Morgan Stanley has cut its riskiest reconciliation work in half – by making its agents less independent

Most enterprise AI deployments to date have focused on coding assistants and customer service bots. Morgan Stanley sent agents to one of the bank’s most important, time-bound operations, the profit and loss (P&L) reconciliation – and cut the job in half. The counter part: it got there by making the system less autonomous, not more.

Humans remain firmly in the loop, and their decisions are repeatedly turned into repeatable rules that the system can implement itself.

“It’s more like a co-worker than a pilot,” said Morgan Stanley Managing Director Todd Johnson at the recent VB AI Impact event. The internal production agent system, known as FIXR, goes beyond simple, straightforward "type AI 1.0" activities. “We think that’s where the opportunity opens up for the most difficult work in the organization.”

FIXR behind the scenes

Every trading day, Morgan Stanley’s trading desks handle significant activity regarding transactions such as equity or debt investments.

And, at the end of each of those days, regulators must put together a P&L for all financial systems, Risk, Operations, and Trade Capture of the financial giant. All that data has to match, and, perhaps not surprisingly, hundreds of thousands of factors often fail to match.

Typically, this means that administrators must manually investigate each discrepancy (or “break”), make decisions about corrections, and sign out before the number goes to the desk. And all this while working on a tough morning deadline.

Previously, this would take six hours for one book. Now, FIXR does the job in two to three hours, Johnson said. For every 100 controllers doing this job, that adds up to about 1,500 hours saved per week.

After the nightly P&L calculations are completed, the system automatically analyzes the “breaks” and suggests decisions based on the rules learned. Several agents work together:

• One interprets previous guidance to improve decisions at the beginning of the day.

• One learns from the controller’s behavior and writes down the rules it uses.

• A person transforms repetitive patterns into a strong, automatic mind.

Over time, the system can automatically clear certain breaks it has encountered before, suggest solutions for others that may be less familiar, ask for help when unsure, and flag for human investigation. When things are solved over and over again in the same way, it can create strong rules.

Sadly, people don’t leave the loop, but stay fully in it, he said. They review, approve or modify all recommendations, and provide those decisions to improve subsequent operations. The agent learns daily from controllers what it finds right and wrong and integrates that information as it iterates.

“You still retain that element of personal responsibility as you start to do it yourself,” Johnson said. Over time you’ll see a lot of those things get resolved automatically.”

He stressed that independence requires a lot of trust; businesses will not see efficiency gains if everyone checks everything the agent does.

The human agent feedback loop was critical to meeting the challenge of controllable, measurable, and repeatable automation. “We realized that all that intelligence sitting in the controller’s mind was going to be difficult to get all of it to the agent on day one,” Johnson said.

Focus on process-first, flexibility

It was important to establish processes first, before any AI came in, Johnson said. His team conducted a “very thorough” process intelligence evaluation process that designed and drilled workflows to identify where automation would be most beneficial: Is it an agent response, traditional automation, or a simple re-do of an inefficient step?

“If we can fix that first before we add agents to the problem, we will be really changing the opportunity,” he said.

The P&L signing process was full of manual steps worth doing, and agents taking on some of these time-consuming tasks freed up regulators for “value-added analysis” and “deeper risk assessment” work, he said.

Expansion, however, was just as important as saving time. Johnson’s team chose this use case for P&L reconciliation because hundreds of controllers are doing this work around the world across the business (America, Europe, Asia).

So start with a use case, prove it, scale it, “and eventually the transformation will be as we roll this out across the organization,” Johnson said.

Determining by design

Johnson said the team also deliberately limited how much workflow depends on the model’s judgment at all. "If you have the opportunity to do things more deterministically and iteratively, that’s cheaper in terms of token usage, it’s more reproducible in terms of control – and you have LLM do things 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 instead of leaving it in the model.

Humans still own morals

An interesting (and perhaps fundamental) question raised at the beginning of the agency era is: Are agents coded or digital workers?

Johnson asserts that they are “probably a bit of both,” and, as such, require flexibility when it comes to governance and direction. Technical teams should still be responsible for maintaining protections and mechanisms such as firewalls or encryption, for example.

But there is a new dynamic about the “performance factor”: People who use agents are responsible for them because we help their business work. For example, when a senior manager works with a junior manager, they don’t just abdicate responsibility because someone is helping them, Johnson notes.

“One of our strong principles in our governance of AI in general is that there should always be human accountability, even if there is some level of automation,” he said.

But often there is no “one person,” and the process is ultimately ongoing. So far, Johnson joked that one “stressful” thing about agent AI is that it will require constant training because the models are constantly changing.

“You can’t say: ‘We’ve done all the checks and balances we need to do. Let’s stop it.’ You have to have an ongoing vision as it changes over time.”

Morgan Stanley targets real business pain points

Morgan Stanley’s experience reflects the patterns VentureBeat has identified across enterprise AI deployments.

In a recent VB Pulse survey by VentureBeat, nearly three-quarters of respondents reported seeing little or no ROI from custom modeling, explaining "Graveyard sandbox" of AI projects proved too expensive to maintain. This suggests that Morgan Stanley’s start-up, buy-and-merge process may be more sustainable than chasing unprecedented models. The survey has 87 respondents and the findings should be considered as a guideline.

Governance emerged as another common challenge: 38% of respondents cited the lack of a single owner responsible for accountability as their biggest obstacle to AI production, while only two of the 87 businesses surveyed had effective monitoring and alerting to detect model failures.