AlphaSense cuts a multi-week data rebuild to hours with dbt and Euno

Why context matters for AI agents on enterprise data

The session was led by Sarah Levy, co-founder and CEO of Euno, and Brad Levy, director of analytics at AlphaSense. Their starting point was that AI demos look clean on stage, while production data is federated, fragmented and full of conflicting definitions. According to the presenters, working in that environment requires context that is automated, governed and scoped, rather than assembled by an agent on the fly.

Euno continuously crawls metadata across a company's stack, including data platforms, ETL and ELT tools, BI, security and observability systems. It rebuilds that metadata into a live context graph that captures how data assets relate to each other, not only that they exist. AlphaSense uses dbt as its governance layer on top of that context.

Use case one: a legacy dashboard model rebuilt in one session

The first use case involved a single table with 287 fields and 68 dbt models upstream of it. It powered 12 dashboards for roughly 300 users, and the real business logic sat in the BI layer rather than in dbt.

Using Euno for context and Cursor as the IDE, Brad Levy rebuilt it as a governed star schema of 13 models. The work took about 15 prompts in a single session, and the new models were reconciled against the legacy ones for parity before shipping. According to the case study, work that would normally take three to five weeks became one focused build-and-validation cycle of a few hours. Roughly half of the resulting diff was tests and documentation rather than net-new logic.

Billing redesign and shared context for many agents

The second use case was a redesign of a billing architecture that had outgrown a single fact table once a second product line arrived. No code was generated. Euno traced the current state down to the column level, and the proposed redesign, including two Mermaid diagrams, was delivered as a Notion document for AlphaSense's data architect to review and comment on directly. The full proposal took about two focused hours with an agent, compared with an estimated day and a half by hand.

The third use case is still in progress. It aims to scale conversational analytics so that many agents across the business draw on the same governed context, instead of each one reconciling metadata independently and disagreeing with the others.

Accuracy, speed and what the Q&A added

AlphaSense summarized the payoff in three points. The models were more accurate because they came out governed and tested, not just generated. The work was faster, measured in hours rather than weeks. And, perhaps unexpectedly, it was more enjoyable, because time shifted from re-tracing what already existed toward judgment and design.

In the Q&A, the speakers discussed how Euno's live labels, which are encoded rules such as "certified if it only queries gold-layer dbt models", turn conflicting definitions into an explainable choice rather than a guess. One example was two different definitions of churn from different teams. They also described a benchmark-question mechanism that continuously checks agent accuracy over time instead of relying on a single point-in-time evaluation.

The figures come from AlphaSense's own account of its work, as reported by dbt Labs. For teams running AI agents on messy production data, the case study points to governed, automated context and built-in tests and documentation as the basis for faster and more reliable rebuilds. The talks from dbt Summit 2026 are available on demand.