The data stack turns agents into a sales motion

Agent-ready is now the organising line

The centre of gravity this week was not a warehouse feature or a dashboard launch. It was the phrase everyone is now trying to own: agent-ready enterprise data. On 17 Sep, Fivetran and dbt Labs introduced new capabilities at dbt Summit 2026 aimed at making enterprise data ready for agents. On 19 Sep and 20 Sep, the same push was carried through multiple press and leadership updates, alongside dbt v2 and dbt State.

That matters because the Fivetran and dbt Labs combination is being presented less as an integration story and more as a new operating layer for AI-era data work. The observed sequence was tight: product announcement on 17 Sep, Summit content and site changes through 19 Sep, further coverage on 20 Sep, and channel planning noted on 21 Sep. This is how a platform narrative gets turned into distribution, not merely documentation.

The wider category followed the same line. Databricks introduced an adaptive instructed-retriever model on 9 Sep, with coverage describing frontier-quality search at twice the speed and lower latency. On 14 Sep, further reporting said Databricks had launched a new model for data retrieval in AI use cases. Snowflake, meanwhile, had multiple reports from 8 Sep to 15 Sep on its expanded AI strategy, dynamic model routing for cost efficiency, and raised guidance tied to AI-driven expectations.

Snowflake and Databricks are building from different ends

Snowflake's fortnight was defined by commercial proof and platform hardening. On 14 Sep, Snowflake reported a 37 percent increase in product revenue to $1.49 billion during its Q2 2027 earnings call and raised its full-year revenue forecast to $6.07 billion. On 15 Sep, coverage also noted a $3.7 billion stock buyback. Customer proof kept arriving: Novo Nordisk was added on 14 Sep with a case study on unifying 900TB of data for 40,000 users, while Mandai Wildlife Group appeared on 19 Sep with a Cortex AI and CoWork case study.

Databricks looked more expansionary. On 17 Sep, multiple articles reported that it would invest more than US$350 million in Singapore to support enterprise AI adoption and double its workforce in the region. On 19 Sep, it also signed a lease at The Franklin in Chicago. Its hiring signals stayed engineering-heavy, with 32 new postings noted on 14 Sep, 43 on 17 Sep, and 14 on 19 Sep, including field engineering, solutions engineering, and AI-forward deployed roles.

The developer surface is becoming a second battlefield. Snowflake added repositories across JDBC, CLI, Kafka connector, Node.js connector, SQLAlchemy, ML Python and Terraform between 12 Sep and 20 Sep. Databricks released CLI v1.15.0 on 9 Sep, CLI v1.17.0 on 16 Sep, Terraform provider v1.131.0 on 14 Sep, and TypeScript SDK v1.3.0 on 19 Sep, when Arrow Flight ingestion was promoted to general availability. In this market, the SDK and connector layer is not plumbing. It is where the AI platform promise either becomes usable or stays in the keynote.

The quiet signals were louder than the launches

The most revealing changes were in the seams. dbt Labs rewrote its homepage in quick succession: on 14 Sep from Summit attendance to watching the AI-era keynote, on 17 Sep to a community keynote, then on 19 Sep to on-demand viewing and a Summit 2027 waitlist. That is a compressed move from event acquisition to product narrative to next-cycle demand capture.

Fivetran's documentation moved with unusual symmetry. On 10 Sep it added 15 troubleshooting pages and removed 15. On 14 Sep it added six and removed six. On 17 Sep it added 19 and removed 19. On 19 Sep it added 11 and removed 11, including new troubleshooting material around AI Connector Agent and applications such as Customer.io and Yotpo, while removing Snowflake destination troubleshooting pages. The signal is not that any single page matters. It is that the connector estate is being actively re-shelved while the company pushes the agent-ready line with dbt.

Sigma's positioning also flickered in public. On 11 Sep its homepage moved from 'Go build it. IT approves.' to 'Vibe-code enterprise business applications.' On 14 Sep it changed back towards 'Go build it. IT approves.' with subtext around analytics, apps and agents on one trusted platform. For a company selling analytics that can become apps and agents, that wording test is the strategy in miniature: make the product feel new without making IT feel bypassed.

Packaging changed too. On 17 Sep, dbt Labs altered tier contents, adding job scheduling and monitoring plus CI checks on GitHub and GitLab to the Developer tier, while also listing job scheduling as removed from Developer. The Starter tier gained five developer seats, Lineage with dbt Catalog, and metrics via dbt Semantic Layer. Whether read as cleanup or repositioning, the movement puts collaboration, lineage and metrics closer to the entry point.

Reliability stayed part of the enterprise story

The category also had a noisy operational fortnight. Snowflake reported two critical incidents on 13 Sep, lasting about 361 minutes and 118 minutes, and another major outage on 15 Sep that lasted about 81 minutes. On 9 Sep, four Snowflake CVEs were published, including improper OCSP response validation and input validation issues. On 17 Sep, CVE-2026-92903 was published with a CVSS score of 8.2, covering improper input validation in Snowflake CLI versions before 3.27.0.

dbt Labs had four incidents reported on 13 Sep, including Discovery API errors, dbt MCP and dbt Wizard error rates, State usage metrics ingestion, and login session termination. On 20 Sep, a major incident titled 'Schema Hydration Panic' lasted about 898 minutes before resolution. Sigma reported a 14 Sep incident involving input table edits for customers who updated their warehouse connection, categorised as no impact but lasting about 1657 minutes, and a 17 Sep major incident involving query timeouts with Databricks in Azure East US that lasted around 201 minutes.

These are not side notes in a market selling agent-ready data. Agents increase the premium on trust, lineage, permissions, connectors and predictable execution. The same fortnight that produced bigger AI claims also produced visible reminders that enterprise buyers will judge the data stack by operational resilience as much as model fluency.