Retail, 21 October 2026
Carvana: A Slack-Native AI Teammate That Triages Its Own Production Alerts
Carvana — a North American online used-car retailer, known for letting customers buy a car entirely from their phone and pick it up from one of its glass “car vending machine” towers.
Problem
Carvana’s engineering teams live in Slack, and production alerts land there constantly: a pipeline breaks, a feature flag misfires, a release ships with an unexplained spike in errors. Before anyone could fix anything, someone had to notice the alert, figure out which team owned the broken code, and dig through logs and tickets to find the actual cause. Carvana had already tried to automate this twice — first with a homegrown “API call plus a few tools” agent loop, then with a Claude Agent SDK app wrapped in a Slack bot. Both worked well enough, but keeping them running ate the time of a small team, and they offered weak per-team access controls. The wholesale platform team had declined to adopt either one, unconvinced anyone would keep supporting it.
Result
56%
- Alert volume in the retail science channel fell 56% — not because alerts were silenced, but because teams actually fixed the recurring root causes Tag kept surfacing.
- The wholesale platform team — the one that had twice refused to adopt a homegrown bot — now reaches an answer 65% faster.
- Two homegrown Slackbots were retired entirely.
- Engineers report filing more issues than before, because they now expect each one to get carried toward an actual fix rather than sitting untouched.
- Half a dozen teams are now building multi-step, multi-person workflows on top of it.
Solution
Carvana switched to Claude Tag, Anthropic’s always-on Slack teammate (currently in beta for Claude Enterprise and Team customers): mention it in a channel, and it works the alert and replies in a thread. Each team gets its own bundle — a scoped set of instructions, tools and data access, tied to a single Slack channel, running under its own identity rather than one shared login:
- The retail science team gets alerts that arrive already investigated — Tag names the team that owns the broken code, tags them, and writes a root-cause analysis in the team’s own fixed format; engineers then ask it to open a pull request straight from that analysis.
- The customer verification team gave Tag a skill, stored in its own repo, that compares each release’s error rate, warning count and log volume against the hour before deployment. When one release produced an unexplained spike with no obvious cause, Tag opened a pull request adding better logging so the next one would be easier to diagnose.
- The wholesale platform team — the one that had said no twice before — gave Tag read-only access to trace issues through tickets, build history and code, and to look up a vehicle’s status by running its VIN through the data warehouse.
- Several teams also run Tag on a schedule, checking pipelines and feature flags hourly and kicking off follow-up work when something looks wrong.
Tag defaults to Claude Opus 5 for the harder root-cause work, with Sonnet and Haiku handling lighter workflows; a handful of non-engineering power users — including a senior attorney reviewing large sets of transaction documents — use Fable instead, precisely because, in her words, “a confident wrong answer is far worse than no answer.” To migrate off the old chatbot, an internal testing group had Tag analyze its output first, to learn exactly what it needed to replace.
Could a small team do this?
Partly yes, and the useful detail is exactly which part. Claude Tag itself — the governed, multi-team product with per-team identities and access-control slices — is a hosted Anthropic beta for Enterprise and Team plans; a solo operator can’t rebuild that infrastructure, and shouldn’t try to. But Carvana’s own first attempt, before Tag existed, was a plain Claude Agent SDK app wrapped in a Slack bot with a handful of tools — and it worked well enough that three separate teams used it. That version is squarely solo-buildable: the Slack Events API to catch a mention, the Claude Agent SDK to read logs, tickets or a repo, and the GitHub API to open a pull request from the findings. A single-channel version of this — watch our own alerts, write a root-cause note, open the fix as a PR — is a realistic one- or two-week build, and it’s exactly the kind of thing worth running on OpenClaw’s own pipeline first before offering it as a service. What’s genuinely hard to replicate solo isn’t the AI step at all; it’s the governance Carvana valued enough to abandon a working homegrown tool for — scoped access per team, and someone else responsible for keeping it running across dozens of teams long-term. The realistic path: build the single-team alert-to-root-cause-to-PR loop, prove it on one real codebase, and only worry about multi-tenant access control once a second paying team actually needs it.