A newly published study introduces the concept of “deferred trust” – a cognitive mechanism in which distrust of human decision-makers redirects reliance toward AI perceived as more neutral or competent. The effect is most pronounced, the researchers find, in financial decision-making contexts, where algorithmic framing reduces perceived personal accountability. That detail turns an interesting psychology finding into a governance problem: it describes a mechanism, grounded in principal-agent theory, by which decisions get made and nobody is left holding them. This article explains the study behind the term, the reason rising AI-trust scores are frequently a mismeasured signal under Goodhart’s Law, the three organizational settings where the damage concentrates, and what the finding demands of boards that are currently reading it as a technology adoption success.
A regional bank’s credit committee declines a small business loan. The applicant asks who made the decision. The loan officer points to a risk score. The risk officer points to the model. The model was built by a vendor nobody in the room has met. The decision was made by everyone, and so, in the way that matters for accountability, by no one. Economists have a name for that outcome: accountability diffusion. New research on what scientists call “deferred trust” suggests this scene is being repeated across organizations for a reason that has nothing to do with how good the algorithm is, and everything to do with how much credibility the people around it have already spent.
Why Employees Trust AI More Than Their Managers
THE STUDY THAT GAVE THE PATTERN A NAME
In November 2025, researchers Johan Sebastián Galíndez-Acosta and Juan José Giraldo-Huertas published a study that gave a precise name to a dynamic most executives have sensed but never measured. “Deferred trust,” as they define it, is a cognitive mechanism in which distrust of human agents redirects reliance toward AI perceived as more neutral or competent. It is not a preference for machines. It is a withdrawal from people, laundered through a technology choice.
The experiment was deliberately human-centered rather than corporate. Fifty-five participants worked through thirty decision-making scenarios spanning factual, emotional, and moral territory, choosing between AI agents, voice assistants, peers, adults, and religious authority figures as their guide. Adults were the single most-selected source overall, at 35.05%, with AI close behind at 28.29% and the remaining share spread across peers, priests, and voice assistants. The pattern that mattered was not the overall split but what predicted it: lower prior trust in human agents, peers, adults, authority figures, consistently predicted higher AI selection. The researchers describe this as a compensatory transfer. When confidence in the people around you erodes, the machine does not have to earn your trust. It only has to be perceived as less compromised than the alternative.
Swap “priests and peers” for “managers and HR” and the finding stops being an academic curiosity and becomes an organizational diagnostic.
Source: Galíndez-Acosta & Giraldo-Huertas, “Trust in Al Emerges from Distrust in Humans,” 2025.
A METRIC BOARDS ARE ALREADY MISREADING
Most AI adoption dashboards track a version of trust: usage rates, sentiment scores, willingness to act on AI recommendations. Rising numbers on that dashboard are almost universally read as a technology success story, proof the change management worked and the workforce has come around. The deferred trust finding says a share of that number is not endorsement of the technology at all. It is evidence that the alternative, a manager, a process, a colleague, was trusted less to begin with.
This is a textbook case of Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure. An AI-trust score was designed to track genuine confidence in the technology. Once leadership teams start optimizing for it, funding tools, running campaigns, and reporting it to the board as a KPI, the metric absorbs a second, uninvited signal: the erosion of trust in management that the deferred trust mechanism is quietly routing through it. A board reading the contaminated number as a pure technology signal is not making a small measurement error. It is missing the fact that a portion of its adoption success is actually attrition dressed up as enthusiasm.
The economic backdrop makes the misreading expensive rather than academic. Deloitte’s current enterprise data shows 74% of organizations want AI to grow revenue, yet only 20% have seen it happen and just 25% of AI initiatives deliver their expected return, even as 86% plan to increase AI budgets again this year. An organization funding more AI on the strength of a trust metric that is partly measuring management credibility is optimizing the wrong variable, at scale, on purpose.
