AI Governance for Institutional Readiness in Finance
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "AI Governance for Institutional Readiness in Finance".
Warren: Here's the gist of the paper. Agentic AI is gaining traction in asset management, yet operational governance has lagged significantly; survey data reveals that 88% of finance professionals lack an operational governance framework for agentic AI, and only a small fraction of large U.S.
Ruth: money managers disclose formal policies in Form ADV filings.
Warren: The authors posit that this gap is fundamentally architectural: governance structures designed for static validation are inadequate for continuously retrained agentic policies.
Ruth: To address this, the paper proposes a four-layer framework: Policy, Engineering, Composition, and Systemic.
Paper summary: Ruth: So, we're talking about this paper, "AI Governance for Institutional Readiness in Finance," and the main thing is that agentic AI is moving into asset management, but the rules for how to manage it are completely out of step.
Warren: Right. The authors point out that eighty-eight percent of finance professionals don't have an operational governance framework for agentic AI, and only a small fraction of those large U <ref:2608.02311#pg1,operational governance framework for agentic AI, and only>.S. money managers disclose formal policies in their Form ADV filings, according to the paper "AI Governance for Institutional Readiness in Finance" <ref:2608.02311#pg1>.
Iris: And the central argument they make is that this gap isn't just a compliance issue; it’s actually an architectural one. They suggest that governance structures designed for static validation simply don't work when you have agentic policies that are continuously being retrained, which is what the paper says on page two of "AI Governance for Institutional Readiness in Finance" <ref:2608.02311#pg2> <ref:2608.02311#pg2>.
Ruth: So they propose a four-layer framework to fix this. It’s Policy, Engineering, Composition, and Systemic. The idea is to build these controls into the structure itself rather than trying to bolt them on later.
Warren: Let's unpack that first layer: the Policy Layer where they talk about what you intend through a reward function. The paper says this reward function needs the same level of review and version control as traditional risk policies, which is really important for things like fraud agents <ref:2608.02311#pg1>.
Iris: And when they look at that policy layer, it involves scrutinizing cost matrices, specifically looking at the false positive and false negative rates for decision-making agents <ref:2608.02311#pg1>.
Ruth: Exactly. Then you have the engineering layer where they focus on enforcement, which means monitoring policy stability using a specific regret formula to track drift in real time <ref:2608.02311#pg2>.
Warren: That regret formula they mention is Regret(T) = X T t=one Cov(ct, pi hat t(ct)) + X T t=one c t bar b t bar which basically measures how the policy is drifting over time <ref:2608.02311#pg2> <ref:2608.02311#pg2>.
Paper summary: Iris: And on top of that, they suggest kill-switch architectures based on inner confidence signals instead of just declared confidence when it comes to fraud agents <ref:2608.02311#pg2>.
Ruth: That leads into the interface and composition validation layer, where they address risk across different components. This layer includes requirements like pipeline level adversarial red teaming and interface contracts with explicit exclusion lists <ref:2608.02311#pg1>.
Warren: They also talk about joint output sampling scaled by an Agentic Risk Score, which is a measure of how risky the combined outputs are <ref:2608.02311#pg1>.
Iris: And for foundation models specifically, they require vendor model version attestation to make sure you know exactly what model version you're running <ref:2608.02311#pg1>.
Ruth: Finally, the systemic layer tackles crowding risk by simulating what they call a "Calibrating the Calm-versus Stress Crowding Simulation" <ref:2608.02311#pg1>.
Warren: And this systemic part demands disclosure of shared data exposure, including who those fraud consortiums or shared fraud scoring vendors are that agents are relying on <ref:2608.02311#pg1>.
Iris: The paper backs all this up with synthetic illustrations, like a regret-covariance drift monitor and crowding simulations showing joint drawdown risk escalating from thirty-nine point two percent to seventy-nine point three percent <ref:2608.02311#pg1> <ref:2608.02311#pg1,from 39.2% to 79.3>.
Ruth: And they use real case studies too, including an LLM-embedding trading strategy and an airline chatbot ruling to show that these failure modes are actually happening in the real world <ref:2608.02311#pg1>.
Warren: The core message they're pushing is that the structural advantage of AI-native systems isn't inherent; it’s something you have to earn by applying these architectural controls <ref:2608.02311#pg1>.
