How Instacart's Blueberry AI Gives On-Call Engineers a Head Start

When the Pager Goes Off

It's 2 a.m. and your phone buzzes with an alert. The grocery delivery app is down, and customers are seeing errors. You roll out of bed, open your laptop, and dive into a sea of logs and dashboards. For most engineers, this is the start of a frantic treasure hunt. But at Instacart, they're trying to make those first minutes less painful with a tool called Blueberry.

Blueberry is an AI assistant that lives in Slack, right where the on-call team is already chatting. When an alert fires, it springs into action, pulling together service ownership details, deployment history, logs, metrics, and past incident reports. Within minutes, it hands the engineer a summary: here's what's changed, here's what's odd, here's a guess at what's wrong.

Sounds nice, right? But let's be honest—AI in ops has a history of being more hype than help. So how does Blueberry actually work, and is it worth the buzz?

The Grunt Work of Context

Think about what happens when an incident hits. The hardest part isn't fixing the bug—it's figuring out where to start. You need to know: which service owns this alert? What changed in the last deploy? Have we seen this pattern before? That's the grunt work of context gathering, and it can eat up precious minutes when every second counts.

Blueberry automates that grunt work. It's like having a research assistant who never sleeps, who can read a thousand logs in a blink, and who remembers every incident from the past 14 years. The system orchestrates about ten sub-agents in parallel, each hunting for clues across Instacart's internal systems. They cross-reference what they find against the company's incident archive, and within about three minutes, they come back with a root-cause hypothesis—backed by evidence, not just vibes.

One of the engineers on the team told me that Blueberry's trick is that it's not a generic chatbot. It's wired into Instacart's specific setup: service maps, on-call rotations, deployment pipelines, debugging signals. So when it suggests something, it's grounded in how Instacart actually runs, not some textbook theory.

Numbers That Make You Look Twice

Instacart shared some stats from Blueberry's deployment, and they're hard to ignore. In April alone, Blueberry completed about 25,000 diagnoses across over 270 Slack channels. The accuracy of its root-cause suggestions jumped from 60% to over 90% after they integrated it with historical data. That's a big leap. And it's not just about getting the right answer—it's about getting it fast. Engineers can go from "what's happening?" to "here's what we should try" in a fraction of the time.

The system also boasts a 99.9% workflow success rate, which means it almost never fails to complete its diagnostic routine. It's made over 58,000 MCP tool calls, and it adapts to about 60 different team workflows. Those are the kinds of numbers that make you think this isn't a toy.

A Partner, Not a Replacement

Here's the thing that sets Blueberry apart: it doesn't act on its own. It collects info, forms hypotheses, and helps with debugging—but it never makes changes to the production environment. The final call, whether to roll back a deploy or tweak a config, always rests with a human.

That's a deliberate choice. AI in ops is scary because it could make things worse. So Instacart keeps Blueberry in an advisory role. It's like having a smart colleague who whispers suggestions while you're the one holding the wrench. And that's probably why engineers actually use it—it's not trying to take over their job, just to make it less awful.

Learning from Every Incident

Blueberry doesn't just use the past; it learns from each new incident. When an engineer works through a problem, the system retains that knowledge and applies it to future responses. Over time, this creates a feedback loop where Blueberry becomes more attuned to Instacart's quirks and patterns. It's like an apprentice who gets better with every job.

But is it a "force multiplier"? That's what Instacart's VP of Engineering, Siby Alappatt, calls it. And honestly, that's a bit of corporate-speak. What it really means is that on-call shifts feel different. Instead of starting from zero, engineers begin with a rich set of context: logs, deployments, related systems. That shift in starting point can make a huge difference in morale and efficiency. I've been on call, and I know the dread of staring at a blank dashboard. Anything that cuts through that fog is welcome.

What Others Can Learn

Instacart's experience with Blueberry offers a lesson for anyone building AI tools for ops. The model alone isn't enough. The real magic is in the integration—how the AI connects to actual data sources, how it fits into workflows, how it learns from feedback. A generic AI might give you a textbook answer, but one that's wired into your environment can give you an answer that actually works.

And context is king. Blueberry's accuracy jumped because it could tap into 14 years of incident history. That's the kind of institutional knowledge that usually lives in people's heads, not in a database. By making it accessible, Blueberry is unlocking a superpower.

Of course, there are limits. Blueberry is only as good as the data it's trained on, and it can't predict every weird edge case. The team is still working on improving its accuracy and expanding its capabilities. But so far, it's a solid example of practical, grounded AI—not just hype.

As more companies look to adopt AI in operations, Blueberry is a case study in doing it right. It's not trying to replace human judgment; it's trying to enhance it. And in the high-stakes world of production incidents, that's a valuable addition to any toolkit. Even if the name sounds like a dessert.