THE MECHANISM: ALGORITHM APPRECIATION MEETS THE FLUENCY EFFECT
The deferred trust finding sits inside a broader and well-established body of research on what behavioral scientists call algorithm appreciation — the documented tendency of people to weight algorithmic judgment more heavily than human judgment, particularly when the human source is one they have reason to doubt. What the new study adds is the causal direction: it is not simply that AI looks credible. It is that a specific deficit in human trust is what activates the transfer in the first place. Remove the deficit and the “AI preference” weakens. The machine was never the primary variable.
This matters because of what the researchers flag as the real risk in the finding. Deferred trust does not eliminate the vulnerabilities that caused people to distrust their human sources: political incentives, inconsistency, self-interest. It relocates them behind a fluency effect: AI output arrives polished, confident, and apparently neutral, which lets it bypass the epistemic scrutiny a skeptical colleague would still apply to a skeptical manager. The employee who no longer trusts their manager’s judgment does not become more careful. They become less careful, in a new direction, because the new source looks like it has no incentive to be wrong.
A workforce is not becoming better at judgment because it is leaning on AI. It may be becoming worse at judgment while feeling more confident about it, the exact inversion an organization needs least while it is also trying to prove AI’s return on investment.
Source: nStratagem analysis, built on Galíndez-Acosta & Giraldo-Huertas (2025) and principal-agent theory.
THREE ORGANIZATIONAL SIGNATURES OF DEFERRED TRUST
The Adjudicated Layoff. Orgvue’s current research finds that 39% of business leaders made roles redundant citing AI deployment, and 55% admit the resulting redundancy decisions were flawed, with roughly a third of the managers who cut an AI-justified role already rehiring for it. Employees watching this cycle repeat do not conclude that AI is unreliable. They conclude that the human calling the decision was unreliable, and AI was the label applied after the fact. The next performance rating or restructuring recommendation carrying an “AI-informed” tag will be trusted disproportionately, not because the model improved, but because the manager’s track record did not survive the last round.
The HR Function Bypass. Deferred trust predicts that employees route sensitive questions, grievances, ethical concerns, career-risk conversations, toward AI copilots rather than the human channels ostensibly built for them, precisely because AI is perceived as less political and less able to hold a grudge. This is a governance exposure with no line item: it produces shadow use of AI on HR-adjacent decisions, undocumented and unsupervised, driven by a confidentiality and fairness deficit the organization has not acknowledged, let alone fixed.
The Silent Board. The dynamic replicates at the top of the house. BCG’s research finds nearly three-quarters of CEOs consider themselves their organization’s chief AI decision-maker, yet only 28% take direct responsibility for AI governance oversight and just 17% say their board does. KPMG and INSEAD’s new AI Board Governance Principles were built against survey data showing nearly three-quarters of boards rate their own AI expertise as moderate or limited. A director who does not trust their own AI judgment, or a peer’s, has every incentive to defer contested calls to an algorithmic risk score instead of building the judgment the seat requires. At board level, deferred trust does not look like a compensatory mechanism. It looks like governance abdication wearing a data-driven label.
Three Organizational Signatures of Deferred Trust
THE FINANCIAL DECISION LAUNDERING EFFECT
The detail in the research that deserves the most board-level attention is the one least discussed: the deferred trust effect is strongest in financial decision-making, specifically because algorithmic framing reduces how personally accountable the human making the call feels for the outcome. This is not a coincidence of the study design. It is a direct hit on principal-agent theory, the framework economists use to describe what happens when a principal, a shareholder, a customer, an employee, depends on an agent, a manager, an underwriter, a lender, to act faithfully on their behalf. The entire discipline of governance, audit, and executive compensation exists to manage the gap that opens when an agent’s incentives diverge from the principal’s interest.
