How AI-Powered Semantics Ensure Trustworthy, Intelligent Agentic Analytics

Generative AI alludes to more than just a technology—it’s a rapidly developing ecosystem of models, architectures, research, techniques, and solutions. And this ecosystem as a whole is demonstrating profound methods to make us more efficient, automate tasks, learn faster, and so much more. 

As we apply new models and agentic approaches to data and analytics, the bar for trust is high. And rightly so—when we query data for analytical purposes, we’re using those outputs to drive critical business decisions. For a concrete example, consider the impact of an agent answering questions about API documentation on a website versus an agent giving a CFO metrics for an SEC filing. If the agent is misleading with API integration details, a developer may be frustrated. But if the agent is misleading with data that makes it into an SEC filing, fines may be issued and people could be fired. While both are bad, one feels more critical than the other, doesn’t it?

We all want trusted, intelligent, relevant, personalized answers from our agentic experiences. And across the industry, we’re all faced with the same fundamental challenge: agents take everything literally. To mitigate this (and more), we use semantics—and not just your grandmother’s semantics. Of course this means measures, dimensions, custom fields, and relationships, but it’s so much more. It's the context, meaning, preferences, lingo, and intentions that drive a business. 

Many of these things will commonly be found in a semantic layer, but they are more often encoded in our minds as tribal knowledge. And because of this, the next generation of AI-powered data and analytics solutions will be differentiated by the richness and capability of the semantics it uses, and more importantly, the tools that it provides to lower the bar to exchanging tribal knowledge into a well-curated collection of semantic models. 

These are the semantic models that agentic AI needs to secure our trust in their outputs: fully customizable to describe the structure and nuance of any organization, with personalized perspectives, and humans in the loop to provide the right guardrails—all while bringing visibility into system quality.

Now let’s dig in a little deeper and talk about different capabilities of Tableau Semantics—the AI semantic layer deeply integrated in Data Cloud—that are doing just this.

Concepts and capabilities of Tableau Semantics

Business semantics

Business semantics is what most people think about as we talk about a semantic layer. These are the core concepts that comprise a semantic model, that all other elements orbit around. This is where logical tables are defined, fields are renamed, new logical (calculated) fields are added, and logical relationships are created. In addition, this is where you can further explain the model through additional metadata like field descriptions, model context, and relationship details. 

Business context 

Business semantics provides the definitions and structures of data elements, while business context interprets the significance and implications of that data for particular individuals or groups. The same data point can hold entirely different values and meanings depending on who is analyzing it and what their goals are. For instance, a jump in the number of cars sold is excellent news for the person in charge of sales at a car dealership, indicating they're meeting or exceeding targets. However, if you consider a hospital setting, a sudden surge in patient admissions could be a cause for concern, signaling a potential public health issue or a problem with efficiency. This disparity highlights the crucial role of business context in understanding and acting upon data. It's not just about the "what" of the information, but also the "why" and "for whom" that truly unlocks its value in decision-making and strategic planning.

To address this, Tableau Next allows its users to augment semantic models with additional metadata, that we call business preferences. Input by authorized stewards, business preferences enhance AI agent accuracy by adding specific rules. Incorporating business knowledge and domain logic is crucial for trustworthy results, as AI agents otherwise lack contextual business understanding. Leading organizations stress the need for agents to understand unique business context, terminology, and data usage. Business preferences drive semantic learning, enabling knowledge input and agent coaching for continued improvements. 

Governance

Building a single source of truth with semantics is extremely important. Doing so requires awareness of the existence of semantics, but also the ability to govern it through many dimensions. From the data perspective, this can be at the object level, field level, or row level. From the semantic perspective, this is about controlling what users and groups can read, update, and delete. Perhaps more importantly, agents need to respect all of these rules as well, acting as an invocated user at all times. The ability to enable self-service analytics on these models is also important, which as we know can sometimes require modifications to the model. The system needs to be able to allow for ad hoc creativity, with governed workflows for introducing those changes from a place of non-linear, self-service analytics back to the governed, single source of truth.

