How C3 AI Data Intelligence Agents Beat State-of-the-Art Text-to-SQL Systemss

By Sina Pakazad, Vice President, Data Science; Anoushka Vyas, Senior Data Scientist; Aarushi Dhanuka, Data Scientist; and Henrik Ohlsson, Vice President and Chief Data Scientist, Data Science, C3 AI

A typical enterprise data estate is a pile of formats: spreadsheets exported from legacy systems, CSVs from operational databases, JSON feeds from APIs, and Excel files passed between teams. Almost none of it arrives structured for analysis. Before any of it can support machine learning or a business decision, someone has to discover it, understand it, integrate it, structure it, and make it accessible. That sequence is one of the most time-consuming steps in the analytics lifecycle, and the hard part is semantic: building a shared understanding of what each field means, which records can be trusted, and how business concepts map to tables. That understanding is harder to produce than the data movement itself.

In practice, the work takes repeated collaboration among data owners, domain experts, data engineers, and data scientists. Data owners know how the data is produced, what the fields mean, and which records are reliable. Data engineers handle ingestion, transformation, validation, orchestration, and scalable access. Data scientists and analysts depend on the resulting structure to ask reliable questions. Once these groups align, they codify the relationships across sources into a semantic layer that turns raw enterprise data into something analysts can query with confidence.

The C3 AI Data Intelligence Suite compresses that workflow. It is a set of three agents (Data Interpreter, Schema Creator, and Query Generator) that let the people who know the data best guide the work directly in natural language, while the agents inspect the data, build a validated semantic layer, and run the analysis. In fully autonomous testing, the same system matches or outperforms specialized, state-of-the-art text-to-SQL systems on standard benchmarks, using a general-purpose model rather than a custom-trained one.

Where Traditional Data Integration Slows Down

Figure 1: Effective data access requires teams to iterate on interpretation, then design and validate a schema over the data.

Enterprise data integration follows a familiar loop (Figure 1). Discovery comes first: what data exists and where it came from. Files and tables are then profiled to understand structure and quality and to agree on business meaning. Schemas are designed next, capturing entities, relationships, constraints, and indexes. The data is loaded and validated, and type mismatches, missing values, and referential-integrity issues are resolved. Only then can analysts answer business questions over the integrated data. Every step is necessary, and each requires several rounds of iteration and handoff between data owners and engineers.

Data owners understand what the data means. Engineers and analysts know how to structure and validate it and how to orchestrate the queries the business needs. When that knowledge moves between people through meetings, tickets, and document revisions, context gets lost and latency accumulates. A single missed relationship or misunderstood field can force schema changes, pipeline updates, and query rework. The process is rigorous, but it is slow, coordination-heavy, and lossy. The opportunity is to give data owners and domain experts a faster, more direct, and more reliable way to drive it.

Three Specialized Agents Inside the C3 AI Data Intelligence Suite

The C3 AI Data Intelligence Suite preserves the rigor of enterprise data integration while removing the slow handoffs that keep it from scaling. It puts data owners and domain experts in the driver’s seat. Rather than translating every clarification, schema decision, or analysis request into an engineering cycle, they guide agents that inspect the data, create executable artifacts, test those artifacts, and expose the results for review. The suite also integrates with the data capabilities of the C3 Agentic AI Platform for ingestion, access, and analysis at enterprise scale.

The suite consists of three specialized agents, each responsible for a distinct stage of the workflow. The agents interact with the user and with one another as they inspect data, create artifacts, execute code, validate outputs, and refine results.

Data Interpreter Agent

The Data Interpreter Agent helps users understand what data they have and what it means. It inspects raw files and tables to identify:

• Available files, tables, and fields
• Data types, distributions, missing values, and anomalies
• Duplicate records and data-quality issues
• Candidate entities represented in the data
• Potential keys, join paths, and relationships across datasets
• Naming inconsistencies and assumptions that may need domain confirmation

It presents these findings in a structured, reviewable form, so data owners and domain experts can validate, correct, and enrich the system’s understanding before any downstream modeling begins.

Schema Creator Agent

Once the data has been interpreted, the Schema Creator Agent builds a semantic representation of it, capturing:

• Core entities and attributes
• Relationships across tables and files
• Primary keys, foreign keys, and referential constraints
• Business-relevant concepts and definitions
• Normalized structures that support reliable analysis
• Indexes and schema choices for efficient querying

The agent creates the schema, loads the underlying data into it, and runs test queries to confirm the schema is functional, executable, and aligned with the actual data. Schema design becomes a tested artifact rather than a static modeling exercise.

Query Generator Agent

After the semantic layer exists, the Query Generator Agent lets users ask analytical questions in natural language and translates each request into an executable analysis plan that can:

• Identify the relevant entities, tables, and relationships
• Decompose complex questions into multiple analytical steps
• Generate and execute one or more queries
• Inspect intermediate results
• Refine the analysis when results are incomplete or inconsistent
• Synthesize a final answer with supporting evidence from the data

The Query Generator Agent goes beyond basic text-to-SQL. It plans, executes, reasons over intermediate outputs, and communicates results in terms users can review and act on.

