AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。
What is AI analytics? Why it only works on governed data | Databricks Blog Skip to main content AI analytics isn't BI with a chatbot bolted on. It shifts who does the analytical work, from a human navigating dashboards to a system that can understand intent, investigate, and increasingly act on its own. Trustworthy AI analytics depends on four things: governed data, shared business context, permission-aware execution, and verifiable evidence. Better models alone don't get you there. Fragmented business definitions were always a BI problem. AI turns it into a bigger one, because the system now has to resolve organizational disagreements on its own, and the wrong answer can look completely reasonable. AI analytics is the practice of applying artificial intelligence and machine learning to data analysis, enabling systems to surface patterns, generate insights, and answer questions in natural language without manual query-building. It shifts analytics from retrospective dashboards toward automated, conversational, and predictive workflows that a broader set of business users can act on directly. Every BI vendor can demo a natural-language query today. Ask a question, get a chart, look impressive in a room. The harder question, the one that separates a category-defining platform from a good demo, is what happens when that same system meets a real enterprise: conflicting metric definitions, five tables that could plausibly answer the question, and a user who technically shouldn't see half the underlying data. Richard Tomlinson leads product marketing for Databricks' business intelligence and analytics products. In this conversation, we talked through what changes when BI becomes AI-native, why governance turns out to matter more with AI rather than less, and what a data leader should be asking vendors instead of which foundation model they use. How is AI analytics different from traditional BI? BI has always been able to tell you what happened. What's actually different about AI analytics, and why is it more than just BI with a chatbot bolted on? Richard Tomlinson: Good BI has always helped people understand more than simply what happened. The real limitation is that traditional BI generally requires a human to navigate the analytical path. Someone has to decide which dashboard to open, which filters to apply, which dimensions to drill into, what follow-up query to run, and often when to involve an analyst. AI analytics changes who does that work. Instead of presenting a predefined view of the data, an AI-powered system can understand the intent behind a business question, determine what data and business context are relevant, generate and execute the necessary queries, and increasingly conduct a multi-step investigation on the user's behalf. That last part is where agentic analytics becomes important. If I ask, "Why did margin fall in the Northeast last quarter?", an agentic system doesn't have to translate that into one SQL query and return a chart. It can form hypotheses, investigate product mix, discounting, customer segments, or costs, learn from each result, and synthesize the evidence into an explanation. So the evolution is roughly: Dashboards: show me what we already decided to measure. Conversational analytics: let me ask a new question. Agentic analytics: investigate the question for me. That's a much more fundamental shift than putting a chat box on a dashboard. So if dashboards show what happened, what does that investigation surface? What can AI-powered analytics tell a decision-maker that a dashboard never could? Richard Tomlinson: The biggest change is moving from observation toward investigation. A dashboard might tell a retailer that gross margin is down three points. That's useful, but it still leaves the decision-maker asking why. An AI-powered system can investigate whether the decline is concentrated in particular stores, products, or customer segments, test whether discounting changed, determine whether product mix shifted, compare the timing against promotions or supplier changes, and synthesize those findings into an explanation. The valuable output isn't another chart. It might be something closer to: margin is down primarily because a higher proportion of sales shifted toward two heavily discounted product categories in the Northeast, while unit volume and acquisition remained stable. That's much closer to the information a decision actually requires. I'd describe the progression as: what happened, what changed, why did it change, what should I investigate or decide next. And I'd stop short of implying that every AI analytics system can autonomously make or execute the decision itself. The meaningful near-term advance is dramatically reducing the analytical work required to get from a signal to a well-supported decision. What does it mean for AI to interpret data semantically? You've talked about AI analytics interpreting data semantically. What does that mean in practice? Richard Tomlinson: AI analytics absolutely still queries data. The difference is that it needs to understand what the data means before deciding how to query it. A database can tell an AI system there's a column called net_rev, another called bookings, and a relationship between two tables. It can't tell it what the company means by "revenue," whether executives use booked or recognized revenue, which fiscal calendar applies, which customer hierarchy is authoritative, or whether cancelled orders should be included. Those are semantic questions. They're the business meaning that sits between a user's language and the physical data. In practice, a strong AI analytics system needs governed definitions of metrics and KPIs, dimensions and relationships between business entities, the terminology and synonyms the business uses, business rules and calculation logic, authoritative sources, permissions and governance, and often institutional knowledge