Morgan Stanley cut its riskiest reconciliation job in half — by making its agents less autonomous

Morgan Stanley deployed an AI agent system called FIXR to handle profit and loss reconciliation, one of banking's most error-sensitive workflows. The system cut the work required in half. The approach reversed conventional thinking about AI autonomy.

Instead of maximizing agent independence, Morgan Stanley kept humans tightly embedded in the process. Decisions made by humans become encoded into repeatable rules the system applies independently going forward. This human-in-the-loop architecture proved more effective than fully autonomous systems.

Managing Director Todd Johnson described the system as "much more like a co-worker than a copilot." The distinction matters. Copilots assist humans who retain full control. Co-workers share responsibility and decision-making authority based on learned patterns.

P&L reconciliation presents a demanding test case for enterprise AI. Banks must match revenue and expense data across multiple systems daily. Errors cascade through financial reporting. Deadlines are fixed. Regulatory scrutiny is intense.

Most enterprise AI deployments focus on lower-stakes applications like coding assistants and customer service chatbots. Morgan Stanley targeted a high-consequence workflow where accuracy directly impacts financial statements and compliance. The choice highlighted a critical insight: maximum autonomy does not always produce maximum results.

The iterative human-feedback loop proved central to FIXR's success. Rather than building rules upfront, the system learns from human decisions on edge cases and exceptions. Those learned patterns gradually expand what the system handles independently. Humans remain available for novel or ambiguous situations.

This model addresses a persistent gap in AI deployment. Fully autonomous systems fail on exceptions. Rule-based systems require extensive upfront work. FIXR bridges the gap by combining human judgment with machine learning. The system grows smarter as humans interact with it.

The 50 percent reduction in required work suggests the approach scales beyond reconciliation. Finance teams spend significant time on routine approval workflows, compliance