73% of Asset Managers Report AI Gains. Only 22% Could Survive an Audit.
The AI productivity story is no longer theoretical.
Asset managers are reporting measurable efficiency gains. Derivatives firms are quantifying time and cost savings. Boards and investment committees are approving larger AI budgets.
But a more important question is being neglected:
Can your firm reconstruct what the AI did when an examiner, allocator, auditor, or risk committee asks?
That is the control gap now separating AI pilots from production-grade operating capabilities.
The latest findings are direct. Grant Thornton’s Asset Management: 2026 AI Impact Survey reports that 73% of asset managers have achieved efficiency gains from AI. Only 14% report accelerated innovation. Just 22% are very confident they could pass an independent AI governance audit.
The conclusion is uncomfortable but clear:
Asset managers are scaling AI for efficiency, not growth, and many cannot prove how their AI works.
The gains are real. The evidence is not.
The Grant Thornton data should not be dismissed as failed AI adoption. Firms are saving time. They are reducing manual work. They are improving throughput.
But the same survey shows that only 25% report higher-quality outputs. Meanwhile, 53% identify governance and compliance barriers as contributors to AI underperformance, and 31% cite poor data quality and integration as a top barrier to scaling.
This is not a technology problem alone. It is an operating model problem.
“We saved 400 analyst hours” is useful information for an operating review. It does not answer the questions a supervisor or allocator will ask:
Which data did the system use?
Was that data complete and current at the time?
Which model or agent version produced the recommendation?
What prompt, workflow state, or tool calls were active?
Which policy rules were applied?
Was human approval required?
Who approved or rejected the action?
What did the system actually execute?
What exceptions occurred?
What was the financial or operational outcome?
An efficiency metric describes the result. It does not establish control.
That distinction matters more as firms respond to real operating pressure. Tariff and trade-policy changes can alter portfolio assumptions and client communications within hours. Higher rates and uncertain funding costs change the economics of financing, liquidity, and execution. Labor scarcity increases the need for agentic automation, not as a novelty, but as a way to expand capacity without expanding headcount at the same rate.
In that environment, productivity is the product. But productivity without evidence is an operational liability.
The market is adopting governance unevenly
Acuiti’s derivatives markets survey, published September 24, 2026, shows the same split from another angle. As reported by The Full FX, 44% of firms report significant, quantifiable time and cost savings from AI.
Governance approaches, however, diverge sharply by firm type:
All surveyed banks control employee AI usage.
Two-thirds of banks have a dedicated AI risk committee.
Hedge funds are more likely to govern AI through existing risk committees.
Around one-fifth of hedge funds allow relatively free staff usage.
Around one-third of proprietary trading firms have no formal AI governance policy.
This is not a minor documentation issue. It is a difference in whether the firm can control AI at the point where it affects decisions, communications, transactions, and risk.
The economic pressure to deploy AI will not slow down. KPMG’s 2026 AI Pulse Survey on asset management and private equity reports that 44% of leaders expect AI ROI in the 10% to 30% range, while approximately 72% expect payback within 24 months. Yet roughly one in three cannot quantify AI ROI at all.
That combination, high expectations and weak measurement, creates a predictable failure mode. Firms approve AI investment based on broad productivity claims, then discover they cannot attribute savings to a specific workflow, establish the control boundary, or reconstruct a material decision.
Why standard MLOps and logging are not enough
Traditional MLOps solves important problems. It can track model versions, monitor performance, manage deployments, and capture system logs.
Those capabilities are necessary. They are not sufficient for capital markets.
A model log may tell you which version was deployed. It may not tell you which version made a particular allocation recommendation after a data refresh, a prompt change, an exception, and a human override.
Application logs may tell you that a workflow ran. They may not preserve the policy boundary that determined whether the workflow was allowed to continue.
A prompt archive may show what a user asked. It may not show the complete workflow state, the tools invoked, the data returned, the approvals requested, or the action executed downstream.
For regulated investment operations, the unit of accountability is not the model. It is the decision and action.
A production-grade control layer must connect:
The inputs and data timestamps.
The model, agent, and workflow version.
The prompt, task context, or workflow state.
The policy applied.
The approval or escalation path.
The action executed in the relevant system.
The outcome and measured value.
Any exception, override, or unresolved status.