captured from how people have historically used the data. That context is what lets the system translate "How are enterprise renewals performing?" into the right interpretation of enterprise, renewal, and performing before it ever constructs a query. This is why semantics become more important with AI, not less. How do you make AI-generated analytics trustworthy? If a data leader asked you point-blank, "how do I make AI-generated analytics trustworthy?", what's your honest answer? Richard Tomlinson: Don't start with the model. Start with the foundation the model is allowed to reason over. I think trustworthy AI analytics comes down to four things. First, trustworthy data: the AI needs governed access to the same enterprise data you'd trust for an executive report, not copies exported into a separate AI environment. Second, trustworthy business context: shared definitions for the metrics, entities, and terminology that matter to the business. Giving an LLM a database schema isn't the same thing as teaching it how the business works. Third, trustworthy execution: governance and permissions need to carry through from the user to the data. An AI system shouldn't circumvent existing access controls just because someone asked a question in natural language. Fourth, verifiability: users and data teams need evidence they can inspect, the source data, the calculations or queries behind the answer, citations where appropriate, and a way to evaluate answers systematically against known-good questions. That last point matters even more as analytics becomes agentic. A single bad query produces one wrong answer. An agent can make multiple analytical decisions during an investigation, so you need mechanisms for grounding and validating that whole process, not just the final output. The key idea is that trust is a system property, not a model property. A more capable LLM alone doesn't make enterprise analytics trustworthy. Why does it fall apart when it isn't grounded in governed data? What goes wrong, concretely? Richard Tomlinson: I'd make one distinction here: governance is necessary, but it isn't sufficient on its own. You can have perfectly governed data and still get a confidently wrong answer if the AI doesn't understand what that data means. Without a governed foundation, several things can go wrong at once. The system may select an outdated table instead of the certified one. Two users may get answers calculated from different versions of the same KPI. Sensitive information may be exposed to someone who shouldn't see it. And when an answer looks wrong, the data team may have no reliable way to trace which data or calculation produced it. AI amplifies these problems because it dramatically expands who can query data and how many questions can be asked. Previously, ambiguity might have meant an analyst asked for clarification. An AI system can instead make a plausible assumption and confidently continue. That's why the foundation needs governance plus semantics: control over what data can be used and who can access it, and a shared understanding of what that data means. The goal isn't merely to stop hallucinations. It's to make sure AI-generated analytics operate under the same definitions, permissions, and standards of evidence the organization already expects from human-generated analytics. What happens when teams define the same metric differently? A lot of organizations have different business definitions living in different BI tools. What's the risk once AI enters the picture, and how should a data leader think about standardizing those definitions? Richard Tomlinson: AI turns a long-standing BI problem into a much larger one. Organizations have always struggled with competing definitions of things like revenue, active customer, churn, or conversion. With dashboards, those inconsistencies are at least relatively bounded. Finance has its dashboard, Sales has another, and people eventually learn which one to use. AI removes those boundaries. A user simply asks, "What was revenue last quarter?" Now the AI has to decide which definition, dataset, and calculation represents the organization's intended meaning. If those definitions are fragmented across individual BI tools, semantic models, and dashboards, the AI is effectively being asked to resolve an organizational disagreement on its own. That's dangerous, because the answer may still look completely reasonable. The answer isn't necessarily one gigantic semantic model that tries to describe the entire company. It's moving critical business meaning into shared, governed semantic definitions that live close to the underlying data and are reusable across analytical and AI experiences. That becomes even more important as companies deploy more AI agents. You don't want every dashboard, copilot, and agent independently learning what "net revenue" means. You want them all drawing from the same shared business context. This is one of the biggest architectural changes AI creates for analytics: semantics can no longer live exclusively inside the presentation layer. Read now Why do users need to see how an AI reached its answer? Beyond accuracy, why does it matter that a business user can see how the AI arrived at an answer, not just the answer itself? Richard Tomlinson: Because analytics isn't only about getting an answer, it's about having enough evidence to make a decision from that answer. If an AI tells a sales leader enterprise churn increased 14%, the natural next questions are: compared with what period, which customers are included, how is churn defined, what data supports that conclusion, and can I drill into it? A trustworthy system should make those questions answerable. I'd be precise about what "show how the AI got there" means. It doesn't mean exposing the model's private reasoning or chain of thought. What users need is provenance and evidence: the data sources used, the calculations or queries executed, relevant [truncated for AI cost control]