Snowflake Cortex AI: My Honest Take, Feature by Feature

I’ve spent the last decade watching Snowflake reinvent itself every couple of years. Cortex is the reinvention I didn’t see coming, and the one I think is most underrated.

Most people still think of Snowflake Cortex as a chatbot bolted onto SQL. I used to think that too, until I actually built on top of it.

My take, after working with it hands-on: Cortex isn’t a feature. It’s a bet Snowflake is making — that the next generation of AI applications gets built on top of governed data, not next to it. Once you see it that way, every individual piece of Cortex reads differently.

Here’s where I’ve landed on each of them so far.

Cortex Functions — the piece I underestimated the most

My first reaction to Cortex Functions was “fine, LLM calls as SQL functions, nice.” I was wrong to dismiss it that fast.

Sentiment analysis, summarization, translation, text completion — all callable directly in a SELECT statement, with no API key, no external call, no data leaving the account. The thing I didn't appreciate until I used it in a real pipeline: the moment an LLM call is just another SQL function, it stops being a special case in your architecture. It's a column transformation like any other. That's a bigger shift than the feature list makes it sound.

Cortex Analyst — the one I trust the most, and the one I’m most skeptical of

Cortex Analyst is text-to-SQL but grounded in a semantic view rather than guessing at raw table names. That distinction matters more than it sounds — I’ve seen enough naive text-to-SQL demos fall apart the moment someone asks a question the model has to interpret rather than pattern-match.

My honest opinion: Analyst is only as good as the semantic model behind it, and building that model well is real work, not a checkbox. I’m still forming a view on how much that work scales past a first few use cases — that’s exactly what I want to dig into next in this series.

Cortex Search — the quiet workhorse

Semantic and keyword search over unstructured data — PDFs, tickets, transcripts — without standing up and babysitting a separate vector database. This is the least flashy piece of Cortex and, in my opinion, the one that will end up doing the most unglamorous heavy lifting in production. Nobody writes a keynote slide about “one less system to keep in sync,” but every engineer who’s maintained a bolted-on vector store knows exactly what that’s worth.

Cortex Agents — the piece I haven’t fully made up my mind on yet

Agents route a single natural-language question across Analyst and Search automatically, answering both “what happened” and “why” in one interface, grounded in your actual RBAC. On paper, this is the orchestration layer that ties everything else together — and it’s the piece I’m least experienced with hands-on so far. I want to reserve judgment here until I’ve actually shipped something with it, rather than repeat what the docs promise.

The thread that connects all four, in my view

A few months into building out an IoT logistics platform, we hit the moment every team eventually hits: someone in the room asked whether the AI layer we wanted to bolt on could see data it shouldn’t. Fleet telemetry, driver records, customer shipment details — all sitting under different role hierarchies for good reason. The honest answer, with a bolted-on external LLM pipeline, would have been “we’d have to rebuild the access control logic a second time, outside the warehouse, and hope it never drifts out of sync.” That’s not an answer you want to give in a security review.

That’s the conversation that made Cortex click for me. Because the models run inside Snowflake’s own compute layer, the question stopped being “how do we replicate our RBAC in a second system” and became “does our existing RBAC already cover this” — which, if you’ve modeled your roles properly, it already does.

If you’ve worked in finance or healthcare, you already know the real blocker to AI adoption was never model quality. It was always:

Can we prove this system only saw what this user was allowed to see?

That’s the boring, unglamorous question Cortex actually answers, and it’s why I rate it higher than most of the flashier AI announcements this year.

Where do I go from here

This is my read on Cortex at the point I’m at with it today — more confident on Functions and Search, still forming a view on Analyst at scale, and reserving judgment on Agents until I’ve built with them directly. As I go deeper into each piece, I’ll keep updating this take rather than pretending I’ve already got it all figured out.

If you’ve formed a different opinion building on any of these, I’d genuinely like to hear where you land.