Iris: They conclude that this gap is really about an architectural literacy problem, meaning governing bodies just don't have the technical knowledge to specify governance for these complex systems <ref:2608.02311#pg3>.
Ruth: So what does this mean for us? It suggests that institutions adopting this capability now won't need to retrofit it later under regulatory pressure because they’ll already have these controls in place <ref:2608.02311#pg3>.
Warren: The paper, "AI Governance for Institutional Readiness in Finance," essentially argues that you can't just use old risk management rules for AI agents; you need a whole new, four-layer system built specifically to handle continuous retraining and emergent risks <ref:2608.02311#pg1>.
Paper summary: Iris: It’s about moving away from static checks to something dynamic, like the regret-covariance monitor they describe which tracks policy drift in real time <ref:2608.02311#pg2>.
Ruth: And the implications are huge because it shifts the focus from just compliance checklists to a deep technical understanding of how agentic systems actually fail under stress <ref:2608.02311#pg1>.
Warren: It moves the conversation from simple pass or fail to continuous monitoring and active control mechanisms embedded in every layer of the AI system <ref:2608.02311#pg1>.
Iris: The paper highlights that existing frameworks simply aren't sufficient for these new types of policies, especially when they deal with adversarial input distributions <ref:2608.02311#pg2>.
Ruth: So for anyone in finance looking at this, it really means you need to start thinking about governance not just as a paperwork exercise but as an engineering challenge from the very beginning <ref:2608.02311#pg1>.
Warren: It’s about making sure the reward functions are robust and that you have those kill-switches ready when things start behaving unexpectedly <ref:2608.02311#pg2>.
Iris: And we need to be careful because they also point out a limitation: the data and code used for their worked numerical examples, like the regret-covariance monitor, were generated synthetically by the code itself <ref:2608.02311#pg3>.
Ruth: That's a fair point—it means we have to trust that their synthetic examples accurately reflect real risks, which is a big assumption in this whole paper <ref:2608.02311#pg3>.
Warren: But they also note that the informal survey data they used was just an aggregate summary from a LinkedIn poll and not individual responses <ref:2608.02311#pg3>.
Iris: So the paper's strength lies in demonstrating computability through those synthetic examples and real case studies, showing what failure modes actually look like <ref:2608.02311#pg1>.
Ruth: So to wrap up, "AI Governance for Institutional Readiness in Finance" is a call for a fundamentally different kind of governance structure that understands how continuously retrained agentic policies behave <ref:2608.02311#pg1>.
Warren: It’s about moving from static validation to real-time drift monitoring and composition risk management across four distinct layers <ref:2608.02311#pg1>.
Iris: And it forces us to confront the literacy gap, showing that technical knowledge is what's missing from current institutional governance structures <ref:2608.02311#pg3>.
Conclusion: Ruth: So, to recap, this paper is looking at how we actually govern AI in finance when those agents are constantly learning and changing their behavior.
Warren: Yeah, they’re talking about this framework called Policy, Engineering, Composition, and Systemic layers that you have to build into the system itself.
Iris: It’s about moving away from just checking boxes for a static system to something that monitors policy drift in real time using a specific regret formula.
Ruth: The authors are arguing that this isn't just some compliance headache; it’s an architectural literacy gap, meaning the people setting the rules don't have the technical knowledge to specify governance for these kinds of AI systems.
Warren: They show how old frameworks fail across four failure modes: policy drift, input distribution drift, composition risk between different agents, and cross-institutional policy correlation.
Iris: It’s not a simple pass or fail test anymore; it’s about continuous monitoring and active control mechanisms embedded in every layer of the AI.
Ruth: The title itself, "AI Governance for Institutional Readiness in Finance," really captures that shift from thinking about compliance paperwork to thinking about engineering what the system actually does.
Warren: And who wrote this? It was written by a team that combines economics, quantitative trading, and econometrics—which is interesting because they’re coming at this from three very different angles.
Iris: I think the combination of an economist looking at incentives, a trader looking at costs and drawdowns, and an econometrician checking identification makes this analysis really robust.
Ruth: Exactly. It’s not just one person telling you what to do; it’s understanding how the people in the system—the reward function designers, the traders using the AI—actually change their minds under pressure.
