Investment bank Morgan Stanley has achieved a remarkable feat in one of banking's most accuracy-critical and deadline-driven workflows: profit and loss (P&L) reconciliation. Instead of pursuing full automation, the institution took a counterintuitive approach, reducing the autonomy of its artificial intelligence agents to cut work time in half. The system, called FIXR, was presented by Managing Director Todd Johnson at a recent VB AI Impact event.
FIXR slashes reconciliation time from six hours to two
Every trading day, after markets close, Morgan Stanley's controllers must reconcile P&L data across multiple internal systems, including Finance, Risk, Operations, and Trade Capture. Hundreds of thousands of attributes often fail to match, creating "breaks" that require manual investigation. Previously, a single book could take up to six hours. Now, FIXR completes the task in two to three hours. With roughly 100 controllers involved, this amounts to about 1,500 hours saved per week.
The key to success lies in keeping humans firmly in the loop. Johnson explained that the system acts more like a coworker than a copilot: controllers review, approve, or correct every recommendation, and their decisions are iteratively turned into deterministic rules the agent can apply autonomously in subsequent runs. This continuous feedback loop allows FIXR to learn day by day, codifying controllers' expertise.
Sponsored Protocol
Less autonomy for greater reliability and token savings
Morgan Stanley's approach is deliberately deterministic. Johnson emphasized that whenever possible, making processes prescriptive and repeatable is preferable, both to reduce token consumption and to ensure tighter control. The large language model (LLM) is used only where flexibility is needed. Over time, recurring patterns are converted into fixed rules, reducing reliance on the generative component.
This strategy aligns with findings from a recent VentureBeat survey, where nearly three-quarters of companies reported little to no return on investment from custom AI models, describing a "sandbox graveyard" of abandoned projects. Governance emerged as another key challenge: 38% of respondents cited the lack of a single accountable owner as the biggest barrier to production AI, and only two out of 87 enterprises had active monitoring to detect model failures.
Sponsored Protocol
Human accountability remains central in agentic AI governance
Johnson also addressed the question of whether AI agents are code or digital employees, concluding they are a bit of both. Consequently, governance must account for this duality. Technical teams must ensure firewalls, encryption, and other protections, but the end user remains responsible for the agent's actions, just as a senior controller cannot delegate responsibility to a junior. "One of our strong principles is that there always has to be human accountability, even with automation," he said.
One aspect Johnson jokingly called "depressing" is that agentic AI requires continuous training, as models constantly evolve. There is no finish line where one can declare the work done. Perpetual monitoring and rule updates are necessary as the system adapts.
Sponsored Protocol
Morgan Stanley started with a concrete use case, P&L reconciliation, and plans to extend the solution to other areas of the organization. The process-first approach, combined with buying existing technologies and blending solutions, appears more sustainable than chasing custom models. With FIXR, the bank has demonstrated that reducing agent autonomy can lead to better outcomes and greater user acceptance.
For more on AI governance, see the article on Anthropic bending security to please Trump. Also, for AI prompting techniques, check AI Image Prompting. Any AI solution must be integrated with solid governance, as Morgan Stanley's experience shows.
External source: Wikipedia - Morgan Stanley.
