AI in asset management: Designing the architecture for a Human-in-the-Loop

As agentic AI takes a greater role in the investment lifecycle, no firm can place a person between every model output and every action. A monitoring agent ingesting borrower information across hundreds of positions cannot wait for sign-off on each data point it processes. But it would be a serious misreading to conclude that human oversight will soon be obsolete. The lesson is the opposite: oversight can no longer be an afterthought bolted onto workflows. It must be designed - deliberately - into the architecture itself.

The current landscape

Adoption is no longer hypothetical. Mercer's 2026 survey of 131 asset managers found that 55% have integrated AI into at least one investment process and 91% plan to expand their use within twelve months. The Cambridge Centre for Alternative Finance reports that 52% of finance firms are already deploying agentic systems.

Yet look at where firms draw the line. Grant Thornton's 2026 survey of 950 business leaders found that only 5% permit AI agents to execute high-stakes decisions without human review. In Cambridge's global study, traditional financial institutions ranked loss of human oversight among their top concerns at 60%, well above the 42% of regulators who said the same. Those closest to the technology are unequivocal — high-stakes lending decisions must be made by human professionals.

Regulation on the horizon?

When the Bank of England's Deputy Governor Sarah Breeden addressed the ECB Forum in Sintra last month, she stated the Bank's position on AI very clearly: "Our frameworks were not built to contemplate autonomous agents and relying on a human in the loop for all agent actions is unlikely to be realistic."

We view Breeden’s comments as an objection to blanket oversight, a person rubber-stamping every action a system takes. Three requirements follow, each telling firms how to build.

A hard stop. She has floated market-wide circuit breakers and kill switches for faulty trading models, a way to halt a system that does not rely on that system agreeing it should stop.

A named individual. Regulators must be able to identify a specific person "accountable for that model". Not a committee in the abstract, a person.

A defence against herding. If many firms' systems react to the same trigger in the same way, they amplify a shock rather than absorb it. The corrective is judgement that varies, applied case by case where it matters most.

The firms that will find agentic regulation painful are those treating the human in the loop as a compliance burden - a person bolted on to sign things after the fact. The firms that treat these three requirements as engineering requirements have already built the answer.

Designing the loop

Oversight at Liquidity is built into the structure. Every decision moves through four stages.

The AI sets out the options. Our systems run across origination, underwriting, monitoring and collections, processing more company data than an analyst team could cover, but crucially, the AI does not make decisions autonomously.

The limits that must never be crossed are enforced deterministically, in code. For example, we do not set fixed rules on how much we can lend to any one borrower No eloquent model output can argue past them. A person signs off. Committing capital, setting terms and resolving a flagged loan go to an investment committee, where a named professional takes responsibility and applies context the model lacks.

Crucially, that sign-off is written back into the system as a record, so for every consequential decision there is always a named human who owns it and an audit trail that proves it. One record holds the reasoning. Each decision and its rationale go to a single shared record, so a credit professional can see what drove a forecast and challenge it.

Ready by design

Regulation for agentic AI is coming. Firms that bolt a human onto the end of an automated pipeline will meet it as a burden - and will discover, under scrutiny, that they cannot point to the circuit breaker, the accountable person, or the explanation.

Firms that build oversight into architecture are better positioned for the investments that follow. The same architecture that satisfies a regulator is the one that protects returns.