Warren: And that brings up a big question for us: if these institutions start adopting this four-layer structure now, what does that mean for how quickly they can actually deploy and scale these advanced AI strategies?
RiskAICenter · AbleAlpha
econ.EM, q-fin.RM, q-fin.ST
Submitted: 2026-08-03
Updated: 2026-11-14
Comments: 50 pages
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Agentic AI is gaining traction in asset management, yet operational governance has lagged significantly…
Key concepts
- Policy Layer
- This layer defines institutional intent through the agent's reward function. It requires rigorous version control and review, treating policy changes like traditional risk policies. For fraud agents, this involves analyzing cost matrices for false positives and performing disparate-impact analyses.
- Engineering Layer
- This focuses on enforcement mechanisms. It mandates monitoring policy stability using a specific regret formula to detect drift. It also requires kill-switch architectures based on inner confidence signals and limits customer-facing agents' unilateral actions via Bounded Action Authority controls.
- Interface and Composition Validation
- This addresses risk arising from combining different AI components. It involves adversarial red-teaming pipelines, using interface contracts with exclusion lists, scaling joint outputs by an Agentic Risk Score (ARS), and verifying foundation model versions from vendors.
- Systemic Layer
- This layer manages crowding risk across institutions. It uses simulations to test stress scenarios and demands disclosure of shared data exposure, including the identities of any fraud consortiums or scoring vendors agents depend on.
Terminology
Summary
Agentic AI is gaining traction in asset management, yet operational governance has lagged significantly; survey data reveals that 88% of finance professionals lack an operational governance framework for agentic AI, and only a small fraction of large U. p. 0 AI Governance for Institutional Readiness in Finance
Irene Aldridge∗and Steve Krawciw†
RiskAICenter
Abstract
Agentic AI is gaining acceptance in asset management, but governance has not kept pace:
88% of surveyed finance professionals report no operational governance framework for agentic
AI, and only 24 of 75 large U. money managers disclosing AI use in Form ADV filings report a formal governance policy. money managers disclosing AI use in Form ADV filings report a formal governance policy. We argue this gap is architectural: governance built for static
validation does not survive continuously retrained agentic policies. money managers disclosing AI use in Form ADV filings report a formal governance policy. We argue this gap is architectural: governance built for static
validation does not survive continuously retrained agentic policies. The
methodology has no analog in traditional model risk management, since a linear regression or a
decision tree has no equivalent ”declared confidence” to distrust. 8 Positioning of This Paper
Table 2 summarizes the coverage gap that this paper’s four-layer framework is designed to close:
each existing literature or framework addresses one or two of the four governance dimensions this paper identifies (policy-level objective review, engineering-level real-time monitoring, cross-component
composition risk, and cross-institutional systemic risk), but none addresses all four jointly, and none
provides a computable, finance-specific instantiation of each layer.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Policy Layer dictates institutional intent via the reward function, which must undergo the same rigorous review and version control as traditional risk policies. For fraud agents, this involves scrutinizing cost matrices (false-positive/false-negative) and performing disparate-impact analyses. p. 5 [2016] and Skalse et al. That literature formalizes when a proxy
objective diverges from a designer’s true intent, but neither paper addresses the specific legal and
empirical form that divergence takes in consumer lending and, by direct extension, fraud decisionmaking: disparate impact on a protected class. 3 Layer 2 Extended: Detecting Fraud-Model Policy Drift via Regret-Covariance 19
5. 3 Layer 2, Extended: Bounded Action Authority for Customer-Facing Agents .Where the paper says it
Where the paper says it
Layer 2.5: Interface and Composition Validation addresses compositional risk through requirements such as pipeline-level adversarial red-teaming, interface contracts with explicit exclusion lists, joint-output sampling scaled by the Agentic Risk Score (ARS), and vendor model-version attestation for foundation models. Finally, the Systemic Layer tackles crowding risk via a Calibrating the Calm-versus Stress Crowding Simulation
and demands disclosure of shared data exposure, including the identity of fraud consortiums or shared fraud-scoring vendors upon which agents rely.
