Morgan Stanley Slashed a Six-Hour Finance Job to Three Hours by Keeping Humans in Control of AI

Quick Facts

• Morgan Stanley’s FIXR agent system cut P&L reconciliation from up to six hours to two or three hours per book, saving roughly 1,500 hours per week across 100 controllers.
• The system deliberately limits AI autonomy, keeping humans in the loop to review and approve decisions before converting repeated patterns into fixed automated rules.
• Morgan Stanley ran a process intelligence assessment before deploying any AI, mapping workflows to determine where agents, traditional automation, or simple re-engineering made the most sense.

Morgan Stanley has cut one of banking’s most time-sensitive back-office jobs in half — by making its AI less autonomous, not more. The firm’s internal agentic system, called FIXR, handles profit and loss reconciliation across Finance, Risk, Operations, and Trade Capture systems. What once consumed up to six hours per book now takes two to three hours.

Across roughly 100 controllers working globally in the Americas, Europe, and Asia, that adds up to about 1,500 hours saved each week. The use case is among the most demanding in financial services: every trading day, hundreds of thousands of data attributes must match across systems before numbers go to the desk.

How FIXR Works

After nightly P&L calculations finish, FIXR automatically analyzes mismatches, known as breaks, and proposes resolutions. Several agents operate in sequence. One interprets past guidance to build start-of-day resolutions. A second learns from controller behavior and documents the rules they apply. A third converts repeated patterns into durable automated logic.

Over time, the system can auto-clear breaks it has seen before, suggest solutions for less familiar ones, ask for help when uncertain, and flag items for human review. When the same resolution appears repeatedly, FIXR codifies it as a firm rule rather than leaving it to the model each time.

‘It’s much more like a co-worker than a copilot,’ said Todd Johnson, a managing director at Morgan Stanley, speaking at a recent VentureBeat AI Impact event.

The Case Against Full Autonomy

Johnson’s team built FIXR on a deliberate constraint: humans stay in the loop on every recommendation. Controllers review, approve, or correct each proposed resolution, and those decisions feed back into the system to sharpen the next run. The agent learns daily from what it gets right and wrong.

‘One of our strong principles in our AI governance generally is that there always has to be human accountability, even if there’s a degree of automation,’ Johnson said.

He also warned that autonomy requires trust earned over time. Enterprises will not see efficiency gains if staff check everything an agent does. The goal is to expand automated clearance gradually, as confidence in each rule type grows.

Process Before Technology

Before deploying any AI, Johnson’s team conducted what he described as a thorough process intelligence assessment. The team mapped and mined workflows to determine whether the right solution was agents, traditional automation, or simply fixing an inefficient manual step. ‘If we can fix that first before we add agents to the problem, then we really will be transforming the opportunity,’ Johnson said.

The P&L sign-off workflow contained numerous manual steps well-suited for automation. Freeing controllers from those tasks gives them more time for risk analysis and value-added work.

Governance as Infrastructure

Morgan Stanley is building governance infrastructure to match the deployment. Johnson framed the question bluntly: are agents code or digital employees? His answer is that they are probably a bit of both. Technical teams remain responsible for guardrails like firewalls and encryption. But controllers using the agents are accountable for their outputs, much as a senior controller remains responsible for the work of a junior colleague.

Johnson also acknowledged that ongoing model evaluation is unavoidable. As underlying models change, firms cannot simply declare testing complete. ‘You’re going to have to have a constant view as it evolves over time,’ he said.

The results arrive as the broader industry struggles to move AI out of the sandbox. VentureBeat’s VB Pulse survey found that nearly three-quarters of enterprise respondents reported little to no ROI from custom model fine-tuning, and only two of 87 enterprises surveyed had active monitoring in place to detect model failures. Morgan Stanley’s constrained, rules-based approach offers a counterpoint to the prevailing push for maximum AI autonomy.

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