Morgan Stanley Cuts Accounting Work in Half with AI

Morgan Stanley has implemented an AI agent system for P&L reconciliation work, one of the most accuracy- and speed-critical tasks in banking operations. As a result, work that previously took up to 6 hours per ledger has been reduced to 2-3 hours. Todd Johnson, the Managing Director who spearheaded this initiative, revealed this at the VB AI Impact event held recently.

P&L reconciliation is the process of comparing data across multiple systems—finance, risk, operations, and transaction records—at the end of each trading day and correcting discrepancies. Because each transaction records a vast number of attributes, hundreds of thousands of data mismatches occur at the close of each business day. Controllers responsible for this must investigate each one and make judgment calls on corrections before the next morning's deadline. The time pressure is extreme, and the environment is prone to human error.

The system deployed for this process is 'FIXR,' an internally developed agent system. Once nightly P&L calculations are complete, FIXR automatically analyzes data discrepancies and proposes solutions based on historical handling records. Multiple agents work in coordination internally, each taking on roles such as 'interpreting past instructions to create next-morning action plans,' 'learning judgment rules from controller behavior,' and 'converting repetitive patterns into automation logic.' For cases repeatedly resolved using the same method, the system generates fixed rules that enable automatic processing.

A distinctive feature is the intentional design to reduce autonomy. FIXR requires human approval for all proposals, and corrections and approvals made by controllers are reflected in the next cycle. Johnson states, 'We maintain the element of human responsibility while advancing automation.' He also described the system as 'more of a colleague than a copilot.'

In terms of actual impact, Johnson explained that approximately 1,500 hours per week in savings have been realized across the roughly 100 controllers involved in P&L reconciliation. This equates to about 15 hours of work reduction per person per week. Johnson also emphasized that 'building autonomy requires accumulation of trust,' indicating that the number of cases that can be automatically processed will gradually increase in the future.

While many companies are leveraging AI for coding assistance and customer support, Morgan Stanley chose core business operations where accuracy errors can translate directly into losses. This case supports the view that in AI agent deployment, 'how effectively one incorporates human judgment into the system' matters more than 'how autonomous to make the system.' Whether AI integration into core business processes accelerates across financial institutions, and how actual accuracy and audit compliance perform, will be key points of interest.