AI proof is replacing dashboard theatre

AI became an execution test

Databricks made the loudest product noise. On 9 Sep it introduced an adaptive instructed-retriever model, with coverage saying it offered frontier-quality search at twice the speed and lower latency. On 14 Sep, the same retrieval push appeared again as a new model for AI use cases, with one report noting the Retriever speed test remained internal. That detail matters: the message is clear, but the public proof is still being staged.

Snowflake answered from the consumption side. Across reports from 5 Sep to 14 Sep, the company tied its Q2 2027 performance to cloud and AI demand, reported product revenue of $1.49bn, up 37%, and raised its full-year revenue forecast to $6.07bn. On 5 Sep, coverage also pointed to CoCo, Snowflake's AI coding agent. Databricks is selling the retrieval layer for agents. Snowflake is selling the revenue evidence that AI workloads are already spending.

dbt Labs and Sigma Computing moved in the same direction, but through content and positioning rather than a single flagship launch. On 5 Sep, dbt videos covered the post-AI data stack and physical AI, and on 11 Sep dbt added an enterprise AI data maturity model. On 11 Sep, Sigma added blog content on AI agents and ChatGPT integration. The category's shared assumption is now visible: analytics is no longer enough unless it can be framed as an AI operating surface.

Commercial proof moved upmarket

Snowflake had the clearest enterprise proof in the fortnight. On 5 Sep, it added Pacific Life to its customer wall and published a case study titled 'Pacific Life Modernizes Data and AI with Snowflake CoCo'. On 14 Sep, it added Novo Nordisk with a case study saying the company unified 900TB of data to power self-service insights for 40,000 users. The public customer story is not about experimentation. It is about regulated, large-scale data estates.

Sigma also refreshed its customer evidence. On 11 Sep, its customer wall added Delta Defense and Dotmatics while removing Ounce and Scribe. On 14 Sep, the same two logos appeared on its case-studies page, with the Dotmatics story highlighting new scientific workflows shipped in days to weeks by embedding Sigma in Luma. Sigma is trying to make embedded analytics and application-building feel operational, not ornamental.

Hiring reinforced the same enterprise tilt. On 14 Sep, Databricks listed 32 new openings, 25 of them in engineering, including field engineering and specialist solutions roles across Seoul, Tokyo and the United States. On the same day, Fivetran listed 10 roles primarily in engineering and internal communications across London, Dublin and remote US options. Snowflake's 13 Sep postings leaned into engineering and machine learning, with roles including GTM analytics and financial services solution engineering. The growth work is happening closest to deployment.

Developer surfaces became the battleground

The most revealing activity was in the developer layer. On 9 Sep, Databricks released CLI version 1.15.0 with a fallback feature for Python interpreter installation and created repositories including databricks-sdk-go and databricks-jdbc. On 14 Sep, it released v1.131.0 of terraform-provider-databricks, adding data sources for IAM external users, service principals and groups. Those are not headline features for business buyers. They are the plumbing that makes platform expansion survivable inside large accounts.

Snowflake moved in parallel. On 4 Sep, it released version 2.2.0 of the Go Snowflake client and added repositories including gosnowflake, snowpark-python and snowflake-rest-api-specs. On 12 Sep, it added five more repositories, including snowflake-jdbc, snowflake-cli, terraform-provider-snowflake and pdo_snowflake. The week after a strong earnings narrative, Snowflake was also widening the surface area for builders.

Fivetran's signal was narrower, but consistent. On 3 Sep, it released v1.9.42 of terraform-provider-fivetran, fixing connection table group mapping. On 4 Sep, it released version 1.22.0 of great_expectations with Oracle and dialect-regex fixes, and added repositories including dbt_ai_reporting, dbt_openai and dbt_shopify. On 10 Sep, it added repositories including connector_sdk_tools and hybrid_deployment. Fivetran is not trying to own the AI story directly this fortnight. It is tightening the ingestion and connector surface underneath it.

The small moves said as much as the launches

Sigma's homepage told a sharper story than a press release. On 3 Sep, its hero changed from 'Vibe-code enterprise business applications' to 'Go build it. IT approves.', with subtext shifting to 'Analytics, apps, and agents on one trusted platform.' On 11 Sep, the headline moved back to 'Vibe-code enterprisebusiness applications', and the subtext became 'Sigma is the layer to build and scaleyour analytics, apps, and agents on trusted data.' In eight days, Sigma tested the tension every analytics vendor now faces: builder excitement versus IT permission.

Snowflake's 4 Sep pricing-page change was easy to miss beside the earnings coverage. Virtual Private Snowflake moved from Contact us to Custom, and the tier copy added that it includes all Business Critical Edition features while being isolated from all other Snowflake accounts. For a company selling deeper into regulated and high-control environments, that is a meaningful packaging signal.

Fivetran's documentation churn was just as instructive. On 3 Sep it added three troubleshooting pages and removed three. On 4 Sep it added seven and removed seven. On 5 Sep it added two and removed two. On 7 Sep it added two and removed two. On 10 Sep it added 15 and removed 15. This was not a launch. It was a support surface being re-cut, connector by connector, while the market's attention sat on AI models and earnings.

Reliability stayed visible

The AI narrative did not erase operational reality. Snowflake had a critical incident on 5 Sep lasting around 156 minutes, then two resolved critical incidents on 13 Sep lasting around 361 minutes and 118 minutes. On 9 Sep, four Snowflake CVEs were published, including CVE-2026-85525 with a CVSS score of 7.4 for improper OCSP response validation and CVE-2026-85528 with a CVSS score of 5.3 for improper input validation in the JDBC Driver.

dbt Labs also had a busy status fortnight. On 5 Sep, incidents included a critical outage in which the dbt Platform was not responding for about 91 minutes, plus minor onboarding and metadata-ingestion issues. On 13 Sep, dbt reported four more incidents, including intermittent Discovery API errors, increased error rates for dbt MCP and dbt Wizard, a major state-usage metrics ingestion issue, and user login session termination.

Sigma reported three incidents on 4 Sep, including a critical issue that prevented GCP-hosted organisations from loading for about 80 minutes. On 14 Sep, Sigma also resolved an incident in which input table edits were failing for customers who updated their warehouse connection, categorised as no impact and lasting about 1,657 minutes. As vendors stretch from dashboards into agents and applications, reliability, identity, drivers and connector support become part of the product promise, not back-office hygiene.