What’s new in Snowflake AI

We were at Snowflake SPN Connect Sydney last month — here’s what’s actually worth paying attention to. The clearest throughline across every session: AI is reshaping how people interact with data, and Snowflake is moving fast. Tools like Snowflake Intelligence are putting self-serve analytics directly in the hands of business users — no data team required for the day-to-day stuff. That’s not a threat to data teams. It’s a relief. Less fielding ad hoc requests, more building data products that people can actually use.

Cortex Code — the announcement worth knowing about

Snowflake’s Cortex Code (nicknamed CoCo internally) was the product that got people talking. Snowflake’s claim is that it outperforms Claude Code for Snowflake-native development. We haven’t stress-tested that ourselves yet, but if it holds up, it makes the platform meaningfully easier to work in — especially for teams spending real time writing and debugging Snowflake-specific code.

Cortex AI — the platform underneath it all

A lot of what was discussed at SPN Connect sits on top of Cortex AI — Snowflake’s suite of LLM-powered functions that run natively inside your Snowflake environment. The core value proposition is straightforward: your data doesn’t move, your governance layer stays intact, and your data team works in SQL and Python rather than learning a new stack. No separate AI infrastructure to manage, no new security approvals to chase.

What’s maturing fast is the combination of Cortex AI with a governed semantic layer. That’s the architecture that makes natural language querying actually trustworthy at enterprise scale — the LLM reasons over your business logic, not raw tables. Organisations that have that semantic foundation in place are already deploying Cortex Analyst and seeing business users get accurate, governed answers without writing a single line of SQL. Organisations that don’t have it yet are the ones finding that AI gives impressive demos and unreliable production outputs.

AI is adding value at two layers — and the distinction matters

Organisations are finding value in AI at two distinct points in their data platform. On the development side, AI tooling is catching issues earlier — query optimisation, pipeline code review, the kind of subtle bugs that historically surface at 2am when a critical ingestion job fails. On the consumption side, business users can now interrogate data through natural language instead of raising a ticket or navigating five dashboards to find one number. Getting the right information to the right person at the right time has been the stated goal of data teams for years. With Cortex Analyst and Snowflake Intelligence, it’s starting to actually happen.

What this means for your data platform

The hard truth: AI won’t fix a broken foundation

The most common mistake organisations are making right now is investing in AI before their data is ready for it. AI doesn’t fix poor data quality or inconsistent schemas — it amplifies them. Organisations that layer AI onto an immature data platform end up with unreliable outputs, eroded trust, and more complexity than value. The version of this we hear most often: teams building infrastructure in anticipation of hypothetical future needs, disconnected from any specific business outcome. At scale, that becomes very hard to unwind. Cost and observability sit alongside this. Cloud-native architectures are flexible, but complexity grows faster than most teams expect. Without observability across the platform, hidden costs accumulate — inefficient queries running unnoticed, unused datasets sitting idle, poorly designed workloads burning compute.

What we’re taking back to client work

Before adopting any new Snowflake-native AI capability, evaluate whether to build within Snowflake or use an external tool — the answer isn’t always Snowflake-native. Monitor credit consumption during the build phase, not just post-deployment. And get data governance in place before rolling out AI tools — classification, tagging, masking policies, row-level security. These aren’t steps to defer.