Most enterprise AI implementations so far have focused on coding assistants and customer service robots. Morgan Stanley has deployed agents to one of the most time-critical and accuracy-critical banking workflows – profit and loss reconciliation – and has cut the work in half. The counterintuitive part: It got there by making the system less autonomous, not more.
Humans are kept perfectly informed and their decisions are iteratively converted into repeatable rules that the system can apply on its own.
“He’s much more of a co-worker than a co-pilot,” Morgan Stanley CEO Todd Johnson said at a recent VB AI Impact event. The in-house agent system, known as FIXR, goes beyond simple and straightforward. "type of AI 1.0" tasks. “We think that’s where the opportunity really is to unlock more complex work in the organization.”
FIXR behind the scenes
Each trading day, Morgan Stanley’s trading desks handle the important work involved in transactions such as cash stocks or debt investments.
And, at the end of each of those days, controllers must reconcile the profits and losses of the financial giant’s finance, risk, operations and business capture systems. All that data needs to be tied together and, perhaps not surprisingly, hundreds of thousands of attributes often don’t match.
Typically, this means drivers must manually investigate each discrepancy (or “break”), make adjustment decisions, and ideally log off before the number hits the desktop. And all this while working on a tough morning deadline.
Previously, this could take up to six hours for a single book. Now, FIXR performs the task in two to three hours, Johnson said. Between the approximately 100 controllers doing this work, that adds up to approximately 1,500 hours saved per week.
Once the nightly profit and loss calculations are complete, the system automatically analyzes “breaks” and proposes resolutions based on the learned rules. Several agents work together:
-
One interprets past guidance to develop resolutions at the beginning of the day.
-
You learn from the behavior of the controller and the rules that apply are documented.
-
One converts repeated patterns into durable, automated logic.
Over time, the system can automatically clear certain faults it encountered before, suggest solutions for others that may be less familiar, ask for help when unsure, and flag for human investigation. When problems are solved repeatedly using the same method, firm rules can be created.
Fundamentally, humans do not leave the circle, but rather remain completely within it, he said. They review, approve or correct each recommendation and then provide feedback on those decisions to improve the next execution. The agent learns daily from controllers what it does well and what it does poorly and encodes that knowledge as it iterates.
“You still have that element of human responsibility even when you start to automate,” Johnson said. “Over time, you’ll see more and more of those elements resolved automatically.”
He highlighted that autonomy requires a lot of trust; Companies will not see efficiency gains if everyone controls everything an agent does.
The feedback loop between humans and agents was critical to addressing the challenge of controlled, measured, and repeatable automation. “We recognized that all that intelligence that’s in a controller’s mind is going to be difficult to incorporate into an agent from day one,” Johnson said.
Focus on process first, extensibility
Johnson said it was critical to establish processes first, before involving AI. His team conducted a “very thorough” process intelligence assessment that mapped and explored workflows to identify where automation would be most advantageous: were response agents, traditional automation, or simple reengineering an inefficient step?
“If we can fix that first before adding agents to the problem, then we will really be transforming the opportunity,” he said.
The P&L approval process was full of manual steps suitable for automation, and agents taking on some of these time-consuming tasks are freeing up controllers for “higher value-added analysis” and “deeper risk consideration” work, he said.
However, extensibility was as important as saving time. Johnson’s team chose this particular P&L reconciliation use case because hundreds of controllers were doing this work globally across the company (in the Americas, Europe, and Asia).
So start with a use case, test it, extend it, “and eventually the transformation will happen as we roll this out more and more across the organization,” Johnson said.
Deterministic by design
Johnson said the team also deliberately limited how much the workflow depended on the model’s judgment. "If you have the opportunity to make things very prescribed and repeatable, that’s cheaper in terms of token consumption, it’s more repeatable in terms of controls, and having the LLM do things where you don’t need that kind of deterministic workflow." said.
As the system sees more feedback from the controller about a given breakout type, Morgan Stanley turns that pattern into a fixed rule instead of leaving it to the model.
Humans still own behavior.
An interesting (and perhaps fundamental) question posed at the dawn of the agential era is: are agents coded or digital employees?
Johnson argues that they are “probably a bit of both” and, as such, require nuance when it comes to governance and oversight. Technical teams should still be responsible for maintaining security protections and barriers, such as firewalls or encryption, for example.
But there is a new dynamic around the “performance element”: the humans who use agents are responsible for them because they help in their business work. For example, if a senior controller is working with a junior controller, they don’t abdicate responsibility simply because someone is helping them, Johnson said.
“One of our strong principles in our overall AI governance is that there always has to be human accountability, even if there is a degree of automation,” he said.
But there is usually no “one person” and ultimately the process is continuous. To this point, Johnson joked that one “depressing” thing about agent AI is that it will require continuous training because the models are constantly changing.
“You’ll never be able to say, ‘We’ve done all the evaluations and tests we needed to do. Let’s just let it go.’ You’ll have to keep a constant eye on it as it evolves over time.”
Morgan Stanley targets real business pain points
Morgan Stanley’s experience reflects patterns VentureBeat has discovered in enterprise AI implementations.
In VentureBeat’s recent VB Pulse survey, nearly three-quarters of respondents reported seeing little to no return on investment in custom model fine-tuning, describing a "sandbox cemetery" of AI projects that were too expensive to maintain. This suggests that Morgan Stanley’s process-first, buy-and-match approach may be more sustainable than pursuing bespoke models. The survey had 87 respondents and the findings should be considered directional.
Governance emerged as another common challenge: 38% of respondents cited the lack of a single responsible owner as their biggest barrier to AI production, while only two of the 87 companies surveyed had active monitoring and alerting to detect model failures.
