage frequency, freshness and certification status. At query time, a ranking system evaluates competing snippets on both relevance and authority, resolves conflicts and delivers only the context the user is permitted to see.

A visualization created using this process isn't just a good-looking chart. It's grounded in the same definitions and logic that the organization's domain experts would use, which makes the difference between a plausible answer and a trustworthy one.

Good governance makes it all possible

What is the role of governance in successfully building data visualizations using natural language instead of SQL?

Richard Tomlinson: Governance is the load-bearing infrastructure that makes natural-language analytics possible at enterprise scale, not a compliance checkbox you bolt on.

When a business user asks a question in natural language, the system must make a series of decisions that shape the answer, including which tables to query, joins to use and metric definition to apply and which rows and columns a specific user is allowed to see. In a SQL world, a trained analyst makes those decisions. In a natural-language world, the AI agent makes them, which means governance must guide every decision the agent makes, not apply after the fact.

Unity Catalog provides that foundation with fine-grained access controls, row filters and column masks enforced per user, certification signals that tell the agent which assets the organization vouches for and deprecation flags that steer the agent away from outdated sources.

Governance also extends to the semantic layer itself. Ontology Snippets follow the same Unity Catalog permissions as the data they describe, so the agent only uses business context the user is allowed to see. With the right governance in place, organizations can open self-service analytics to a much broader audience with confidence because the guardrails are enforced by the system.

Keeping conversational threads on track

How does AI-powered analytics keep track of context in an ongoing conversation, including follow-up questions?

Richard Tomlinson: This is one of the most underappreciated technical challenges in conversational analytics, and it's where most NL-to-SQL systems fail. When a user asks "What were sales last quarter?" and then follows up with "Break that down by region," for example, the system has to understand that "that" refers to the prior query's metric, time window and filters and use that context in its answer.

Genie retains conversational context within a thread by maintaining the conversation state. Table and column metadata, instructions provided by agent editors and context from prior conversational turns all inform how the next question is interpreted. If the follow-up is ambiguous, Genie can ask a clarifying question rather than guessing.

Genie can also draw on past conversations to form answers. A user can reference earlier work ("Remember when we talked about ARR last quarter? Let's drill down from that by region") and the system pulls that context into the current conversation without the user navigating away.

The combination of semantic grounding and conversation filtering keeps answers consistent as conversations get longer. The ontology provides stable definitions that persist across turns, so "revenue" means the same thing in turn five as it did in turn one. Meanwhile, conversation management selects relevant context rather than naively appending everything, which would eventually overwhelm the context window and cause the type of drift users experience in less sophisticated systems.

Bringing Genie into everyday tools with governance built in

How can organizations bring Genie into the apps and AI tools their teams already use, and how are data access permissions enforced?

Richard Tomlinson: Databricks provides multiple ways to integrate. The Genie Conversation API enables developers to embed natural-language data querying directly into their own applications, chatbots and agent frameworks. Through the API, users can ask questions, retrieve generated SQL and query results and get answers in multi-turn conversations that remember earlier context.

Organizations can also use Genie from the AI assistants they already rely on, like Claude, ChatGPT, Microsoft Copilot or Cursor, through the Genie One MCP server. Answers use the same business definitions and data permissions as within Databricks.

Genie Agents can also be embedded in other applications with permissions intact. Queries run on a SQL warehouse using credentials set up by the agent author, but Unity Catalog determines what each end user can see based on their access rights. Row filters, column masks and table-level privileges all apply per user.

The recommended authentication pattern for external integrations is on-behalf-of (OBO) OAuth, where the external assistant acts on the user’s behalf and Genie applies that person's permissions to every request. So two people can ask the same chatbot the same question and get appropriate answers based only on the data they're allowed to see, without special setup for each user.

A raw database connection just hands over data. A governed analytics integration controls who sees which metrics and how they're calculated.

Delivering on the promise of natural-language analytics

Natural-language analytics has long promised that anyone could ask a question and get a data visualization. In reality, the results often weren't reliable enough to act on. As AI takes on decisions traditionally made by analysts, underlying context becomes critical. Trustworthy answers depend on governed data, shared definitions, clear permissions and a way to verify how each answer was generated. Systems also need to keep track of context as conversations continue and bring AI-powered analytics into the tools employees already use. With that foundation in place, organizations can confidently give employees across the business access to analytics they can trust, wherever they work, to help them make smart decisions.

See how Genie turns business questions into governed answers and action.