Together, the three agents mirror the traditional data-integration process while compressing its feedback loops. Discovery, schema design, validation, and analysis stay rigorous, and they become interactive, executable, and directly guided by the people who know the data best.

Built on Execution: Why Data Intelligence Agents Validate Against Real Data

The Data Intelligence (DI) agents are built around execution. They are directly built on autonomous coding agents (ACAs) running in remote cloud sandboxes, where they create, execute, test, debug, and refine concrete artifacts.

This matters because data-integration work depends on correctness. A schema is only useful if data loads into it; a transformation, only if it runs; a query, only if it executes and returns results grounded in the underlying data. An agent that relies on a language model alone can propose what should be done; because the DI agents run inside ACAs, they validate their work directly against the data.

The outputs are verified executable artifacts: schemas, data-loading scripts, transformations, queries, validation checks, and result summaries. Each can be inspected, tested, reused, critiqued, and audited. Together they form a concrete record of how the system interpreted the data, what assumptions it made, what it executed, and what it produced.

This architecture also changes how the agents use memory. Traditional LLM agents rely on text held in the model context or a retrieval layer. ACAs work with persistent artifacts in their sandbox: code, schemas, data samples, validation reports, query logs, and prior outputs. That gives the system a richer, more operational memory and lets it build on previous work instead of reconstructing context from text each time.

The result is a workflow that moves past static, pre-scripted automation. The agents generate an artifact, execute it, observe failures, correct errors, and validate the result. They coordinate through shared artifacts rather than text-based handoffs, which makes the whole process more traceable, reproducible, and reliable. For data owners and domain experts, that removes the long handoff chain: they guide the work in natural language while the agents inspect the data, build and test artifacts, and expose the outputs for review. The path from raw enterprise data to trusted insight gets shorter without giving up the rigor enterprise data work requires.

Benchmarking DI Agents in Fully Autonomous Mode

This section is written for technically inclined readers. If you are focused on system-level concepts or use cases, feel free to skip ahead.

The Data Intelligence agents are built to collaborate with users, and we also evaluate them in a fully autonomous mode, with no user input. In that setting, the Query Generator Agent independently interprets database schemas, reasons over natural-language questions, generates SQL, executes queries, observes failures, and refines its outputs end to end.

We measure performance on established text-to-SQL and database-grounded reasoning benchmarks, which test whether an agent can turn natural-language questions into correct, executable queries over real databases. Because these benchmarks include ground-truth queries, execution results, or task-level test cases, they measure correctness by execution: whether the generated SQL actually returns the right answer. Table 1 summarizes how the DI agents perform against current state-of-the-art systems.

Table 1: Across text-to-SQL and database-grounded reasoning benchmarks, DI Agents match or outperform leading systems.

State-of-the-art figures reflect the best published leaderboard results for each benchmark at the time of writing.

The DI agents match or exceed state-of-the-art results across a range of SQL and database-reasoning benchmarks. That is notable because many leading systems rely on specialized architectures, benchmark-specific techniques, custom-trained models, or multi-model ensembles tuned for text-to-SQL. The DI agents reach these results with a general-purpose commercial model, Anthropic’s Claude Sonnet 4.5, paired with an execution-grounded architecture: the agents reason over schemas, generate candidate queries, execute them, inspect errors and intermediate results, and refine their outputs. Careful agent design, execution-based feedback, and schema-aware reasoning can match or beat specialized systems on complex data-analysis tasks.

In Practice: Nursing-Staff Analysis, Start to Finish

To make the workflow concrete, consider an example from hospital nursing operations. The data covers wards, shift structures, nurses, attendance records, and work history. As in many enterprise datasets, it is spread across multiple files and tables, with relationships that are not explicit at the outset.

The user provides sample data files to the Data Intelligence agents and starts the workflow (Figure 2). The Data Interpreter Agent inspects the files, identifies key fields and entities, detects data-quality issues, and proposes candidate relationships: how nurses relate to shifts, how shifts relate to wards, and how attendance records connect to broader staffing patterns.

Figure 2: Users upload data and start the workflow in the C3 Intelligence interface, where the Data Interpreter Agent begins inspecting the files.

The interpretation results come back in a structured, reviewable form. This step matters because the system should not silently guess what enterprise data means. Data owners and subject-matter experts review the agent’s findings, correct assumptions, clarify field meanings, and validate the inferred relationships, creating a shared understanding of the data before any modeling or analysis.

Once the user is satisfied with the interpretation, the Schema Creator Agent constructs a semantic layer over the data (Figure 3). In this example it lays out a star schema: a central fact table of roughly 122,000 shift assignments linked to dimension tables for nurses, wards, shift patterns, and time windows (Figure 4). The agent creates the schema, loads the data, and runs validation checks and test queries, reporting full referential integrity and no orphaned records.

Figure 3: The Schema Creator Agent summarizes the interpreted data as a star schema and reports data-quality checks, including full referential integrity and no orphaned records.

