Instacart's Blueberry Bot Helps On-Call Engineers Fight Fires in Slack

The On-Call Struggle Is Real

When production goes down, the first thing an on-call engineer does is not fix the problem. It's figuring out what the problem even is. You have to check which service owns the failing component, scan recent deploys, dig through logs and dashboards, search internal docs, and try to remember if you've seen this symptom before. That context-gathering phase can eat up precious minutes—or worse, hours—while customers are staring at error messages.

Instacart knows this pain well. The grocery delivery company runs a massive, distributed system, and their engineers were spending too much time just getting up to speed during incidents. So they built an AI assistant they call Blueberry.

What Blueberry Actually Does

Blueberry is not another chatbot that spits out generic advice. It's an AI-powered incident response system that integrates directly with Slack, where Instacart's on-call workflows already live. When an alert fires, Blueberry spins up about ten sub-agents in parallel. These agents pull information from internal tools—service ownership data, logs, metrics, deployment history, and even 14 years of past incident reports—and then post a root-cause hypothesis right into the Slack thread where engineers are already working.

The whole thing takes about three minutes. In April alone, Blueberry ran roughly 25,000 diagnostics across more than 270 Slack channels.

From 60% to 90% Accuracy

One of the biggest challenges with AI in production operations is making sure the advice is actually grounded in reality. Instacart says Blueberry's diagnostic accuracy jumped from 60% to over 90% once they connected it to their historical incident data. That's a huge leap, and it shows that the quality of the underlying data matters just as much as the model.

Blueberry doesn't take action on its own. It gathers information, generates hypotheses, and helps with debugging—but a human engineer always makes the final call. That's a smart design choice, because in a production environment, you don't want an AI making unilateral changes.

Built on a Tool-Aware Framework

Instacart's engineers designed Blueberry with a persistent, tool-aware framework. The system maintains a catalog of MCP (Model Context Protocol) tools, keeps state between steps, and includes a review loop for continuous improvement. That means Blueberry can call out to various systems, retrieve what it needs, and remember where it left off if the investigation spans multiple steps.

In one month, Blueberry made over 58,000 MCP tool calls and achieved a 99.9% workflow success rate. It also adapts to the working patterns of about 60 different teams, which is no small feat given how varied engineering cultures can be.

Why This Matters for Collaboration Tools

Blueberry is a prime example of how AI is reshaping collaboration tools. It's not just about making meetings easier or automating calendar invites. It's about bringing intelligence directly into the tools where work actually happens—in this case, Slack threads during an incident.

For on-call engineers, the benefit is huge. Instead of starting from zero, they now get a head start. As Instacart's software engineering director Alan Wong put it, Blueberry changes the starting point for on-call engineers by providing relevant context before the deep dive begins. Engineers can begin troubleshooting with logs, deploys, and related system info already in hand.

The Power of Institutional Knowledge

Another key aspect of Blueberry is that it doesn't forget. The system captures operational knowledge from each incident and integrates it with past cases and team-specific context. That means the next time a similar issue pops up, engineers can draw on previous experience without having to dig through old tickets or ask around.

This is a big deal for collaboration. In many organizations, tribal knowledge lives in people's heads or in scattered documents. Blueberry makes that knowledge accessible and actionable, right when it's needed.

Not Just About the Model

Instacart's experience with Blueberry highlights a broader lesson: effective AI for operations isn't just about the large language model. It's about the surrounding engineering framework—the context, the workflows, the tool integrations, and the feedback loops.

Without those, an AI assistant is just a fancy autocomplete. With them, it becomes a true force multiplier. As Instacart's VP of Engineering Siby Alappatt said, Blueberry has proven to be a force multiplier, changing the on-call experience and helping teams quickly troubleshoot and mitigate complex production issues.

What's Next for AI in Incident Response?

Instacart's CTO Anirban Kundu sees Blueberry as part of a broader exploration into AI agent systems. The company is likely to expand this approach beyond incident response, maybe into other areas where engineers have to sift through large amounts of data quickly.

For now, Blueberry is making a tangible difference in how Instacart handles outages. It's not replacing engineers—it's giving them superpowers. And that's exactly what good collaboration tools should do.

If you're building or evaluating AI-powered collaboration tools, take a page from Instacart's playbook: ground your AI in real data, integrate it where people already work, and always keep a human in the loop.