Most enterprise AI deployments so far have focused on coding assistants and customer service bots. Morgan Stanley has deployed agents into one of banking’s most accuracy-critical, deadline-driven workflows – profit and loss (P&L) reconciliation – and cut the work in half. The counter-intuitive part: It got there by making the system more, not less, autonomous.
Humans live in a tight loop, and their decisions are turned into iteratively repeatable rules that the system itself can enforce.
“It’s more like a co-worker than a co-pilot,” Todd Johnson, managing director at Morgan Stanley, said at the recent VB AI Impact event. The Internal Production Agentic System, known as FIXR, is beyond simple, straightforward "General AI 1.0" Work. “We think this is where the opportunity is to really unlock more complex work across the organization.”
FIXR behind the scenes
Each trading day, Morgan Stanley’s trading desks handle important tasks related to transactions such as cash equity or debt investments.
And, at the end of each of those days, controllers must reconcile the P&L to the finance giant’s finance, risk, operations and trade capture systems. All that data must come together, and, perhaps not surprisingly, hundreds of thousands of attributes often fail to match.
Typically, this means that controllers must manually investigate each mismatch (or “break”), decide on an adjustment, then ideally sign off before going to the numbers desk. And all this while working to tight early morning deadlines.
Previously, one book could take up to six hours. Now, FIXR completes the task in two to three hours, Johnson said. Approximately 100 controllers performing this function save approximately 1,500 hours per week.
After the nightly P&L calculation is completed, the system automatically analyzes the “breaks” and proposes solutions based on the rules learned. Multiple agents work together:
- The person interprets previous guidance to develop resolutions to begin the day.
-
One learns from the controller’s behavior and documents the rules they enforce.
-
One converts repeated patterns into sustainable, automatic logic.
Over time, the system can automatically clean up some of the interference it encountered earlier, suggest solutions for others who may be less familiar, ask for help when unsure, and flag for human investigation. When objects are solved repeatedly through the same method, it can create rigid rules.
Crucially, he said, humans do not leave the loop, but remain in it entirely. They review, approve or correct each recommendation, then feed those decisions back to make improvements in the next round. The agent learns from controllers daily what it thinks is right and wrong and as it repeats the knowledge it codifies it.
“When you start to automate you still retain that element of human accountability,” Johnson said. “Over time you’ll see more and more of those items resolved in an automated way.”
He emphasized that autonomy requires a great deal of trust; Enterprises won’t gain efficiency if everyone checks everything an agent does.
The human-agent feedback loop was critical to addressing the challenge of controlled, measured, and repeatable automation. “We recognized that whatever intelligence was in a controller’s mind would be difficult to convey to an agent on day one,” Johnson said.
Process-first, focus on extensibility
Johnson said it was important to establish processes before incorporating any AI. His team ran a “very thorough” process intelligence assessment that mapped and mined workflows to identify where automation would be most beneficial: was the answer agents, traditional automation, or simple re-engineering of an inefficient step?
“If we can fix this before we add agents to the problem, we’ll really change the opportunity,” he said.
The P&L sign-off process was filled with manual steps amenable to automation, he said, and agents handling some of these time-consuming tasks are freeing up controllers for “more value-added analysis” and “deeper risk consideration” work.
However, expandability was just as important as time savings. Johnson’s team chose this particular P&L solution use case because hundreds of controllers were doing this work globally (in the US, Europe, Asia).
So start with the use case, prove it, expand it, “and then ultimately the change will happen as we implement it more and more throughout the organization,” Johnson said.
deterministic by design
Johnson said the team also intentionally limited how much of the workflow depends on the model’s decisions. "If you have the opportunity to make things very deterministic and repeatable, it’s cheaper in terms of token consumption, it’s more repeatable in terms of control – and LLM works where you don’t need that kind of deterministic workflow," He said.
As the system sees more controller feedback on a given break type, Morgan Stanley converts that pattern into a fixed rule rather than leaving it to the model.
Man is still the master of behavior
An interesting (and perhaps fundamental) question is being raised at the dawn of the agentic era: are agents code or digital workers?
Johnson argues that “they’re probably a little bit of both,” and, as such, nuance is needed when it comes to governance and oversight. For example, technical teams must still be responsible for maintaining security and guardrails, such as firewalls or encryption.
But there is a new dynamic around the “performance element”: the humans using the agents are responsible for them because it is assisting their business function. For example, if a senior controller is working with a junior controller, they don’t give up responsibility just because someone is helping them, Johnson said.
“One of our strong principles in our AI governance generally is that there should always be human accountability, regardless of the degree of automation,” he said.
But there is usually never “the same person”, and the process is ultimately ongoing. At this point, Johnson joked that one “frustrating” thing about agentic AI is that it will require constant training because the models are always changing.
“You’ll never be able to say: ‘We’ve done all the evaluation and testing we need to do. Let’s give it a go.’ As it evolves over time, you need to take a consistent approach.
Morgan Stanley takes aim at real enterprise pain points
Morgan Stanley’s experience reflects the patterns highlighted by VentureBeat in enterprise AI deployments.
In VentureBeat’s recent VB Pulse survey, nearly three-quarters of respondents reported seeing little or no ROI from custom model fine-tuning, describing a "sandbox cemetery" AI projects that proved too expensive to maintain. This suggests that Morgan Stanley’s process-first, buy-and-mix approach may be more sustainable than pursuing a bespoke model. There were 87 respondents to the survey and the findings should be considered directional.
Governance emerged as another common challenge: 38% of respondents cited the lack of an accountable owner as the biggest barrier to production AI, while only two of the 87 enterprises surveyed had active monitoring and alerting in place to detect model failures.
<a href