Figure 4: The validated schema, with the central shift-assignments fact table linked to dimension tables for nurses, wards, shift patterns, and time windows.

With the semantic layer in place, the Query Generator Agent lets the user ask analytical questions in natural language. A user might ask which wards are most affected by staffing gaps, whether certain shift patterns are associated with absenteeism, or how nurse experience relates to attendance. The agent turns each request into an executable plan, runs the queries, inspects intermediate results, and returns answers grounded in the validated schema. It also produces visual outputs, including dashboards, so results are easier to communicate, while still exposing everything for review.

The demo above uses sample files the user uploaded. The same workflow runs over enterprise source systems through the C3 Agentic AI Platform, which connects to source systems, ingests data, orchestrates workflows, manages access, and scales execution across enterprise datasets. The agents validate logic and analysis in their execution environment, then apply those validated workflows to broader enterprise data through the platform.

The C3 AI Data Intelligence Suite at a Glance

• Three agents, one workflow. The Data Interpreter, Schema Creator, and Query Generator agents cover discovery and profiling, semantic-layer creation, and natural-language analysis.
• Domain experts drive. Data owners and subject-matter experts guide the work in natural language and review every result, instead of filing requirements and waiting on engineering cycles.
• Execution-grounded. The agents run as autonomous coding agents in cloud sandboxes, validating schemas, transformations, and queries against the real data.
• Auditable artifacts. Every step produces inspectable, reusable outputs: schemas, load scripts, transformations, queries, validation checks, and result summaries.
• Benchmark-tested. In autonomous mode, the suite matches or outperforms state-of-the-art text-to-SQL and database-reasoning systems using a general-purpose model.
• Enterprise scale. Through the C3 Agentic AI Platform, the same validated workflows run over production source systems with managed ingestion, access, and orchestration.

What Changes When Domain Experts Drive Data Work

Together, the Data Interpreter, Schema Creator, and Query Generator agents form a collaborative toolbox for understanding, integrating, and analyzing enterprise data at scale. The most important change is who does the work. Data owners and subject-matter experts participate directly, validating assumptions, refining interpretations, and asking increasingly sophisticated questions, while the agents inspect data, build schemas, execute code, and produce reviewable artifacts.

That shortens the path from fragmented raw data to trusted insight. It cuts the handoffs that slow teams down, keeps the rigor enterprise data work requires, and makes the process more transparent, reproducible, and scalable. Domain experts stop waiting for every clarification or analysis request to move through a long technical queue and start working alongside agents that do the work with them. For enterprises, that means faster time to value from existing data, more reliable analytical foundations, and a more direct route from operational knowledge to decision-ready intelligence.

If your organization is dealing with fragmented data, slow integration cycles, or growing pressure to make data accessible to the people who understand it best, we would like to show you what the C3 AI Data Intelligence agents can do with your own data.

Want to put these agents to work on your own data? Contact us and we’ll put the C3 AI Data Intelligence Suite to work on a sample from your environment.

About the Authors

Sina Khoshfetrat Pakazad is Vice President of Data Science at C3 AI, where he leads research and development in generative AI, machine learning, and optimization. He holds a Ph.D. in Automatic Control from Linköping University and an M.Sc. in Systems, Control, and Mechatronics from Chalmers University of Technology. He has worked at Ericsson, Waymo, and C3 AI, contributing to AI-driven solutions across healthcare, finance, automotive, robotics, aerospace, telecommunications, supply chain, and process industries. His recent research, published at venues including ICLR, ICML, and EMNLP, focuses on multimodal data generation, instruction-following and decoding for large language models, and distributed optimization. He has also co-invented patents on enterprise AI architectures and predictive modeling for manufacturing.

Anoushka Vyas is a Senior Data Scientist at C3 AI. She holds a master’s degree in computer science from Virginia Tech and a bachelor’s degree in electronics and communication engineering from IIIT Hyderabad. Her background spans graph neural networks, geometric deep learning, and optimization-integrated machine learning, with publications on hyperbolic graph neural network architectures at The Web Conference (WWW) and on graph-based forecasting for precision agriculture in the IJCAI-ECAI AI for Good track. Her current work focuses on scalable AI and optimization for complex operational decision-making in enterprise environments.

Aarushi Dhanuka is a Data Scientist at C3 AI, where she focuses on agentic development and intelligent workflow automation, applying data science and engineering techniques to build AI-driven tools for enterprise operations. She is pursuing an M.S. in Computational Science and Engineering at the Georgia Institute of Technology, specializing in applied machine learning and computational data systems.

Henrik Ohlsson is Vice President and Chief Data Scientist at C3 AI. Before joining C3 AI, he held academic positions at the University of California, Berkeley; the University of Cambridge; and Linköping University. With more than 70 published papers and 30 issued patents, he is a recognized leader in artificial intelligence and its industrial applications. He is also a member of the World Economic Forum, where he contributes to discussions on the global impact of AI and emerging technologies.