The paper substantiates these points with synthetic illustrations—including a regret-covariance drift monitor and crowding simulations showing joint drawdown risk escalating from 39. p. 0 We argue this gap is architectural: governance built for static
validation does not survive continuously retrained agentic policies. We propose a four-layer
framework (Policy, Engineering, Composition, Systemic) grounded in two distinct kinds of evidence, kept explicitly separate: two calibrated synthetic illustrations (a regret-covariance drift
monitor; a crowding simulation showing joint drawdown risk rising from 39. 2 The counter-case: this paper’s own evidence that ”AI-native” is not sufficient 35
5. 3 The synthesis: the advantage is real but earned, not structural . 3%), and
three real, documented cases (a deployed LLM-embedding trading strategy, a $45 billion discretionary fund’s forced-deleveraging blowup, and a tribunal ruling holding an airline liable for its
chatbot). The synthetic examples demonstrate computability from observable data; the cases
demonstrate that the failure modes are not hypothetical. 2 Payments Track . 39
7 Discussion: The Literacy Problem Underneath the Governance Gap 41
8 Limitations 41
9 Conclusion 41
10 Appendix: Robustness of the Regret-Covariance Monitor 43
Data availability statement
The data and code used to produce the worked numerical examples in Sections 5. 5 Policy Drift Under Adversarial Input Distributions . 6 Compositional Risk: Emergent Failure Across Validated Components . [2016] define frame reward hacking as an agent
finding an unintended way to maximize a proxy objective that diverges from the designer’s true
intent. Reward hacking is one of a small set of concrete, near-term AI safety problems requiring
engineering solutions rather than philosophical ones.Where the paper says it
Where the paper says it
Where the paper says it
earned, not structural.
The framework is presented as a novel unification of reward-function review, real-time drift monitoring, cross-vendor composition risk, and cross-institutional crowding disclosure. Crucially, the authors argue that this governance gap is an architectural literacy gap—the governing bodies lack the technical literacy to specify governance for these systems—rather than a mere compliance failure. p. 2Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Policy Layer Review: Implement a formal review process for agentic strategy reward functions with
the same review, approval, and version control as any other risk policy document.
This ensures that the objective function driving an agent is scrutinized for its time horizon mismatch with vehicle liquidity and its guardrails around pursuing it. -
Engineering Layer Monitoring: Deploy a model-free policy stability monitor using regret-covariance decomposition to detect drift by computing Regret(T) from observed costs and decisions alone, triggering an alarm when the statistic rises sharply after a change point.
-
Engineering Layer Safety: Implement inner confidence kill-switches based on pre-decoding signals like entropy of logits or distance from the validated training distribution to halt actions before they are executed, as
declared confidence is biased by the decoding process.
-
Engineering Layer Bounded Action Authority: Establish a hard ceiling on unilateral action scope, defined by v(a) <= κ(s), ensuring that a customer-facing agent cannot execute actions whose consequence magnitude exceeds an independently reviewed and version-controlled threshold.
-
Composition Layer Validation: Apply pipeline-level adversarial red-teaming at the entry point and exit point, combined with vendor model version attestation for every production inference to trace performance discontinuities to specific vendor updates.
-
Systemic Layer Disclosure: Require firms to report the identity of any fraud consortium, shared device-fingerprint provider, or concentrated upstream fraud-scoring vendor on which their agents depend, along with the approximate share of decision-making weight attributable to that shared signal versus proprietary signals.
Abstract
Agentic AI is gaining acceptance in asset management, but governance has not kept pace: 88% of surveyed finance professionals report no operational governance framework for agentic AI, and only 24 of 75 large U.S. money managers disclosing AI use in Form ADV filings report a formal governance policy. We argue this gap is architectural: governance built for static validation does not survive continuously retrained agentic policies. We propose a four-layer framework (Policy, Engineering, Composition, Systemic) grounded in two distinct kinds of evidence, kept explicitly separate: two calibrated synthetic illustrations (a regret-covariance drift monitor; a crowding simulation showing joint drawdown risk rising from 39.2% to 79.3%), and three real, documented cases (a deployed LLM-embedding trading strategy, a 45 billion discretionary fund's forced-deleveraging blowup, and a tribunal ruling holding an airline liable for its chatbot). The synthetic examples demonstrate computability from observable data; the cases demonstrate that the failure modes are not hypothetical. We provide a 90-day implementation sequence spanning trading and payments/customer-facing systems.
Sources
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs