• Jouk Pleiter, CEO and Founder at Backbase


• 26.06.2026 10:15 am
#AgenticAI #AIGovernance

There has been a surge of AI pilots and point solutions in banking in 2026, with agents as one of the must-haves.Those not deploying them are already behind but for those that have started to, have they considered the risks? What happens when an agent makes a wrong call inside a regulated institution? 

Banking is not like other industries. When something fails, regulatory accountability lands with the bank. Not the model provider, nor the implementation partner.

Yet, despite the pressure that comes from operating in a highly regulated industry, banks are deploying agents into an operating model that was never designed to govern them. They're trying to make AI fit within an infrastructure built for human workers.

Why governance means pilots stay pilots  

The AI industry has been very good at talking about what agents can do. Less so about who answers when they don't. And, that’s the biggest reason why AI is paralysed inside most banks right now - governance.

Risk and compliance teams are taking a step back and asking, if this agent makes a consequential decision, can we explain it to a regulator? Can we show what it read, what policy it followed, what it decided? And, do we have a kill switch?

Right now, that answer, more often than not, is “no”. So the pilot stays a pilot, the board gets another excuse, and the AI-native competitors keep claiming ground.

Agents need the same infrastructure as human employees

Banking already has the model for solving this. Know Your Customer (KYC) exists because accountability requires registration, defined permissions, and a decision trail. The same logic applies to employees - every person inside a regulated institution has a fixed role, a defined scope, and a record auditors can follow. Agents need identical infrastructure. Banks need a Know Your Agent framework.

Every agent operating inside your institution should be registered. Its permissions explicitly defined. Every decision it makes logged - what data it reads, what policy governed the action and what it produced. This needs to happen from the moment it touches a live workflow.

The discipline that makes this work is the decision token. When an agent holds a token for a specific action within a specific boundary, it can act. Outside that boundary, it cannot. This protection is worth the time it takes to build.

The banks scaling AI built their governance layer first

Banks talk about needing a kill switch. That instinct is right but it's the wrong focus. A kill switch is a last resort. They should focus on defining boundaries tightly enough so that agents never cross them. A complete audit trail where a policy applied, scope operated within and a decision produced is what gives risk teams something to sign off on. It is what turns a regulator conversation from a confrontation into a demonstration.

If you go look at the banks leading the charge in rolling out agentic capability, you’ll find that they built their governance infrastructure first. Once the governance layer is there, you can start putting the agents to work in your bank.

AI ambitions stall when agents get deployed into an operating model built for humans. Fingers get pointed at the models  with claims that“it’s not ready”. But the operating model itself is non-compatible with the new actor on the frontline.

The AI industry will keep shipping capable agents and vendors will keep winning contracts but when something goes wrong, they will point to the small print.

The bank that registers its agents, defines their authority, and maintains a full decision trail is the bank that scales thousands of agents without losing control. The accountability is not going to move from the bank. The time to act is now or lose out to the AI natives.

You know your customer. You know your employee. It's time to know your agent.