How Instacart's Blueberry AI Could Reshape Property Management Incident Response

When a server goes down at 2 a.m., the on-call engineer doesn't get to sleep in. They wake up, open Slack, and start hunting for context: Which service is this? What changed in the last deploy? Are the logs screaming anything useful? That scramble for background info can eat up more time than the actual fix.

Instacart just announced an internal AI system called Blueberry that's designed to shorten that scramble. It lives inside Slack, spins up a bunch of sub-agents the moment an alert fires, and comes back with a root-cause hypothesis in about three minutes. In April alone, it ran roughly 25,000 diagnostics across 270+ Slack channels.

Here's what's interesting for people in real estate: the same operational pain exists in property management, especially as buildings get smarter and more connected. A failing HVAC sensor, a security camera going dark, or a water leak detector that won't stop chirping—each one forces a facilities manager to dig through vendor dashboards, maintenance logs, and past work orders before they can even start troubleshooting.

Blueberry's approach offers a template for how real estate teams could handle those alerts without losing their minds.

The core problem: context is scattered

Instacart's engineers were drowning in context gathering. When something broke, they had to figure out which team owned the service, review recent deploys, comb through logs and metrics, search internal docs, and then compare the current symptoms against past incidents. That's a lot of clicking before you even form a hypothesis.

Property managers face a similar mess. Let's say a smart thermostat in a retail space reports a temperature spike. To diagnose it, you might need to check the tenant's lease for after-hours HVAC rules, pull up the building management system (BMS) logs, look at recent maintenance tickets, and see if the weather is just being dramatic. That's a lot of tabs.

Blueberry's whole reason for existing is to collect that context automatically and hand it to the engineer in the same thread where they're already talking to their team. No jumping between tools. No re-typing the same question into five different dashboards.

How Blueberry works: parallel agents, persistent memory, tool access

When an alert triggers, Blueberry launches around ten sub-agents in parallel. Each one is wired to a different internal system—one checks deployment history, another looks at service ownership, a third digs into logs, and so on. They all report back into the same Slack thread where the incident is being handled.

The system doesn't just guess. It's grounded in 14 years of incident history, which is apparently what pushed its accuracy from 60% to over 90%. It also uses the Model Context Protocol (MCP) to connect to internal tools, and it keeps a persistent state so it can remember what's already been tried.

For real estate, the analog would be an AI that ties into your property management software, your IoT sensor network, your work order history, and maybe even your lease agreements. When a sensor goes haywire, the AI could pull up relevant past incidents (like the time the same sensor caused a false alarm), check if any recent maintenance was done, and even note whether the tenant has a history of complaining about temperature.

Humans stay in the loop

One of the more reassuring details: Blueberry doesn't make changes to the production environment. It's strictly an advisory system. It gathers info, generates hypotheses, and helps with debugging. A human engineer still decides what to actually do.

That's a smart line to hold. In real estate, you don't want an AI that automatically shuts off a building's water supply because a leak sensor went off. You want it to tell you the sensor is located in the third-floor janitor's closet, the last leak there was due to a broken pipe in 2021, and here's the contact info for the plumber who fixed it. Then a human decides whether to send someone out.

What Blueberry's numbers actually show

Instacart is pretty open about the results. In April, Blueberry handled about 25,000 diagnostics. The workflow success rate hit 99.9%. MCP tool calls exceeded 58,000. And it's adaptable enough to fit the working patterns of roughly 60 different teams.

Those numbers matter because they suggest the system isn't just a demo. It's handling real incidents, at scale, and it's being used across the company. The fact that it's integrated into Slack means it's not adding another dashboard to check—it's living where the work already happens.

For a property management company with hundreds of buildings, something similar could handle thousands of sensor alerts a month. Instead of a facilities manager waking up to 15 emails about a single false alarm, they'd get one Slack message with a clear summary and a few likely causes.

What real estate teams can borrow from this playbook

You don't need to build a custom AI to steal the ideas. The key principles are simple:

• Connect your data sources. The AI is only as good as what it can see. If your incident history is scattered across email, spreadsheets, and a half-used CMMS, start there.
• Build a memory of past incidents. Blueberry's accuracy jumped because it had 14 years of history to reference. Even five years of well-tagged work orders would give you a solid baseline.
• Meet people where they are. If your team lives in Slack, put the AI there. If they use Teams or email, adapt accordingly. The point is to reduce friction, not add another tool.
• Keep humans in charge. Use AI to summarize, correlate, and suggest. Don't let it hit the emergency stop button on your building systems without a human in the loop.

Why this matters for property management

The real estate industry is sitting on a lot of data—sensor readings, maintenance logs, tenant requests, energy usage—but most of it is siloed and underused. The problem isn't that we don't have information; it's that we can't find it fast enough when something breaks.

Blueberry is a reminder that the value of AI isn't just in predicting when something will fail. It's in helping you respond faster when it does. For a facilities team managing a portfolio of smart buildings, that could mean the difference between a quick fix and a tenant complaint that escalates.

The takeaway: context beats raw intelligence

Instacart's engineers didn't need a smarter model. They needed a system that could bring the right context to the right conversation at the right time. The AI's job was to connect dots across tools and history, not to be a genius.

Real estate teams dealing with increasingly complex building systems could use the same logic. The next time a rooftop unit starts acting up, wouldn't it be nice to have an assistant that says, "This unit had a similar fault last July, and the fix was replacing the capacitor. Here's the vendor's contact info and the last service date."

That's not science fiction. It's just a matter of connecting the dots—and letting an AI do the boring part.