Snowflake Engineering Chief Vivek Raghunathan Details 5-Stage AI Rollout Model

Snowflake engineering chief Vivek Raghunathan argued this week that most enterprise AI rollouts fail because they skip the middle work of discovering, naming, and spreading effective practices before redesigning roles, a pattern he laid out in a five-stage maturity model shared during a recent Leaders of Code conversation.

The problem isn’t licensing or talent, Raghunathan said. Nearly every engineering organization has bought AI coding tools and posted strong adoption numbers. Six months later, outcomes are wildly uneven: a handful of engineers ship solo projects that once required teams, while the majority treat the agents as glorified autocomplete.

That gap gets misdiagnosed as a hiring problem or a tooling deficit. The real issue, he told the audience, is the absence of a deliberate stage that turns individual experiments into organizational knowledge.

Raghunathan borrowed a maxim from Intel’s Andy Grove: during any platform shift, you have to let chaos reign before you can rein it in. Step five cannot happen until the previous four have done their work. Most companies, eager to show return on investment, compress or skip stages two and three entirely.

The trap of early metrics

“phase, the only metric that matters is adoption itself: Are people actually using it? Are they using it for more than one kind of task? Everything else is noise until you know what” Raghunathan said.

Lines of code, pull requests merged, and tickets closed are easy to game and don't reveal whether people are getting better at using the tools. Everything else is noise until the team knows what “good” looks like.

Building a pattern language from the inside

The stage most rollouts skip entirely is codification. Once people experiment, some build genuinely better ways of working. If nobody watches for those successful experiments, they stay locked in individual heads.

Raghunathan’s team went looking for them, asking engineers about the best practices they had developed. They ended up with roughly 14 named patterns, deliberately modeled on the “Gang of Four” software design patterns, giving the team something concrete to teach and measure against.

“A pattern language gives a team something concrete to teach and measure itself against.”

— Vivek Raghunathan, SVP of Engineering, Snowflake

Mechanisms beat mandates

Once a set of practices works, the instinct is to announce them and expect adoption.

Org charts come last

The last stage—changing roles, responsibilities, and org charts—is the one everyone wants to talk about first because it’s visible and dramatic. Skipping ahead usually leads to restructuring around guesses. The proverbial cart precedes the horse.

These middle stages are unglamorous and genuinely hard to run. Codifying practice requires someone senior enough to translate findings into something teachable—a real time investment that feels skippable when plates are full.

Buying licenses is a decision you make once. Building a pattern language, a mechanism to spread it, and eventually a new org structure is an ongoing practice, and it’s much harder to build a practice than to make a purchase.

Snowflake, founded in 2012 by former Oracle architects Benoît Dageville and Thierry Cruanes alongside Vectorwise co-founder Marcin Żukowski, pioneered the separation of compute and storage in a cloud-native data warehouse. After emerging from stealth in 2014 with 80 customers, it expanded across AWS, Azure, and Google Cloud, reaching $4.72 billion in annual product revenue by fiscal 2026 and counting more than 9,100 accounts using its AI features.

Raghunathan’s closing guidance: resist grading the rollout before you understand the range. Go looking for what’s already working, on purpose. Give the pattern language a name and a distribution mechanism. Hold off on the org chart.

If your rollout has adoption numbers but no shared vocabulary for what “good” looks like, ask which stage you’re actually in—and who is responsible for getting you to the next one.

Photo of author

Technology Desk

Technology Desk at Archyde.

Lehmann Denies Knowledge of Sandpaper Use, Contradicts Bancroft’s Account