Aniruth Narayanan, product manager at Snowflake, said Snowflake CoCo now works inside Microsoft Excel and Google Sheets as a governed coding agent for data teams.
The demo shows CoCo building a leveraged buyout model at 4x speed, with live Snowflake data instead of CSV exports. The affected surfaces are Excel, with support for up to 1 million rows, and Google Sheets, with a 20 million-cell limit. There is no remote exploit here. Access requires a signed-in Snowflake user and that user's role, masking and row-access policies apply to every query.
Narayanan said CoCo removes ODBC and third-party connectors and runs queries in Snowflake's warehouses, then streams results directly to the sheet. He said the add-in saves query text in workbook settings and offers manual refresh via a button or /refresh skill, plus automatic refresh on open for Excel. If a SELECT * changes shape, CoCo keeps formulas and assets intact. That lineage matters because Snowflake said all queries stay visible in the workbook to reconcile which number to trust.
Matt Wyman, vice president of information technology at The Chefs' Warehouse, said CoCo understood joins across his company's complex tables and felt intuitive enough to use without training.
Kate Savina, head of data technology at investment firm GTCR, said many AI tools work in a spreadsheet but CoCo starts with Snowflake data directly. She said that removed exporting and switching tools and cut time spent rebuilding reports. The two customers did not disagree on utility, but they pointed to different payoffs. Wyman emphasized correctness on messy relational data, Savina emphasized speed to answer.
The move places Snowflake against the spreadsheet owners themselves. Microsoft pitches Copilot in Excel and Google pitches Gemini in Sheets as native AI that writes formulas and analyzes tables, but both sit outside Snowflake governance. Snowflake is betting analysts will prefer an agent that honors Semantic Views, Business Ontology and Cortex Sense and offloads heavy aggregation to its warehouses. That pitch follows Snowflake wants your model training off the cluster on Nov 3, which pushed warehouse-based training over cluster-bound jobs.
Snowflake said internal tests on 25 spreadsheet tasks using Opus 5.5, Sonnet 5.5 and Haiku 5.5 in Excel for Web, across three trials between Oct. 26 and Oct. 28, 2026, showed similar accuracy for CoCo for Excel versus an AI harness using the Snowflake MCP, but lower median seconds per task and fewer output tokens for CoCo. The company said CoCo streams results without using the model's context window, which kept performance near-constant as data size scaled, while the MCP harness hit a 900-second cap on failed or timed-out tasks.
Narayanan said CoCo can also call Cortex AI functions such as CORTEX.CLASSIFY and CORTEX.SUMMARIZE directly in Excel cells, and build live charts and pivot tables that exceed local sheet limits by computing in Snowflake. He did not say what CoCo for Excel and Sheets will cost after public preview, how Horizon Catalog lineage will appear, or what happens when two users with different roles open and refresh the same workbook.
