Morgan Stanley’s FIXR cut P&L reconciliation time in half, by limiting autonomy

Morgan Stanley Managing Director Todd Johnson says its internal agentic system, FIXR, now completes P&L reconciliation in two to three hours instead of up to six hours for a single book. That is a “cut it in half” outcome, delivered in one of banking’s most accuracy-critical, deadline-driven workflows.

The counterintuitive lever is that FIXR gets results by being less autonomous, not more. Johnson describes a tight human-in-the-loop setup where controllers review, approve or correct every recommendation, then feed decisions back into the system so the next run improves. In Johnson’s framing, FIXR is “much more like a co-worker than a copilot.”

So what exactly is happening in the workflow? Every trading day, Morgan Stanley’s trade desks execute transactions such as cash equities or debt investments. After the trading day, controllers must reconcile P&L across multiple Finance, Risk, Operations, and Trade Capture systems. The process is messy by nature: hundreds of thousands of attributes frequently fail to match. Typically, that turns into manual “break” investigations, adjustments decisions, and sign-off before the number goes to the desk, all under a hard morning deadline.

Previously, Johnson says the work could take up to six hours for a single book. With FIXR, the task runs in two to three hours. Across roughly 100 controllers doing this work, Johnson estimates about 1,500 hours saved per week. That matters because reconciliation is not just busywork, it is a control point that links multiple systems into one defensible financial picture. Cutting cycle time without losing correctness is the entire game.

Technically, the system is agentic without being “guessy.” Johnson says after nightly P&L calculations complete, FIXR analyzes breaks and proposes resolutions based on learned rules. Several agents work together: one interprets past guidance to develop start-of-day resolutions; one learns from controller behavior and documents the rules controllers apply; one converts repeated patterns into durable automated logic. Over time, Johnson says the system can auto-clear certain breaks it has encountered, suggest solutions for less familiar ones, ask for help when it is unsure, and flag items for human investigation. When items are repeatedly resolved the same way, the system creates firm rules.

But the governance design is the real story. Humans do not leave the loop, even as automation increases. Controllers review, approve or correct recommendations, then feed those decisions back so the system iterates daily based on what it gets right and wrong. Johnson emphasized that “autonomy requires a great deal of trust,” and that efficiency gains will not appear if everyone is still checking everything an agent does. The system gets trust through a measured feedback loop rather than a leap to full automation.

Morgan Stanley also started by fixing the process before adding agents. Johnson says his team ran a “very thorough” process intelligence assessment, mapping and mining workflows to decide where automation would be advantageous. The assessment aimed to answer a practical question: was the answer agents, traditional automation, or re-engineering an inefficient step? The team’s takeaway: if they can “fix that first before we add agents,” then automation can transform the opportunity instead of just digitizing chaos. In this case, Johnson says the P&L sign-off process contained manual steps suitable for automation, freeing controllers for “more value-added analysis” and “deeper risk consideration.”

Extensibility drove the choice of use case. P&L reconciliation was not a niche workflow, but one hundreds of controllers were doing globally across the business in the Americas, Europe, and Asia. The plan, per Johnson, is start with one use case, prove it, extend it, and then roll transformation across the organization. That rollout logic matters in enterprise AI because the winning deployments are the ones that survive contact with new teams, new edge cases, and new oversight needs.

Johnson adds a design principle that reads like a direct response to how LLMs can go off the rails: deterministic by design. He says the team deliberately limited how much of the workflow depends on the model’s judgment. In his words, making tasks prescribed and repeatable is “cheaper in terms of token consumption,” more repeatable in terms of controls, and suitable for the LLM to handle only where deterministic workflow is not required. As the system sees more controller feedback on a given break type, Morgan Stanley converts patterns into fixed rules instead of leaving it to the model.

This approach also reframes the governance question behind the agentic era: are agents code or digital employees? Johnson argues they’re “probably a little bit of both,” which changes how oversight works. Technical teams still own protections and guardrails such as firewalls or encryption. But there is also a “performance element,” meaning humans using agents are responsible because the agents are aiding business work. Johnson notes that responsibility does not simply vanish when one controller has a system assist them, including scenarios where a senior controller works with a junior controller. His governance principle is clear: there always has to be human accountability, even with automation, and there typically is not just one single person. It is a continuous process.

There is also an operational reality for the boardroom: agentic AI will require ongoing training. Johnson jokes that agentic AI’s “depressing” part is you cannot declare victory after one round of evaluation and testing, because models change over time. So the control system must be continuous, not one-and-done.

If you zoom out beyond Morgan Stanley, the company’s playbook lines up with the broader enterprise AI pain that VentureBeat has reported. In a VB Pulse survey of 87 respondents, nearly three-quarters reported little to no ROI from custom model fine-tuning, describing a “sandbox graveyard” of AI projects that proved too costly to maintain. Governance also emerged as a recurring bottleneck: 38% of respondents cited lack of a single accountable owner as their biggest barrier to production AI, while only two of the 87 enterprises had active monitoring and alerting in place to detect model failures. FIXR’s approach, with human accountability built into the workflow and repeatable rule conversion, is positioned as a way to avoid chasing fragile custom models while still delivering measurable speed.

For executives and board members, the strategic stake is simple: the reconciliation deadline is not optional, and accuracy is not negotiable. FIXR suggests a path where automation is allowed to expand, but only through process intelligence, deterministic rule formation, and explicit human approval. In other words, the fastest way to scale AI in regulated, money-critical work may not be giving agents more freedom. It may be giving humans better structure, and giving the system only the kinds of decisions it can safely make and prove.