.png){kind=small}
Snowflake Summit 2026: What Matters Most for P&C Carriers
Part 2 of the Snowflake Changes that Matter Series
Snowflake Summit 2026 generated more than two dozen distinct announcements, two product rebrands, a $6 billion cloud commitment, and an acquisition. This article aims to summarize the new feature announcements and understand what is available today versus what is in preview.
The most useful insight of the week didn't come from a product demo. It came from CEO Sridhar Ramaswamy's framing in the opening keynote. A frontier model is something anyone can rent, so on its own it buys you no advantage; rivals can run the identical one too. The asset your competitors can't replicate is the data you've spent years accumulating. For a P&C carrier, that should register differently than it does for a retailer. Your loss history, your premium pricing history, the adjuster narratives buried in your claim files, the underwriting judgment encoded into your rating plans, that is the moat. Everything Snowflake shipped this year is, functionally, plumbing to let an AI agent act on that moat without breaking it.
1. The conversation moved from governing data to governing actions
The pivotal shift this year was conceptual. The old question was who's allowed to see a given table. The 2026 question is what an autonomous agent is permitted to do once it has access. Snowflake's answer is AI Agent Identity, now in private preview. Each agent gets a unique, recognizable identity called the Agent Principal, scoped to the user's session. The agent still operates on the user's behalf, but governance policies can now detect when an agent is driving the session and restrict it accordingly, such as limiting it to read-only access or blocking sensitive data entirely. Every action it takes stays attributable. Snowflake reinforced this with its acquisition of Natoma, whose MCP gateway inspects every individual tool call, verifies who's behind the request, checks whether policy allows it, and records the outcome.
For a regulated carrier, this is not merely an IT improvement; it is a compliance issue. An agent that drafts coverage, triages a submission, or touches a claim falls under the same scrutiny by NAIC and state regulators. The ability to show which agent took which action, and under whose authority, is quickly becoming a prerequisite for deploying agents anywhere near underwriting or claims. This is where AI Agent Identity would play a pivotal role.
2. Context is where the real contest is, and it's the dull layer that settles everything
Having an agent correctly answer a question like “what's our catastrophe-exposed premium in coastal Florida, net of reinsurance?” is almost entirely a question of semantics. Snowflake gave this real airtime at the Summit all week.
Horizon Context layers business meaning onto the Horizon Catalog, pulling metadata from outside tools (Power BI, Tableau, dbt, and SQL Server among them), enriching it with lineage and generated documentation, and serving it to agents when they query.
Cortex Sense assembles that context on the fly for each user and role, with Snowflake reporting a meaningful lift in answer accuracy when it's active.
Semantic Studio lets business users state what a metric means without touching SQL.
Snowflake CoWork (formerly Snowflake Intelligence) is a conversational agent for business users. It reasons across both structured and unstructured data, runs deep multi-step research, and returns cited answers rather than a dashboard you still must interpret. For a carrier, the capability that matters most is the one that reaches unstructured data. An agent that can read across claims notes, submissions, and policy contracts, then provide summarized information, is doing work no dashboard reaches. With Snowflake CoWork every chart and conversation can be stored as an Artifact. What one analyst figures out stops being a one-off result and becomes something the rest of the team can reopen, reuse, and build on and continue the conversation.
Our caution is to build the semantic layer first. An agent built on loose definitions doesn't fail loudly; it answers with confidence and gets it wrong, which in reserving or underwriting is precisely the expensive failure mode. One caveat is that several of these pieces (Cortex Sense, Agent Studio) remain in private preview.
Snowflake CoCo, the AI coding agent for data, is the counterpart for your data and platform teams. It builds pipelines, semantic views, and migrations under the same RBAC governance, across Snowsight, the IDE, and the terminal.
3. Putting all your data in one tool is over
The clearest architectural signal was Snowflake choosing open interoperability over lock-in; one governed copy of your data that any engine can use in place, rather than a second copy trapped in a proprietary format. Three pieces make that real. Apache Iceberg v3 reached general availability and pushed the open table format into the data types that used to force workarounds, namely, semi-structured (Variant), geospatial, and change-data-capture. Horizon Catalog, built on Apache Polaris and the open Iceberg REST standard, is the control plane, with outside engines now able to both read and write Snowflake-managed Iceberg tables, both generally available. And Snowflake Storage for Apache Iceberg makes that a single managed table you stand up with one command, with predictable cost and enterprise resilience, tuned for Snowflake but open to any engine.
No industry runs a more tangled data estate than a P&C carrier: policy admin on Guidewire or Duck Creek, legacy mainframe, and a number of third-party feeds from geospatial to weather. The payoff lands precisely there. One governed copy every engine can reach, with Iceberg v3 now handling the geospatial and semi-structured data those feeds carry, quietly resets the build-versus-buy calculus on moving data and cuts the risk of betting everything on one platform.
4. The AI cost counterweight
AI consumption drives spend upward. That's a certainty, not a risk. Snowflake's response is Adaptive Compute, which takes warehouse sizing out of human hands and tunes it automatically; the company cites gains in the range of 1.6x on analytical work and up to 3.5x on transformation-heavy jobs.
While Adaptive Compute automates the warehouse, Snowflake Optima automates the query, on the company's wager that it should be impossible to run inefficiently. It bundles automatic optimizations that learn from your query history with no DDL, no configuration, and no separate billing. Optima Indexing builds hidden search indexes for recurring point-lookups; Optima Metadata materializes custom pruning metadata for hot predicate patterns to cut I/O, and Optima Planning tunes execution from prior performance. All three are GA on Gen2 standard and Adaptive Warehouses, included in every Snowflake edition at no additional cost. One pricing note: Adaptive warehouses bill on a per-query basis rather than compute-based pricing for standard warehouses. One caution for clients weighing the switch: Adaptive Warehouses are not a guaranteed cost saver and can raise spend, since the platform sizes every query for the most performance-sensitive workload, leaving you without the per-query control to trade speed for savings.
Conclusion
Two sobering realities to plan against. First, a meaningful share of the most exciting capabilities is still in preview rather than GA, useful for shaping a roadmap but not for committing to production. Second, and weightier, Accenture's leaders made the point from the mainstage that in most of their engagements (they put the figure near 85%), the binding constraint is the condition of the data, not the AI sitting on top of it. For carriers carrying real legacy debt, that is the gate. No degree of agentic polish offsets inconsistent loss codes, undocumented lineage, or a semantic layer nobody trusts.
The posture that wins
The carriers who pull ahead won't be the ones who switch on the most features. They'll be the ones who get the unglamorous foundation right and then add agents on top with governance built in from the start rather than bolted on after an audit finding, and that foundation means clean, governed, clearly defined data.
Comments