Stack OverflowEira MayProblem & solution
Chaos, then control: the maturity model most AI rollouts skip over
Summary
Many engineering organizations see uneven results from AI coding tools despite high adoption because they skip the crucial middle stages of a proper rollout. Vivek Raghunathan, SVP of engineering at Snowflake, advocates a five-stage progression: let chaos reign, discover what works, codify it into a shared language, rein in chaos by scaling that language, then reorganize. The skipped stages—discovering effective practices and naming them as patterns—turn individual competence into organizational knowledge. Raghunathan’s team built a pattern language of roughly 14 named practices, including planning in writing before coding and fencing agents into separate git worktrees. They spread these practices through structural mechanisms like focus weeks rather than mandates, and resist reorganizing until the earlier stages have produced reliable knowledge about what AI-assisted work looks like.
Context
Vivek Raghunathan, SVP of engineering at Snowflake, describes a five-stage AI rollout maturity model in a Leaders of Code conversation, emphasizing the skipped middle stages of discovery and codification.
Approach / What changed
A five-stage AI rollout progression: let chaos reign, discover what works, codify into a shared pattern language, rein in chaos by scaling that language via structural mechanisms like focus weeks, then reorganize around the new reality.
Takeaways
The gap between high adoption and uneven results is usually caused by skipping the middle stages of discovery and codification, not by tooling or talent.
Raghunathan’s team identified roughly 14 named patterns from engineers’ successful experiments, including planning in writing before coding and using separate git worktrees for parallel agent work.
Spreading effective practices requires structural mechanisms like dedicated focus weeks rather than mandates, because new habits cannot be internalized during an overloaded sprint cadence.
FAQ
Why do AI rollouts produce uneven results despite high adoption?
Most orgs skip the stages of discovering what works and codifying it into a shared language, leaving individual competence unshared. This produces a gap between engineers who become dramatically more productive and those who use AI as glorified autocomplete.
What metrics should be tracked in early AI rollout phases?
In the early months, only adoption itself should be tracked—whether people use the tool for a range of tasks like writing code, reviewing code, and drafting design docs. Output metrics like lines of code or PRs merged are easy to game and don't indicate whether people are getting better at using the tools.
How can an organization spread effective AI practices without mandates?
By building structural mechanisms like focus weeks where the entire org sets aside dedicated time to learn the paved path or forge new ones. Mandates don't work because new habits cannot be internalized during an overloaded sprint cadence.
Source and freshness
- Source
- stackoverflow.blog
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.