To date, most enterprise AI deployments have focused on coding assistants and customer service bots. Morgan Stanley put agents in one of banking’s most time-critical, deadline-driven workflows—the profit and loss (P&L) reconciliation—and cut the work in half. The counter-intuitive part: it got the system by making it less autonomous, not more.
Humans remain firmly in the loop and their decisions become repetitive rules that the system can enforce on its own.
“It’s more like a co-worker than a pilot,” Morgan Stanley managing director Todd Johnson said at the VB AI Impact event. The in-house production agent system, known as FIXR, goes beyond simple "AI 1.0 type" assignments. “We think that’s where the opportunity really lies to unlock more complex work in the organization.”
FIXR behind the scenes
Each trading day, Morgan Stanley’s trading desks handle important business around transactions such as cash equity or debt investments.
And at the end of each of those days, controllers must reconcile the P&L between the finance giant’s Finance, Risk, Operations and Trade-Capture systems. All of this information needs to come together, and perhaps not surprisingly, the hundreds of thousands of attributes often don’t match.
Typically, this means that supervisors must manually investigate (or “crack”) each discrepancy, decide on corrections, and then ideally sign off on the number before it goes to the table. And all this while working on a difficult morning.
It used to take up to six hours for one book. Now, FIXR completes the task in two to three hours, Johnson said. Between about 100 supervisors doing this job, that saves about 1,500 hours a week.
After the nightly P&L calculations are completed, the system automatically analyzes the “breaks” and suggests decisions based on learned rules. Several agents work together:
-
One interprets past instructions to make decisions about the start of the day.
-
The person learns from the controller’s behavior and documents the rules it applies.
-
One transforms repetitive patterns into continuous, automated logic.
Over time, the system can automatically remove certain interruptions it has encountered before, suggest solutions for others that may be less familiar, ask for help when unsure, and flag them for human investigation. When items are solved the same way over and over again, it can create solid rules.
Critically, people don’t leave the loop, but stay completely in it, he said. They review, confirm, or correct each recommendation, then feed back those decisions to improve the next run. The agent learns what is right and what is wrong from the controllers on a daily basis and encodes this knowledge as a repeater.
“Even when you start automating, you still retain that element of human responsibility,” Johnson said. “Over time, you’ll see more of these elements being automatically resolved.”
He emphasized that autonomy requires great trust; If everyone is checking everything an agent does, businesses won’t see an increase in efficiency.
A human-agent feedback loop was critical to solving the automation challenge in a controlled, measurable, and repeatable manner. “We realized that it would be difficult to get all the intelligence that a supervisor has in mind into an agent on day one,” Johnson said.
First of all, focus on the process, the expansion
Before committing to any AI, it’s critical to establish processes first, Johnson said. His team conducted a “very comprehensive” process intelligence assessment that mapped and mined workflows to determine if automation would be most cost-effective: Was the response agents, traditional automation, or a simple redesign of an inefficient step?
“If we can fix this before we add agents to the problem, then we will have really changed the opportunity,” he said.
The P&L underwriting process was full of manual steps suitable for automation, and having agents take over some of those time-consuming tasks frees up supervisors for “more value-added analysis” and “deeper risk consideration,” he said.
Expansion was as important as saving time. Johnson’s team chose this particular P&L reconciliation use case because hundreds of controllers were doing it globally across businesses (Americas, Europe, Asia).
So start with one use case, prove it, scale it up, “and then as we roll it out more and more across the organization, it’s going to be transformational,” Johnson said.
Deterministic by design
Johnson said the team deliberately limited how much of the workflow depends on the model’s decision at all. "If you have the ability to make things very deterministic and repeatable, it’s cheaper in terms of token consumption, more repeatable in terms of controls – and make LLM do things you don’t need that kind of deterministic workflow," he said.
As the system sees more supervisor feedback on a given break type, Morgan Stanley turns that pattern into a fixed rule rather than leaving it to the model.
People still have manners
An interesting (and perhaps key) question raised at the beginning of the agency era is: are agents code or digital workers?
“They’re probably a little bit of both,” Johnson argues, and thus nuance is required when it comes to management and control. Technical teams must still be responsible for maintaining protections and safeguards, such as firewalls or encryption.
But there’s a new dynamic around the “performance element”: people who use agents are responsible for them because it helps their business. For example, if a senior supervisor works with a junior supervisor, they don’t simply abdicate responsibility because someone is helping them, Johnson noted.
“One of our strong principles in AI management is that there is always human responsibility, even if there is a degree of automation in general,” he said.
But there is usually no “one person” and the process is ultimately continuous. To this point, Johnson joked that one “depressing” thing about agent AI is that it requires constant training because the models are constantly changing.
“You’ll never be able to say, ‘We’ve done all the evaluation and testing we need to do.’ Let’s release it.” You’ll have to have a constant vision as you evolve over time.”
Morgan Stanley targets real enterprise pain points
Morgan Stanley’s experience mirrors the patterns VentureBeat has uncovered for AI applications in enterprises.
In VentureBeat’s recent VB Pulse survey, nearly three-quarters of respondents reported seeing little or no ROI from custom model fine-tuning. "sandbox cemetery" AI projects that are too expensive to maintain. This suggests that Morgan Stanley’s process-first, buy-and-mix approach may be more sustainable than pursuing bespoke models. 87 respondents took part in the survey and the results should be considered indicative.
Governance emerged as another common challenge: 38% of respondents cited the lack of a single responsible owner as the biggest barrier to AI production, while only two of the 87 enterprises surveyed have active monitoring and alerts to detect model failures.
