Carvana turns Slack alerts into production fixes with Claude Tag

Overview

Carvana, the online car retailer, replaced two homegrown Slack bots with Claude Tag, giving individual teams their own scoped data-warehouse access via per-team identities and bundles of instructions, tools and access. Claude Tag monitors Slack channels for production monitoring alerts, automatically investigates root causes, identifies the owning team, and can open a pull request for review before a human merges the fix. The rollout cut alert volume in the retail science channel by 56% by resolving recurring root causes, and reduced time-to-answer for the wholesale platform team's issue and alert triage by 65%.

The challenge

Production alerts at Carvana used to wait for an engineer to manually investigate. Carvana had already built two homegrown Slack-based AI tools internally, but maintaining them required time and resources that could be better spent running the business, and extending them with fine-grained, team-specific access controls would have required significant additional engineering work. Carvana's wholesale platform team declined to adopt the homegrown tools over uncertainty about ongoing support.

The solution

Claude Tag monitors the Slack channels where production monitoring alerts land and begins investigating as soon as an alert arrives, identifying the team that owns the affected code and posting a root-cause analysis in that team's preferred format. Teams use "bundles" - packaged instructions, tools and access scoped to one team's use case - dropped into a single Slack channel, and Claude Tag runs under its own scoped identities so each team gets its own slice of Carvana's data warehouse. Workflows run from a Slack alert through investigation, a pull request, human review and merge; some teams also schedule Claude Tag to run hourly checks on pipelines and feature flags.

Reported business value

Claude Tag cut alert volume in Carvana's retail science channel by 56% by resolving root causes rather than repeatedly clearing the same alerts, and reduced time-to-answer for the wholesale platform team's issue and alert triage by 65%. It replaced two homegrown Slackbots, gave each team its own scoped data-warehouse access, and now runs hourly checks that catch failing pipelines and stale feature flags.