Only 8% of the 26 systemically important banks in a Boston Consulting Group benchmark had a comprehensive, standardized framework for monitoring and controlling AI agents. The study, published November 13, found that 73% of the banks allowed agents in production for specific uses. Authorization for broad production use stood at zero across the sample.
The benchmark covered banks designated as systemically important at the global or domestic level. BCG found that 92% were piloting or permitting agentic AI. At 88% of the banks, agents accounted for 10% or less of the AI portfolio. The figures describe how widely banks are experimenting with agents and what they authorize in production. Actual use still makes up a small share of most banks’ AI portfolios.

Tracking AI agents after approval
Most banks in the study already had policies for AI. BCG found that 96% handled AI risk through existing risk frameworks and 77% had a dedicated AI policy. Fewer than half kept a centralized AI inventory in a dedicated tool. The benchmark recorded zero banks with a comprehensive human oversight framework applied consistently across their AI systems.
That gap matters when an agent can take several actions in sequence. Consider a bank that authorizes an agent to gather documents for a case review. Giving that same agent permission to update the case record changes what the bank needs to monitor. The operations team needs a record of what the agent changed, a person responsible for reviewing exceptions, and a way to stop further changes if something goes wrong.
BCG said oversight also has to account for software providers that change AI features after an initial approval and for business users who build agents themselves. In either case, the scope of an agent’s access or actions can change after it first enters the bank’s inventory. Anne Kleppe, BCG’s global lead for AI safety and governance and a coauthor of the study, described the challenge as “turning governance from a framework into an operating capability.”
BCG reported that 46% of participating banks had experienced an AI-related issue or incident in the preceding 12 months. That figure covers AI broadly. More than a quarter said AI systems were falling short of their expected business case, a separate finding about performance against expectations.
Preapproved controls were linked to faster deployment
Banks in the top third of BCG’s AI risk management assessment were 2.5 times as likely to qualify as advanced AI adopters as those in the bottom third. BCG also found that banks using preapproved designs for lower-risk applications were nearly twice as likely to move those applications from idea to production in under six months. Both findings show associations within a sample of 26 systemically important institutions. The benchmark’s design leaves the direction of cause and effect open.
A preapproved design gives teams a starting point for recurring work. For example, an internal document assistant could use an approved set of data sources, access permissions, and testing procedures. BCG recommends reusing controls for applications that fit an established pattern and reserving closer review for unfamiliar or higher-risk uses. It also calls for ongoing production testing and monitoring, so banks can detect changes in how an agent behaves after it goes live.
Sources

ERP News Editorial Team
We cover ERP and enterprise technology for the people who select, implement, and run these systems.
Pitch a guest article or work with our editorial team on a news story, customer case study, or executive interview.
Editorial Calendar · Pitch a Guest Article · Commission a Story
Advertising Options · ERP News Magazine · ERPNews.com
Editorial Inquiries · [email protected]
Copyright & Usage · Our original editorial content is produced by the ERP News Editorial Team, with expertise in ERP and enterprise technology.
© ERP News. All rights reserved. Except as permitted by law, reproduction, republication, and redistribution require prior written permission or a licence from ERP News, along with clear attribution to ERP News as the source.
ERP News Content Licensing · License full-text ERP News articles for your internal research tools, delivered through a private JSON feed. [email protected]









