Accounting Teams Can Learn from Instacart's AI Incident Response

When Your Ledger Goes Down, Who Do You Call?

Imagine it's 2 a.m. and your accounting system just flagged a reconciliation mismatch that could be a million-dollar error. The usual drill? You'd pull up service ownership docs, dig through transaction logs, compare metrics, and maybe scroll through old incident reports. That's hours of context-gathering before you even start diagnosing the actual problem.

Instacart faced a similar pain—but for production outages. Their engineers were wasting precious time collecting background info during incidents. So they built Blueberry, an AI assistant that does the heavy lifting of context collection and hypothesis generation. In April alone, Blueberry ran about 25,000 diagnostics across 270+ Slack channels. Diagnosis accuracy jumped from 60% to over 90% after hooking into 14 years of incident history.

Now, I'm not saying accounting is exactly like running a grocery delivery platform. But the underlying problem—slow, manual triage when something breaks—is painfully familiar. And the solution's architecture has some lessons for accounting teams drowning in data and deadlines.

The Context Problem in Accounting

When an auditor or a controller spots an anomaly, the first instinct is to gather context. Which entity does this belong to? Was there a recent deployment of a new ERP module? What do the logs say? Are there historical precedents? That context-gathering can eat up hours, especially in large organizations with fragmented systems.

Instacart's engineers faced the same issue. They had to determine service ownership, review deployments, analyze logs and metrics, search documentation, and compare symptoms to past incidents. It was a manual, time-consuming process that delayed actual troubleshooting.

For accounting, the stakes are high. A delayed response to a compliance issue or a misstated financial report can mean regulatory fines, audit failures, or lost investor confidence. The faster you can zero in on the root cause, the better.

How Blueberry Works: A Crash Course

Blueberry is built on a multi-agent AI system. When an alert triggers, about 10 sub-agents spin up in parallel. They pull information from internal resources—incident history, service ownership data, logs, deployments, and other debugging signals. Within about three minutes, they generate a root-cause hypothesis, right inside the Slack thread where the engineer is already working.

What's key is that Blueberry doesn't act autonomously. It collects info, generates hypotheses, and assists debugging, but the engineer stays in charge. No unauthorized changes to production. That's a governance model accounting teams can appreciate.

From Incident Response to Accounting Diagnostics

So how does this translate to accounting? Think of your own 'incidents': a failed audit reconciliation, a sudden spike in expense report anomalies, a vendor invoice that doesn't match a PO. The diagnostic process is similar—you need to find the root cause fast.

Here's a practical mapping: Blueberry's incident history becomes your archive of past accounting issues and resolutions. Service ownership data becomes a chart of which team or system owns each data source. Logs and metrics become your transaction logs and control monitoring outputs. And the Slack integration? That's your team's collaboration hub, whether it's Slack, Teams, or email.

Building Your Own 'Blueberry' for Accounting

You don't need to build a full-blown AI agent network to start. But you can adopt the principles:

• Centralize your knowledge: Create a searchable repository of past accounting incidents, audit findings, and resolutions. Include specifics like system versions, process owners, and timestamps.
• Define data ownership: Map out which system or team is authoritative for each data element. This cuts down the 'who do I ask?' time.
• Automate context collection: Use scripts or RPA to pull relevant logs, transaction lists, or control test results when an anomaly is flagged.
• Integrate with your workflow: If your team lives in Slack or Teams, build a bot that can fetch context on demand. Even a simple slash command that pulls recent transactions for an account helps.

These steps won't magically fix everything, but they'll cut down the time from anomaly detection to root-cause analysis.

AI Accuracy Depends on History

One of the most interesting tidbits from Instacart's experience is that accuracy improved dramatically once they fed Blueberry 14 years of incident history. That's a reminder that AI is only as good as the data you give it.

For accounting, that means your historical audit data, past adjustments, and even successful reconciliations are gold. If you haven't been documenting your troubleshooting steps, start now. Future you—or future AI—will thank you.

Instacart's CTO, Anirban Kundu, said Blueberry is part of a broader AI agent exploration. The system uses a persistent, tool-aware framework with an MCP catalog and a review-improvement loop. In one month, it handled ~25,000 diagnostics with a 99.9% workflow success rate and over 58,000 MCP tool calls, adapting to about 60 different team workflows.

Human-in-the-Loop Is Non-Negotiable

Here's a cautionary note: Blueberry never makes changes to production on its own. It suggests, it informs, but the engineer decides. That's a model worth copying in accounting, where the cost of a wrong automated action can be severe.

In accounting, you can't let an AI post journal entries or approve write-offs without human sign-off. But you can let it flag potential issues and provide evidence. The final call stays with a trained professional.

Instacart's VP of Engineering, Siby Alappatt, called Blueberry a 'force multiplier' that helps quickly troubleshoot and mitigate complex production issues. For accounting, a similar tool could be a force multiplier for audit readiness and compliance monitoring.

Start Small, Think Big

You don't have to wait for a vendor to sell you 'AI for accounting.' Start with a simple chatbot that queries your incident database. Or set up automated alerts that bundle context when a variance exceeds a threshold.

Alan Wong, Instacart's Director of Software Engineering, noted that Blueberry changes the starting pattern for on-call engineers—they no longer start from zero. Instead, they begin with collected context like logs, deployments, and related systems. That's exactly the shift accounting teams need: move from 'where do I even begin?' to 'here's what I know, let's dig in.'

So next time you're staring at a blinking cursor on a reconciliation issue, remember: you don't need to be Instacart to think like them. With a little foresight and some process engineering, you can build your own Blueberry—maybe just without the cute name.