Art Amador, QuantumStreet AI
A chief investment officer told me last quarter that his committee had walked away from a promising AI-driven mandate, not because it underperformed. It backtested well and the returns were impressive. They passed because one member of the investment committee asked a straightforward question that nobody in the room could answer. “If this strategy draws down 15 percent in a quarter, what do we tell our beneficiaries about why it happened?”
That conversation has become increasingly common as allocators evaluate AI-driven quantitative investment strategies. I believe the technology is maturing, and allocators are becoming more comfortable using AI in both their professional and personal lives.
SimCorp’s 2026 InvestOps Report, which surveyed 200 executives at asset managers, pension funds, and insurers each overseeing at least $10 billion, found that AI has become a business-critical part of front-office work. A year earlier, only about one in 10 were actively using AI tools. Adoption is no longer the question.
What hasn’t kept pace is trust. Using AI inside an organization is very different from trusting it to make critical decisions.
PwC’s 2025 Global CEO Survey found that only 8 percent of financial services leaders expected AI to be significantly integrated into their core business strategy. Many firms are running the tools. Far fewer are willing to stand behind them when capital and fiduciary exposure are real. That gap is where many institutional AI investment programs stall, and the cause is rarely the quality of the model.
Many AI-driven strategies still do not give an allocator enough information to explain and defend an investment decision. For institutions, that creates a serious governance and fiduciary concern. The duty of care obliges a manager to understand the basis of a recommendation, test it, and monitor it as conditions change. Handing the decision to a machine does not lift that obligation.
Legal commentators have begun addressing the issue directly. Relying on an opaque model without independently validating its output or reasoning raises difficult fiduciary questions. It’s not fundamentally different from relying on an outside analyst whose work was never vetted. The more opaque the model, the harder its reasoning is to reconstruct when the decision is under scrutiny. The need for reconstruction is almost always demanded after performance has deteriorated, when an explanation matters most to investment committees and boards.
The issue isn’t whether AI can make investment decisions, but whether institutions can understand and defend them.
The CFA Institute’s 2025 report, “Explainable AI in Finance,” Identified another problem. Most of us have quietly noticed the pattern in our own working habits. We ask an AI tool for a view, and when it agrees with the position we already hold, we take the agreement as confirmation and move on. That’s what we call AI confirmation bias. On the flip side, when it flatters our thesis, we rarely interrogate it. The CFA Institute report’s warning lands on this reflex.
A clean explanation invites people to trust the answer, but a plausible account of why a model acted is not evidence that the model was correct. Explainability cuts both ways. It gives allocators something they can defend, while also requiring them to test whether the explanation is actually supported by evidence. Anyone or any platform offering certainty on this point is worth treating with suspicion.
Drug regulators settled a version of this question decades ago. A compound that performs well in trials is not approved on the strength of the outcome alone. Reviewers want the mechanism of action, the account of what the molecule does, and why the effect follows, because an effect nobody can explain is an effect nobody can predict the limits of. Investment allocation has not reached that standard. The models perform. Whether anyone can say why is still treated as optional, and that is what will separate the firms that can scale these strategies from the ones that walk away after the first shock.
This is the problem my firm works on, so I hold a clear view of it.
The approach we take at QuantumStreet AI is to make every forecast add up in a way a person can follow. The method, called SHAP, short for Shapley Additive Explanations, comes out of cooperative game theory, and the underlying idea is old and intuitive. If several players combine to produce a result, how much did each one contribute? Applied to a forecast, the model starts from a baseline, its average expected return, and every signal it weighs then moves the number up or down from there. Performance can be attributed to specific and explainable signals in granular detail.
Momentum might add half a point and a deteriorating sentiment reading might take a point away, or a valuation signal might add a quarter. Together, those contributions add up to the model’s forecast. A portfolio manager can look at any single position and see which factors drove it and by how much, in terms an investment committee can question directly. When those contributions shift as the market regime changes, the shift is visible too, which turns the explanation into an early warning that the model’s logic may be drifting. None of this makes the underlying machinery simple. The models are complex, and the honest description is that this attribution sits on top of that complexity rather than removing it. What it delivers is narrower and more useful to a fiduciary. You can see, and defend, why every position was taken.
A serious objection runs the other way. Many allocators will say they are judged on returns, not mechanisms, and that a manager who beats the benchmark across a decade is never asked to show the wiring. In a calm market, fair enough. But the demand for an explanation is counter-cyclical. It arrives precisely when a strategy is losing money, when the board is uneasy and beneficiaries are asking pointed questions, which is the worst possible moment to learn that the reasoning was never legible to begin with.
The question for allocators in 2026 and certainly going into 2027 is not whether to adopt AI. That decision is largely settled, as the survey data confirms. The question is sharper, and it will separate the institutions that compound an edge from those that stall. When your AI-driven strategy has its first difficult quarter, and it will have one, can you sit in front of your board and explain what the model saw and why it acted? If the honest answer is no, you have not acquired an investment strategy. You have taken on a liability that simply has not come due yet.
The author, Art Amador, is Co-founder and President of QuantumStreet AI