Without that chain, an audit becomes a forensic reconstruction project. The firm has to search email, tickets, application logs, model registries, chat records, and system databases, and still may not be able to prove what happened.
Run this operational test this quarter
Choose one production AI workflow. Do not choose a low-impact experiment. Select a workflow that influences a trade process, compliance review, client communication, exception queue, reconciliation, or investment research decision.
Now select one material decision and attempt to reconstruct it end to end.
Ask for the evidence:
Inputs: What data entered the workflow, and when was each source captured?
Version: Which model, agent, prompt template, and workflow definition were active?
State: What did the system know at each decision point?
Policy: Which permissions, thresholds, limits, or restrictions applied?
Approval: Who reviewed the action, and what did they approve?
Execution: What action was taken in the OMS, EMS, CRM, email, ticketing, or other system?
Outcome: What happened afterward, and how was value or loss measured?
Exceptions: Were there overrides, failures, escalations, or unresolved items?
Set a time limit. Give the team one business day, not one month.
Weak firms usually fail on lineage or approval. They can show that an AI workflow ran, but not precisely what it used or who authorized the consequential step.
That is the test. If the evidence is scattered, incomplete, or dependent on one engineer’s memory, the workflow is not audit-ready.
Three control plane use cases that close the gap
1. A scoped kill switch
A kill switch should not be a dramatic “turn off all AI” button. It should be scoped to a specific agent, desk, workflow, data source, or action type.
If a model begins producing anomalous recommendations, a data vendor changes format, or a policy condition changes because of a market event, the firm should be able to pause the affected workflow without disabling unrelated automation.
The control must also be tested like business continuity. Who can activate it? What happens to in-flight tasks? Are pending approvals preserved? How does the workflow resume? Is the event logged?
A kill switch that exists only in architecture diagrams is not a control.
2. Policy-as-code gates
Policies should be executable, not merely written in PDFs.
A control plane can enforce approval requirements based on role, asset class, notional, counterparty, communication channel, strategy, or risk threshold. It can prevent an agent from sending a client communication, releasing an exception, or initiating an operational action until the required conditions are met.
The important point is not just blocking bad actions. It is recording the decision boundary:
Which rule was evaluated?
What threshold was triggered?
What data supported the decision?
Was the action approved, rejected, or escalated?
Who was responsible for the final decision?
This is how a firm moves from “the AI is supervised” to evidence that supervision actually occurred.
3. One-click audit export: the Audit Bomb
The final control is the ability to export the complete evidence set for one decision or workflow without launching a forensic project.
The export should package the inputs, timestamps, model and agent versions, prompts or workflow state, policy evaluations, approval records, tool calls, executed actions, outcomes, and exceptions in a coherent chain.
Call it an Audit Bomb if the name helps people remember the requirement: when the question arrives, the firm should be able to produce the evidence in one controlled export.
Not screenshots. Not a collection of disconnected logs. Not a narrative assembled after the fact.
A decision record.
The ROI reframe
KPMG’s findings point to the central financial problem: firms expect meaningful AI returns, but many cannot quantify them.
The answer is not another dashboard with a larger productivity estimate. It is workflow-level attribution.
A governed AI workflow can show:
Hours removed from a defined process.
Exceptions prevented or resolved.
Cycle time reduced.
Errors avoided.
Human review concentrated on higher-value decisions.
Funding, market, and operational risks surfaced earlier.
Cost per completed workflow before and after automation.
That is the difference between an AI story and an operating capability.
QUANTEX approaches this as a Neurosymbolic AI Control Plane: a governed brain that combines flexible AI reasoning with explicit policies, permissions, approval gates, and audit evidence. The AI can interpret complex information and coordinate work. The symbolic control layer defines what it is allowed to do. Humans remain accountable at the points that require judgment.
The standard is not vague trust. It is 100% explainability for every material workflow decision: what happened, why it happened, under which policy, with whose approval, and with what result.
The firms that win the next phase of AI adoption will not be the firms with the most pilots. They will be the firms that can prove their production workflows are controlled, measurable, and reversible.
If your firm is moving AI from pilots to production, explore the QUANTEX AI Control Plane. To discuss a workflow, control boundary, or audit-readiness challenge with the team, get in touch.
Comments