At Education Analytics, we strive to build standard data infrastructure that democratizes data access, so educators can gain insights and improve student outcomes. In the age of AI, we are expanding that goal to cover not just data infrastructure, but “AI infrastructure” as well. That means building a system for AI-friendly context and deploying tools that utilize LLMs to answer prompts, so educators or agency staff who otherwise don’t have the time or technical expertise can access the data insights they need.
In pursuit of that goal, we have built Stadium AI to test the hypothesis that a natural language interface on top of a well-structured education data warehouse could let more people, not just analysts, ask questions about their data and get answers they can trust. We started with testing a beta version of Stadium AI with real analyst users at the South Carolina Department of Education (SCDE) to see how they would use it, and if that could inform how a broader audience may use the tool. Here’s what we learned.
One pattern stood out: users were not only retrieving data, they were also asking how to interpret it. For example, they wanted to know why a number meant what it did, how a business rule worked, or which table had the data needed for a specific use case. These are not the questions standard analytics tools are designed to answer, as one analyst at SCDE explained:
Transparency was the other element users kept coming back to. The ability to trace a calculation to the source system is something that typically requires a data engineer. When the tool can do it conversationally, it changes what's possible for an analyst or staff member to do without being as dependent on a data engineer. It could also provide more staff the ability to discover the source of an error or understand the rules behind a metric. In turn, this transparency can increase staff trust in the reliability of the data. The structured context, or the semantic model, that tells the AI what the data means, draws on validated knowledge rather than improvising general training.
The concern we hear most often about AI tools applied to student data is: can you trust what it tells you? Our approach is designed specifically to address that. The validity of responses to queries in the attendance domain, which was our first use case for context development, held up well in the pilot.
But the feedback that stuck with us the most and validated that we were on the right track was about who else might eventually use a tool like Stadium AI.
This is the access problem Stadium is built to solve at every layer, and what the pilot confirmed we’re on track to address. The warehouse lowers the barrier to working with education data by providing a standardized (i.e., predictable) structure and decreasing silos by pulling multiple sources into one warehouse using interoperability technology. Stadium AI lowers the barrier further — to the point where the people who have questions and the domain expertise can get answers without needing to learn SQL or understand warehouse architecture to do it. Stadium AI is also more efficient and reliable because it is built on that standardized base and context layer.
Digging Deeper into Stadium AI
Stadium AI is a natural language interface built on top of Stadium, EA’s managed education data warehouse. A user types a question; the tool works to answer it by drawing on a curated context layer — documentation, business rules, and state-specific resources, such as the South Carolina Accountability Manual, to help the AI agents make sense of the raw data.
There are two interaction modes we’ve developed and tested. The first is documentation and schema lookup: questions like “Where does demographic data live in the warehouse?” or “What business rules sit behind the graduation rate calculation?” The second is analytical querying: questions like “What school had the highest average attendance rate this year?” and “Which students who are currently chronically absent have enough time remaining in the school year to recover and cross the 90% threshold?” These modular tools build and run an SQL query, interpret the results, and return an answer that is accompanied by context and documentation.
What separates Stadium AI from a generic AI tool applied to any database is that context layer. Most importantly, the data and context are governed within the state- or LEA-owned infrastructure. EA currently utilizes Snowflake as well as Snowflake’s LLM tooling, so data never leave the existing cloud data platform or are used for training models, and users’ role-based access control (RBAC) is respected. Furthermore, the state can configure agent behavior, e.g., to only recommend interventions from an approved list. This offers more control than a generic external AI tool separate from the state-governed system. Without it, an LLM is more likely to hallucinate, misinterpret schema, or produce answers that are technically plausible but wrong.
The Data Foundation is the Point
What makes Stadium AI possible is that the data it runs on has already been standardized and well-documented. Stadium is built on the analytical framework Enable Data Union (EDU). The data Stadium AI reasons over have been extracted from Ed-Fi operational data stores, staged, transformed through validated dbt models, and organized into an analytics-ready dimensional warehouse — standardized, structured, documented, and governed.
This matters enormously for AI. An LLM applied to messy, ungoverned data produces messy, ungoverned answers. The education data space has seen too many tools that generated impressive-looking outputs built on inputs that nobody fully understood or that tried to be one-size-fits-all.
The context layer we’ve built for Stadium AI extends that same philosophy: curate the knowledge once, build in site-specific configuration, make it available consistently, and guide the AI to draw on it rather than improvise. Context built once also scales: it works for any other partner using EDU.
What We’re Still Working Through
Our initial findings from this pilot are very positive. Curated context on data and documentation has enabled AI agents to provide helpful and accurate answers that can improve the efficiency of analysts and accessibility of data-informed insights. The pilot also provided a variety of new research topics for us to continue to explore.
The most fundamental question we are still working through before rolling out this feature widely: how do we verify that our answers are reliable, at scale? It is impossible to test and evaluate the LLM’s answer to every question that a user might ask. Even with strong context and thorough testing before release, an LLM can still generate a wrong response. How do we know that the LLM's answers to questions we haven't yet tested are reliable enough, and how do we communicate that to users?
We’ve learned that different user types interact with the tool differently and need different configurations. A state analyst running queries across many districts has different needs than a district administrator checking a single school’s data, or a researcher trying to understand how the warehouse data are structured in order to build statistical code. We’re working toward a model where the tool can be configured for different roles — surfacing the right capabilities and the right scope of access for each, without requiring every user to understand the full system.
Users want to know what to do next. Which interventions should I try to improve attendance? What resources are available to my students? This requires a different type of context that is more specific to the partner. How do we build a process to responsibly incorporate that data?
On the cost side, running AI tools on complex, multi-step analytical questions involves compute and token costs we need to understand before we can scale responsibly. The pilot has given us data on this, and optimizing for efficiency — getting more accurate answers at lower cost per prompt — is part of the work ahead.
Our Next Focus
Our priority is getting this right before getting it everywhere. The education sector doesn’t need more AI tools that generate enthusiasm before they generate results. It needs tools that work when and how they say they will.
The attendance domain is where the context is strongest, and the pilot results reflect that investment. Expanding coverage to other domains — assessment, demographics, accountability, program data — requires the same work of building, testing, and validating context before we can be confident in the answers, and that work is ongoing. That work can move faster now that we have a template.
We’re also building toward a future where the context layer itself — the structured, governed semantic knowledge that makes AI tools trustworthy in education — can be community-driven and shared across agencies, the same way EDU code is shared.
In this future, we envision district data staff, school principals, and educators directly interacting with and making decisions on their data in a way that works for them. School principals and educators will be able to quickly ask AI agents questions to understand current student performance, trusted interventions to implement, and how to evaluate effectiveness of those interventions. With the context we add, the answers provided will follow analytic best practice and district policies around interventions and equity. This self-service ability can free up the time of district data staff to turn the most popular and useful queries into an updating analysis that alerts educators when new students are off-track or when there is new information on which interventions are working. Those queries and results can be saved and securely shared across schools or districts to create an easy-to-access library of practices.
What comes next is grounded in EA’s vision of building tools for the public good so that people have the right data, at the right time, to make the right decision.