Algorithmic framing does not close that gap. It severs the line the gap is measured against. When a lending decision, a claims denial, or a compensation call is described as “what the model recommended,” the human agent has interposed a third party into a two-party accountability relationship, one that cannot be sanctioned, promoted, or held to a fiduciary standard. The principal still bears the consequence of the decision. The agent has functionally exited the relationship while remaining in the room. This is why the mechanism is sharpest in finance: lending, underwriting, and pricing are precisely the decisions where principal-agent theory is best developed and most tightly monitored, which means overlaying algorithmic framing onto them does not add a new risk so much as it disables the oldest and best-understood safeguard against an existing one.
The organizational health warning the source research points to is specific: a lending team, a claims unit, or a pricing desk that shows rising AI-trust scores alongside declining willingness among staff to put their name behind a contested call is not becoming more efficient. It is quietly redistributing accountability away from every person the principal would otherwise be able to hold responsible.
None of this requires bad faith on anyone’s part. A loan officer who defers to a risk score is not trying to escape accountability any more than the manager who lets an “AI-recommended” rating stand in for a difficult conversation is trying to dodge one. Deferred trust is compelling precisely because it does not feel like abdication from the inside. It feels like prudence: the model has more data, less ego, and no stake in the outcome. That is exactly the framing that makes principal-agent erosion invisible to the people living through it, and exactly why it has to be caught structurally rather than left to individual conscience.
THE ECONOMIC COST OF MISREADING THE SIGNAL
The financial exposure here is not hypothetical. Trust erosion inside an organization carries a documented cost independent of AI: research from BetterUp finds a 10% decline in trust can reduce financial performance by as much as $115 million over four years for a $500 million company. Deferred trust is a symptom of exactly that kind of erosion, arriving dressed as a technology metric that makes it harder for a board to see, and therefore harder to price.
Layer in the current AI capital cycle and the exposure compounds. Hyperscalers alone are pushing 2026 AI infrastructure spending toward $725 billion, and markets have already shown they will punish companies that cannot pair the spend with visible growth. An organization that reads its rising AI-trust numbers as adoption success, funds the next round of tooling accordingly, and never addresses the interpersonal trust deficit underneath it is compounding two failures at once: a capital allocation problem and a decision-quality problem, both hidden behind the same misread metric.
There is a talent dimension to this cost that rarely reaches the board pack. Employees who have quietly concluded that the algorithm is more trustworthy than their manager have not resolved their trust problem. They have parked it. The research literature on psychological safety consistently shows that unresolved distrust does not sit still: it either surfaces as attrition among the people with the most external options, or it calcifies into the kind of disengagement that shows up two years later as a retention crisis with no obvious trigger. Deferred trust is a leading indicator sitting in plain sight, months or years ahead of the exit interviews that will eventually confirm what the AI-usage data already implied.
WHAT THE DATA ACTUALLY DEMANDS
None of this argues for slowing AI adoption. It argues against treating a rising AI-trust score as self-explanatory. A board that separates the two variables, genuine technology confidence and relocated management distrust, is a board that can price its AI investment correctly and see a workforce credibility problem before it shows up in attrition data. A board that pairs its AI-trust metric with an independently tracked management-trust metric turns Goodhart’s Law back in its own favor: two measures are harder to game than one, and the gap between them becomes the signal that matters.
The same logic applies to the accountability question raised by the financial decision-making finding. Naming the human who owns a lending, underwriting, or performance call, and requiring that ownership to survive contact with an algorithmic recommendation rather than dissolve into it, is not a compliance add-on. It is the restoration of the principal-agent link that algorithmic framing is currently permitted to sever by default. Organizations that treat this as a governance principle, not a technology feature request, will be the ones still able to answer the applicant in that credit committee when the question comes back around: who made this decision, and why should I trust them.
Deferred trust is not a compliment to your AI stack. It is a receipt for what your leadership has already spent, and most boards have not yet been handed the bill.
© 2026 Norm Murray. All Rights Reserved. No part of this publication may be reproduced, distributed, or transmitted in any form without the prior written permission of the author.
