Snowflake-Anthropic Deal Drives Governed AI Agent Adoption

What happened

Snowflake and Anthropic signed a multi-year, $200 million agreement in December 2025. Ten months later, the deal has become one of the clearest examples of how enterprise AI is now sold. The model provider brings reasoning ability. The data platform brings governance. Revenue follows where the two meet.

The agreement made Anthropic's Claude models available to more than 12,600 Snowflake customers on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Azure. It also set up a joint go-to-market effort aimed at putting AI agents into the world's largest companies.1 The two companies had first worked together in November 2024, when Claude models arrived in Snowflake's Cortex AI development suite.2 Under the expanded deal, Claude Sonnet 4.5 became the engine behind Snowflake Intelligence, the company's natural-language agent for business users. Claude Opus 4.5 became available through Cortex AI Functions for multimodal queries written in SQL.68

The relationship has kept growing since then. At Snowflake Summit in June 2026, the companies announced that Snowflake was one of ten launch partners for Anthropic's Claude Marketplace. Through the marketplace, customers can apply their existing Anthropic spending commitments to Snowflake AI features.11 Claude models now power both of Snowflake's renamed flagship agents: CoWork for knowledge workers (formerly Snowflake Intelligence) and CoCo for developers (formerly Cortex Code).11 In early October, Claude Sonnet 5.5 reached Cortex AI in public preview through Cortex Inference, with support in CoWork, CoCo and Cortex Agents still to come.14

Why governance became the selling point

Across the coverage, one argument shows up again and again: enterprise AI pilots stall because of data access, not model quality. One technical guide describes the usual pattern. Each attempt to connect governed data to an outside model brings a new security review, another copy of the data and another bill, so months pass before a model answers anything useful.12 Snowflake's answer is to run the models inside its own service perimeter. Existing role-based access controls, row access policies and column-level security still apply, and Snowflake says customer data is not used to train models.1215

Anthropic CEO Dario Amodei made the same point when the deal was announced. Companies have spent years building trusted data environments, he said, and now want AI that works inside them without compromise.4 Both companies said regulated industries such as financial services, healthcare and life sciences are the main target. Snowflake's AI research chief Dwarak Rajagopal called the investment fuel for collaboration in those sectors.24

Snowflake now uses the phrase "agentic control plane" for a governance layer that coordinates data, models and enterprise systems. At its Paris World Tour in October, the company argued that the other parts of its platform would remain disconnected projects without that layer.16 The June acquisition of Natoma fits the same approach. Natoma adds governance for agents connecting through the Model Context Protocol (MCP), with audit trails and verified agent identities.11

The adoption numbers

The best evidence that governed AI is selling comes from Snowflake's earnings, and the numbers have risen steadily. In the quarter when the Anthropic deal was signed, Snowflake reported a $100 million annual AI revenue run rate, reached a quarter earlier than planned.25 Snowflake Intelligence had 1,200 customers at that point. Management said AI influenced half of the bookings signed in the quarter.27

By February, Snowflake Intelligence had more than 2,500 accounts, and more than 9,100 accounts were using Snowflake AI in some form.26 Snowflake also announced its largest contract ever, worth more than $400 million.22

In September, Snowflake reported fiscal second-quarter product revenue of $1.49 billion, up 37% from a year earlier. It was the third straight quarter of faster growth, and management raised full-year product revenue guidance to $6.07 billion.3134 Executives said AI products accounted for about half of the acceleration.31 Several reports say CoCo passed 9,100 accounts after adding more than 2,000 in the quarter, while CoWork reached 5,800.353940 The figures are not fully consistent. One published transcript of the call attributes the 5,800 figure to Cortex Code (CoCo's former name).31 Most coverage, including the company's results summary, assigns it to CoWork.40

The usage data supports the same picture. When the deal was announced, both companies said thousands of Snowflake customers were already processing trillions of Claude tokens a month through Cortex AI.15 Snowflake reported that customer use cases on the platform rose 89% year over year.34

Case studies, and their limits

The return-on-investment evidence remains mostly anecdotal and vendor-supplied, which matters for anyone trying to judge the payoff. Intercom uses Claude through Cortex AI to run its Fin AI Agent. Its engineering leadership says the setup raised automation rates for customer support volume and made work with its largest customers more efficient.1 Simon Data says it uses Claude on Snowflake to find patterns in its data while keeping governance controls in place.1 Anthropic cites Snowflake's internal benchmarks claiming more than 90% accuracy on complex text-to-SQL tasks. Those are the company's own tests, not independent ones.1

Snowflake is also its own case study. It uses Claude Code across its engineering organization, and a Claude-powered sales assistant built on Snowflake Intelligence is meant to shorten deal cycles.59 On the September call, management linked internal AI tools to a 43% rise in use cases won per account executive. It also said bringing search optimization in-house with AI cut $400,000 a year in agency spending.31 Snowflake added that new customers are reaching 80% of the consumption they bought more quickly, because AI is speeding up project delivery.31

These results come from the vendor or from customers it chose. The more independent signal is the financial one. Net revenue retention was 126%, and the number of customers spending more than $1 million a year rose 27%.3140 Both measures suggest existing customers are spending more, which is harder to fake than a showcase customer story.

Where the coverage diverges

Analysts largely agree that the partnership matters, but they disagree on why. ISG's David Menninger argued that the dollar figure matters less than the commitment of scarce resources. He noted that Anthropic "stays in its lane" and does not compete with Snowflake's platform business.2 Omdia's Mike Leone read the deal more pointedly: Snowflake needed a top-tier model to close the reliability gap in agentic workflows and to keep the heavy processing on its own platform.2

The deal is not exclusive, and some coverage gives that too little weight. Anthropic also works with Databricks. Snowflake signed a separate $200 million agreement with OpenAI.222 Cortex AI also offers models from Google, Meta, Mistral, DeepSeek and xAI.12 In Paris, Snowflake openly warned customers against depending on any single model provider.16 The more accurate reading is that Snowflake is selling governed access to whichever models perform best. Claude currently has a favored position within that offering, not a lock on it.

The promise that data never leaves the perimeter also deserves scrutiny. A practitioner analysis points out that Snowflake's guarantee about data staying in its governance boundary applies mainly to Cortex Analyst's default models. Those defaults are Snowflake-hosted Mistral and Meta models. Anthropic and OpenAI models can involve cross-region inference.13 When Claude or ChatGPT queries Snowflake through its managed MCP server, results go back to that outside client.13 One developer walkthrough also found that misconfigured access tokens can let unauthorized roles see tools they were never granted.19 Governance works only when it is configured correctly.

The cost question

AI margins are the clearest risk. Snowflake cut its full-year product gross margin target to 74%, citing a larger share of fast-growing AI workloads that are less profitable for now.31 As early as February, executives were fielding questions about "sticker shock" as agents multiply. CEO Sridhar Ramaswamy promised per-user spending caps to make costs more predictable.26 Separately, observers note that Snowflake does not report AI revenue as its own line.39

The takeaway

The Snowflake–Anthropic partnership shows a structural change in enterprise AI. Models are becoming interchangeable, and governed data with business context is becoming the scarce asset. The Summit recap made the same point, describing models as a commodity and governed enterprise data as the asset that is not.11 Anthropic gets a large distribution channel into regulated enterprises. Snowflake gets a top model that brings more workloads onto its platform. The open question is no longer whether companies will use agents on governed data. It is whether vendors can keep those agents affordable, auditable and correctly configured as deployments grow beyond early adopters.