Reading the AlphaSense case study: what it shows and what it does not

Most vendor case studies in this category say a customer saved time. Euno's AlphaSense write-up, presented at dbt Summit by AlphaSense's Director of Analytics, publishes the size of the task and the number of prompts. That makes it worth reading closely.

AlphaSense's agents connect to Euno's MCP server for context and query BigQuery and Snowflake. The stack includes dbt Cloud, Tableau and Cursor. Euno says it integrates with each system's APIs and query logs, crawls metadata continuously and computes its context graph ahead of time, rather than gathering context through MCP at query time.

Source: Euno blog: AlphaSense x Euno at dbt Summit

Euno reports two tasks. First, a legacy reporting table with 287 fields, feeding 12 dashboards, 41 views and 303 users and built from 68 dbt models, was rebuilt as a 13-model Kimball star schema in about 15 prompts in one session. The manual estimate was 3 to 5 weeks. Second, a billing model redesign proposal took about 10 prompts and two hours, against an estimate of a day and a half, and a column-level trace from a workbook to the warehouse took seconds.

Source: Euno blog: AlphaSense x Euno at dbt Summit

It shows the kind of task where context matters most: refactoring. Rebuilding a wide legacy table safely requires knowing every downstream dashboard, view and user. An agent with column-level lineage and usage can plan that work; an agent with only schema cannot. The case also shows the value of business logic stored in BI: Euno says its context extends to definitions deep in BI workbooks, outside curated semantics.

Source: Euno blog: AlphaSense x Euno at dbt Summit

It does not show a controlled comparison. The baselines are estimates, not timed manual runs. It is one customer with an experienced analytics team using Cursor, which is a favorable setting. It does not report how many of the prompts produced output that needed correction, or how the result was validated. None of this makes the figures wrong; it means they are an example, not a benchmark.

Pick a comparable task from your backlog: a wide legacy table that many dashboards depend on. Write down your team's estimate before you start. Run the task twice with the same agent, once with schema access only and once with the context platform connected, and record prompts, elapsed time, corrections and how you validated the result. Keep the raw transcripts.

Do the same with every shortlisted vendor, so the comparison is between products on your work rather than between case studies.

Three additions would turn a good example into evidence. A timed manual baseline for a similar task, rather than an estimate. A record of how the agent's output was checked, for example tests that passed or a reviewer's sign-off. And a second customer with a different stack, to show the result does not depend on one team's skills. We would welcome case studies from any vendor on this site that include those three elements, and we will link them from the relevant review when they are public. Until then, read every case study in this category, including the one above, as a description of what is possible rather than a forecast of what you will get.

It does not change any score by itself. Our lineage and automation scores for Euno rest on what its product pages document. The case study appears on Euno's review as a vendor claim, labelled as such, because readers asked for concrete examples rather than adjectives.

If you run the test described above and want to share the result, the contact page accepts corrections and evidence. We do not publish private customer results without permission.