Databricks vs Snowflake & A Fairly Objective, Opinionated Journey: Let's Talk GenAI (Written Mid April, 2025)

An Objective But Opinionated Take On Available AI Options From Both Vendors

Note: Article was originally published about a month ago as I post it on Substack. Some of the elements may have changed since the time of writing.

Let's Be Clear & Upfront About My Background

-I work for a consulting firm (zeb) that does a lot of work on AWS & also Databricks. Part of my contract’s terms are that I have 100% editorial freedom on what I write about, including the freedom to both speak positively or negatively about Databricks.

-Part of the Databricks MVP program & Product Advisory Board, both recognitions received after being a customer of Databricks across different companies for 3 years.

Since there is a lot of Databricks above, here is what you should know about my experience & knowledge of Snowflake:

-Instructor-led training three years ago, a couple of Udemy courses, a few weeks of production work, and a POC (Snowflake was my first choice before Databricks, but Databricks won). I regularly move data between the two platforms, have done non-pipeline client work, and spend ~10–20 hours a month on Snowflake hands-on or researching its documentation. While I haven’t lived its day-to-day for long, I have a decent grasp of its strengths and limitations.

-I generally believe Databricks offers the best business value and hence why I am so invested in the ecosystem. At the same time, this article is written with the intention of helping you gain understanding & empowering you to ask the right questions when making your choice, no matter which one it is.

Today’s Topic: Let’s Talk GenAI

Both platforms offer AI features you can use with SQL, support multiple LLM models, offer chat-like capabilities, and offer “Agentic” capabilities.

Models

Both platforms offer multiple LLM models and are both racing against each other to make sure they support the latest and greatest. If you see an announcement from one saying they support a new model, chances are, you will see a similar post by the other making a similar statement.

The way they implement them might be different at times, but generally, you can’t go wrong either way in this subcategory.

SQL AI Capabilities

Both platforms have a handful of SQL AI functions that have out-of-the-box capabilities such as sentiment analysis or classification. They also offer the availability to utilize custom prompts inside of SQL with the model and prompt of your choices.

I will call out, that I appreciate Snowflake’s “COMPLETE Structure Output”, as well as Databricks’ approach to forecasting with ai_forecast().

In this category, I believe that cost & performance are key, and I have not seen a published comparison that felt robust enough to persuade me one way or the other. So if you are evaluating between the two platforms, I’d run tests on your own data to help you make the right decision. Keep in mind both token related costs, as well as the cost of running the compute.

Conversational AI

When it comes to chatbot/conversational AI experience, they differ meaningfully in how usable and integrated those experiences are, especially for business users.

Databricks offers AI/BI Genie with a built-in, stand-alone UI as well as an API for embedding it into your own applications. You can “train” Genie through custom instructions, have it leverage functions you can create outside of Genie, and most importantly, teach it to give you what you need through incremental chat-like interactions that are not overtly technical. It leverages the “Comments” (descriptions) of your data that are part of your Unity Catalog, a central repertoire of knowledge.

Snowflake’s approach is through a collection of components: Cortex Analyst (natural language to SQL via semantic modeling), Cortex Search (document-level hybrid search), and Cortex Agents for orchestration/routing. While there’s no polished, production-grade UI out of the box, it integrates well with Streamlit as seen in Snowflake’s quickstarts to help you build a basic UI. “Training” here really means defining a semantic model in YAML which contains business terms, table mappings, example queries, etc.

In terms of the chatbot-type experience, I have a hard time with Snowflake’s approach to configuration/training. I find it unintuitive and favors the engineering type much more than the business and analyst types. The lack of leveraging the centralized field definitions is also something I quite don’t understand as this could lead to a sprawl of definitions of the data. The one positive is when a high level of customization is required around definitions for different business use-cases, this could be an asset, but generally speaking, I am not a fan of this approach.

Databricks’ Genie does have its flaws. I’ve historically had little success with generic instructions inside of Genie working as intended and sometimes I’ve found it to be too conservative with some of its answers, though on the other hand, this can be a good thing.

At the end of it all, you can probably do most of the things that you can do with Snowflake’s approach inside of Databricks. On the flip side, you’d have to spend a lot of time and money building something in Snowflake that recreates what Databricks’ Genie does as well as it does, both out-of-the-box as well as with proper configuration & to a certain degree, governance.

Would I switch platforms purely for this subcategory? Unlikely. But if I were starting fresh or migrating from something else, this could carry significant weight in the decision.

Agentic

Agentic AI is a loaded term with much marketing, hype, and skepticism around it. Often, I roll my eyes when people mention it, and other times, when I see solutions addressing real problems, I am impressed and optimistic. Because of that, depending on who you ask, agentic AI means something different and therefore, expectations differ from person to person.

So, here is my best shot at comparing Databricks vs Snowflake when it comes to the more general agentic capabilities, without going into the industry fluff.

You have a lot more freedom as to what is possible with Databricks, from satisfying basic needs, all the way to complex needs than what Snowflake has to offer, specially with its approach to custom functions. It is important to note still, that Snowflake has done a good job at incorporating a lot of core functions/elements to handle common customer use-cases.

A lot of the success stories i've heard around both of these platforms also reflect this. From Snowflake's side, I tend to hear more about streamlined situations that all tend to sound the same, while Databricks' tend to be a mix of streamlined situations, but more so, novel use-cases.

In the end, personally, the flexibility of being able to leverage Genie's API and everything it is capable of doing out of the box, in addition to the general Databricks agentic capabilities + Databricks' approach to the AI Playground give Databricks an edge in this category. This is one where I would argue that unless Snowflake covers a need out of the box and with a quickstart, the category would be heavily weight toward Databricks.

It is important to call out: Neither platform makes building agentic workflows a breeze for analysts or non-engineers. So, if the magic of robots that do all your work with minimal configuration & pain is what you are looking for, neither of the two options offer that, unless the conversational AI piece is a huge part of it, in which, I believe Databricks takes it home.

Closing

I hope you enjoyed reading this article, with hopefully more articles to come in this series with a similar approach.

Very much welcome respectful commentary and requests on areas to cover next.