Composable data models

Semantic models are important for many different experiences. Furthermore, they tend to have layers of foundational importance and a hierarchy of importance to different teams and divisions within the organization. Perhaps your organization wants some global semantics, with models extended and redefined for a given team. For example, ARR (annual recurring revenue) might be important for finance, while ACV (annual contract value)  is important for sales. Yet both might both be based on the same derivative calculation, all of which needs to remain consistent.

Metrics

A metrics store provides centralized views into the semantics that power consistent, intelligent, personalized, and contextual insights delivered to users in the flow of work. Metrics help everyone in your organization integrate data into their daily jobs to make better, faster decisions. Without having to learn a new tool or build comprehensive visualizations, metrics help you go beyond the “how” and “what” and show you the “why” behind your data.

Tableau Published Data Sources

Many organizations (including Tableau), have spent years creating highly curated semantics, meticulously built into data and analytics platforms—in most cases, as Tableau Published Data Sources. Migrating or moving these can be tedious and error prone. But now, bringing Published Data Sources into Tableau Semantics for use in Tableau Next is just a click away. Note: This capability is in development for release later this year.

Semantic learning

A customer’s knowledge of their business and domain-specific logic plays a crucial role for achieving more trustworthy and accurate data results in general—and this particularly true with AI. However, much of that knowledge remains outside of the system, in the minds of the people that make up the business. To bridge this gap, a dual strategy is needed for capturing semantic and business insights. This involves expert both curation—to allow professionals to create new semantic definitions, data object metadata, business terms, and specific instructions—as well as coaching agents so they learn customer data through the proposed definitions and preferences from users during agent interactions.

Concierge testing center

Concierge, a pre-built agentic analytics skill in Tableau Next, enables conversational analytics so business users can ask questions of their data in plain language. To ensure people receive accurate and reliable answers from agents, there needs to be provisions provided by subject matter experts with self-service agent testing and calibration tools. This will allow these experts to fine tune the agents to align with their organization's specific analytics needs, driven from their data, semantic models, terminology, and other custom definitions—ultimately creating a more accurate and personalized Q&A experience for their users. Note: This capability is in development for release later this year.

Verified answers

To enhance out-of-the-box agent accuracy and address the cold-start challenge, a common industry practice involves equipping AI agents with analyst-verified question responses. This approach enables agents to provide precise answers to semantically similar queries and clearly indicate when a response has been verified. Expert-verified questions serve as a coaching tool, and a baseline for regression testing, boosting the agent's response accuracy and consistency, which in turn builds user trust. Verified questions, specifically supporting inline coaching of agents, through multi-turn conversations, empower data stewards and analysts to verify both questions and results. Additionally, it allows the agent to draw upon verified questions stored within the semantic layer repository when responding to semantically similar inquiries and grants data professionals the ability to verify questions through the testing center.

AI readiness for semantic data models

AI readiness provides visibility into the overall completeness of your semantic models, so you know where to take time to curate, and what the impact of those curations are. Furthermore, it helps accelerate model preparation with agent-driven UX, quantifies semantic model readiness for AI workflows, and identifies and resolves structural and semantic inefficiencies. Together, this provides an automated evaluation functionality that strengths your semantic models on key AI quality dimensions. Note: This capability is in development for release later this year.

Bringing it all together

Each of these elements of a semantic layer are important, of course. But it’s how they’re managed and curated that makes them the most valuable companion for agentic AI. Truly, your semantics are the brains behind your AI agents. Taking the time to curate and test for conflicts, correctness, relevancy is paramount. With tooling provided, you get transparency into the efficacy of your semantics as your agents respect them. Furthermore, having these be composable to you, as all elements of Tableau Next platform are, you can choose which tools are the most important for you to use and how they each influence one another.

Ready to get started? Learn more about Tableau Semantics and see it in action.