AI Assistant for University Staff to Ask the Data

Post University is a private, for-profit university in Waterbury, Connecticut, and one of the first US institutions to offer full degrees online. It enrolls more than 20,000 students a year, most of them online, and recruits them through admissions advisors.

Its own reporting and analytics team runs the data platform it built with DLabs.AI from its core systems the year before this project. The plan was to start with clean, connected data. With that in place, Post University turned to us again for the next step: building an AI assistant on top of it.

Every answer meant writing a database query, something only a few people at the university could do. Everyone else turned to the analytics team, by chat, email, phone, or a ticket, and waited their turn. For a small team already stretched, every such request was time taken from its own work.

Every department had its own way of counting students, so “how many students do we have?” had as many answers as people you asked. One team counted everyone enrolled, another only students active in the current term, a third split the count by campus or program. Any number pulled from the platform followed one of those definitions, and anyone used to a different one took it for an error.

A question that cut across systems, such as how many graduates had also completed an internship, meant pulling several reports and stitching them by hand. Some were hard enough that people stopped asking.

Together with the university’s analytics team, we built Ava, an AI assistant that lives in the messaging app staff already use. Ava translates the question into a database query and answers as text or a table, or, when asked for charts, as an Excel file. When a question is ambiguous, Ava asks a follow-up first. Analysts can also see the query behind each answer in the same conversation and check how a number was produced.

Post University decided which groups may see which records. We tied each question to the identity of whoever asks, so Ava returns only what that person may see on the platform. We also built three versions of the assistant: one for public data, one for internal data, and one for restricted records. Each group gets the version that matches its access.

The ready-made “chat with your data” tools we tested at the start could not be taught how the university calculates its numbers. That is why we built Ava to work from its own tables, described in plain language, so it picks the right data for each question. Every update to Ava is checked against a set of questions with known answers before users see it.

Ava came with what the university’s analytics team needs to run it day to day. A usage view shows which questions staff ask and which come up most, so the team can see what to add next. A library holds the questions the analysts have verified against their own reports, so new users start from answers known to be right.

Four months after kickoff, Ava was in production with its first users, the university’s analysts. It answered from data in all five core systems: student records, finance, two CRMs, and the learning platform.

Every business question Post University brought to the project is answered from the platform’s data instead of being pieced together from reports. Some of those questions had simply gone unasked, because the piecing together was too much work. Analysts now ask Ava those too.

When two reports disagree on a number, analysts now ask Ava how each one was calculated and find the cause of the difference themselves. Before, a discrepancy like that was simply taken for an error.

One analyst timed a request: pulling the data and building a chart, at least an hour of work by hand, took 15 minutes with Ava.

Close to a year after go-live, Ava is in production and in use by the university’s analysts. Post University called the first phase a success and is planning Ava’s rollout to a first business department, with the university’s new CRM as the next source.

• Microsoft Azure
  Azure OpenAI Service
• Azure Databricks (Unity Catalog, Delta Lake)
• Microsoft Teams
• Microsoft 365 Agents SDK
• Adaptive Cards
• Microsoft Entra ID
• Langfuse
• Python
• SQL