Skip to content

Content Library

Back to Articles

Article March 25, 2026

AI Approval Crisis Exposes Hidden Failures in Bank Governance

AI Approval Crisis Exposes Hidden Failures in Bank Governance

Michael Talbert, Head of Professional Services

As banks accelerate AI adoption, model risk teams are struggling to keep pace with approval demands. Michael Talbert argues the real issue is not speed but flawed governance design. Structural failures, misclassification, and late involvement are creating
  • AI approval pressure rising as deployment accelerates and scrutiny increases
  • Uniform governance frameworks creating bottlenecks across use cases
  • Leading firms classify AI by approval burden before validation begins
  • Structural failures driving most approval breakdowns not model quality
  • Late MRM involvement and poor documentation key issues
  • Misclassification exposing gaps between perceived and actual risk
  • Regulated functions require higher standards of governance
  • Agentic AI introducing new validation and accountability challenges
  • Early governance involvement critical to approval success
  • Model risk shifting from gatekeeper to governance partner 

As artificial intelligence becomes embedded across financial services, model risk teams are facing intensifying pressure to approve systems at pace while maintaining rigorous governance standards.

Yet, according to Michael Talbert, Head of Professional Services at leading unified compliance control platform Behavox, the industry is focusing on the wrong problem.

“The instinct is to ask how do we approve AI faster,” Talbert says. “But the right question is which AI actually requires deep scrutiny, and which doesn’t.”

This distinction is becoming critical as institutions struggle with mounting approval backlogs. Many banks continue to apply broadly uniform governance frameworks to all AI use cases, regardless of their risk profile.

The result is a system that creates friction at both ends. Low-risk tools become trapped in validation processes designed for more complex systems, while genuinely high-risk applications do not always receive the depth of scrutiny they require.

Talbert argues that leading institutions are beginning to address this by rethinking classification before governance even begins.

“They classify AI use cases by approval burden before the governance process begins,” he explains. “That changes everything.”

By doing so, model risk teams can focus their efforts on systems that genuinely matter, particularly those operating in regulated control functions where errors carry regulatory, financial, or reputational consequences.

Contrary to common assumptions, most AI approval failures are not driven by model complexity or technical limitations.

“The conventional diagnosis focuses on model quality,” Talbert says. “That’s real - but it’s not where most failures actually originate.”

Instead, Talbert points to structural weaknesses as the primary cause. The most common failure is late involvement from model risk teams. In many cases, technology or compliance teams build or procure AI systems independently, with model risk brought in only at the final stage.

“By that point, documentation is incomplete, architecture decisions are locked, and explainability has been treated as an afterthought,” he says. “Approving it becomes extremely difficult. Rejecting it creates significant business friction. Nobody wins.”

Another recurring issue is use case misclassification. Teams may believe they are developing decision-support tools, but in practice, the outputs are driving regulatory decisions.

When properly assessed, these systems fall into much higher-risk categories than initially assumed. The gap between perceived and actual risk can be significant, exposing institutions to governance failures.

Documentation failures further complicate the process. “That is not a governance process - it is archaeology,” Talbert says, describing situations where validators are forced to reconstruct model logic and assumptions after the fact.

These challenges become more acute in regulated control functions such as communications surveillance, market abuse detection, anti-money laundering, and conduct risk monitoring. “The stakes are categorically different,” Talbert emphasizes.

Systems in these areas must meet far higher standards of transparency, auditability, and explainability than general-purpose AI tools.

This requires model risk involvement much earlier in the lifecycle. “The choice between a large foundation model and a specialized, deterministic system should be informed by governance requirements,” Talbert explains, “not decided by the technology team and handed to validation.”

Events you may be interested in

View all of our upcoming events

Sign Up to Our Newsletter

Sign up

Connect Plus

Unlock more with Connect Plus membership

This content is available to Connect Plus members.