Skip to main content

AI that works for regulated financial institutions

Capital markets run some of the most advanced technology in the world, in the front office. Substantial investment goes into research tools, pre-trade analytics, scenario analysis and execution optimisation. However, once the trades are executed, they wait on emails, spreadsheets and portal logins to be booked, margined, reconciled and settled.

In a companion post, Himanshu wrote about how much the hedge fund COO’s job has grown: scale without headcount, key-person risk, controls they can evidence, and AI they’re now accountable for. I want to pick up that last point, because I think it’s particularly important given the current hype around AI. The critical question is how a COO puts AI into production in a manner that works within a sophisticated regulated environment.

Decades of investment, mostly at one end of the trade lifecycle

At Oliver Wyman I spent years advising banks, buy-side firms, and market infrastructure providers on where to invest. The answer was almost always the front office. Any edge on alpha generation would be immensely valuable, hit P&L in a palpable way and drive ROI.

The middle and back office got the remainder. The result, at most funds, is a stack where a single trade touches an OMS, several prime broker portals, an administrator, a collateral system and a dozen inboxes after having executed. None of these were designed to talk to each other, so operations teams need to do the translation across these systems. Every exception becomes a small investigation.

Why automation never reached the exceptions

The last generation of tools, RPA, workflow engines and rules, pushed STP rates up on the standardised cases. However, automation fell apart on the exceptions, and in operations, the exceptions are the job.

The reason is structural. The rules that decide how an exception gets resolved are written down nowhere. How a particular prime broker formats its FX file. Which counterparty disputes margin calls and which pays. What a dealer means by a particular phrasing. How the same break was settled last quarter. That tacit knowledge lives in years of email threads and in the heads of a few experienced people. A rulebook can only encode the breaks somebody has already seen, not conventions nobody has ever stated.

What COOs now ask about AI

AI has moved from an innovation project to something the COO co-owns. In SS&C’s 2026 survey of multi-manager funds, about three in five named AI as the trend with the biggest impact on the year, and nearly half already run it in production.

The questions COOs ask have changed with it. Less “what can the model do?” and more:

  • What exactly did it do, and on what evidence?
  • Who approved it?
  • What data can it see, and what can’t it?
  • Can I show this to my board, an allocator’s ODD team or a regulator?
  • What happens when it’s wrong?

Those are the right questions. Most AI in operations today can’t answer them. General-purpose assistants can summarise a break once manually provided with the information from the platform, but without the firm’s conventions, memory of past edge cases or an audit trail, the work still falls to a person. Moreover, few pilots and initiatives run inside the firm’s own environment, with governance that would get sign off from a CISO and compliance.

Five principles for AI a COO can defend

When I led ventures in market infrastructure, I learned that building the technology is only part of it. Getting it trusted inside a regulated institution is the harder part. These are the principles I think matter.

  1. Operate within the systems and permissions in the firm. No rip-and-replace. The agents connect to the OMS, platforms, databases, files and mailboxes the team already uses, run in the firm’s own environment with its chosen model provider, and see only what their role-based permissions allow.
  2. Provide evidence on every action. Every proposal comes with the evidence attached and its reasoning visible. The audit trail is a by-product of the work, down to every tool call, not a report written afterwards.
  3. Build in sign off by people, with autonomy earned progressively. On day one the agent prepares each action, a person approves it in one click, and code executes it. Autonomy is extended per workflow and per exception type, as the record justifies it, and the COO decides where the boundary between recommendation and action sits.
  4. Ensure SOPs and institutional memory are captured and owned by the firm. How the firm actually resolves issues, today held in a few heads and inboxes, is written down as a by-product of the work and becomes an artefact the team owns. When someone is on holiday or leaves, the knowledge stays. Approved changes are versioned, so you can see how the agent behaved on any given day and why a particular conclusion was reached. That directly reduces key-person risk.
  5. Build resilience to provide leverage when it is most needed. Agents work around the clock and absorb spikes in volume, so on the day volatility hits or three people are out, the team spends its time on judgement, not triage.

Repeatability and composability are a must

If every new process means a new build, with its own controls and its own sign-off, you have a consultancy, not a platform. In Yantra, we run every agent on our harness: Fulcrum. Memory, guardrails, observability and tool configurations are inherited, so each new workflow takes a fraction of the effort required to set up the first one. The effect compounds: every resolved exception teaches the agents something, and fewer items reach an operator at all.

Why now

The models caught up, and they keep improving every couple of months. Cost pressure on operations is real and lasting. The best ops people are scarce and mobile. Settlement cycles are compressing too, with the UK and EU moving to T+1 in October 2027.

That’s what we’re building at Yantra, starting with Arch. I’ll be writing here about where AI is actually changing capital markets infrastructure, and what it takes to deploy it inside regulated institutions. If you run operations at a fund and want to see what is the art of the possible, get in touch.

Further reading on the Yantra blog