Morgan Stanley has deployed agentic AI in one of banking's most accuracy-critical, deadline-driven workflows -- profit and loss (P&L) reconciliation -- and cut the associated work in half, according to VentureBeat reporting published June 30, 2026. The counterintuitive part, per Morgan Stanley Managing Director Todd Johnson: the bank got there by making the system less autonomous, not more, keeping humans tightly in the loop while their decisions get iteratively turned into repeatable rules the system can apply on its own.
The internal system, known as FIXR, operates against a specific, high-stakes daily process: every trading day, Morgan Stanley's desks handle transactions across cash equities and debt investments, and at day's end, controllers must reconcile P&L across the firm's Finance, Risk, Operations and Trade Capture systems -- a process where hundreds of thousands of attributes frequently fail to match. Historically, this required controllers to manually investigate each mismatch (a "break"), decide on an adjustment, and sign off, all against a hard morning deadline, taking up to six hours for a single book. FIXR now performs the same task in two to three hours, and across roughly 100 controllers doing this work globally, that translates to approximately 1,500 hours saved per week.
FIXR's architecture uses several specialized agents working together: one interprets past guidance to develop start-of-day resolutions, one learns from controller behavior and documents the rules they apply, and one converts repeated patterns into durable, automated logic. After nightly P&L calculations complete, the system automatically analyzes breaks and proposes resolutions based on learned rules, auto-clearing certain familiar breaks, suggesting solutions for less-familiar ones, and flagging genuinely uncertain cases for human investigation. "It's much more like a co-worker than a copilot," Johnson said.
“"It's much more like a co-worker than a copilot," Johnson said.”
Critically, humans never leave the loop -- they review, approve or correct every recommendation, and those decisions feed back into improving the next run. "We recognized that all that intelligence that's sitting in the mind of a controller is going to be difficult to get all into an agent on day one," Johnson said, describing a deliberate, iterative approach to codifying institutional knowledge rather than assuming an agent could absorb it immediately. "You still preserve that element of human accountability even as you start to automate," he added, emphasizing that autonomy requires earned trust, and that efficiency gains disappear if every agent action still needs to be checked by a human anyway.
Johnson's team also deliberately limited how much of the workflow depended on the model's judgment at all, favoring prescribed, repeatable logic wherever possible: "If you have an 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 have the LLM do the stuff where you don't need that kind of deterministic workflow," he said. As the system observes more controller feedback on a given type of break, Morgan Stanley converts that pattern into a fixed rule rather than continuing to leave it to the model's judgment -- a deliberate design choice to minimize unnecessary model reliance rather than maximize it.
The broader context reinforces why this design philosophy matters: VentureBeat's own VB Pulse survey of 87 enterprises found nearly three-quarters reporting little to no ROI from custom model fine-tuning, describing a "sandbox graveyard" of AI projects too costly to maintain, while 38% cited the lack of a single accountable owner as their biggest barrier to production AI, and only two of the 87 surveyed enterprises had active monitoring and alerting in place to detect model failures. Morgan Stanley's process-first, human-accountability-centered approach directly addresses the governance gap that survey identifies as the industry's most common failure point.
For founders and enterprise operators building agentic AI, Morgan Stanley's experience is a genuinely useful counter to the instinct to maximize agent autonomy as quickly as possible: deliberately constraining a system to deterministic, prescribed logic wherever the workflow allows, while reserving genuine model judgment only for cases that need it, appears to be both cheaper and more auditable than a fully autonomous approach. For investors and executives evaluating enterprise AI deployments, a specific, verifiable claim -- 1,500 hours saved weekly across 100 controllers, in one of the highest-stakes workflows in banking -- is a far more useful data point than the vague productivity claims common across the industry.
What to watch: whether Morgan Stanley extends FIXR's process-first, human-in-the-loop design to other accuracy-critical workflows across the firm, whether other financial institutions adopt similarly constrained-autonomy approaches rather than chasing maximal agent independence, and whether the broader industry's governance gap (highlighted in VentureBeat's own survey data) narrows as more enterprises study deployments like Morgan Stanley's.