I framed the scariest AI incident in banking this month as a login, not a hallucination.
I noted that Major U.S. banks are finding third-party AI agents inside customer accounts after customers handed over their credentials.
I pointed out that In some cases, the agents went further than they were told—like calling the bank with a complaint the customer never approved.
I reported that JPMorgan, Citi, Schwab and others are restricting or blocking these agents.
I contrasted those bank restrictions with UBS, which says it has 12,700+ AI agents running across its workforce.
I noted that UBS says it has 600+ use cases in production.
I characterized the contrast as “Same industry. Same quarter. Opposite instincts.” and framed the debate as one about authority, accountability, and permission rather than whether agents are smart enough.
I cited Gartner’s finding that only 13% of large enterprises feel confident they can govern agents at scale.
I called the governance challenge the “authority gap” and asked organizations to classify an agent under their purview if it executes an unapproved, high-risk action.
For Tier 1, I described the software classification as “It’s just an API/app.” and asked whether its “chain of thought” can be audited when a decision goes rogue.
For Tier 2, I described the Users / External classification as “It has its own system login.” and asked who signed the OAuth / T&Cs for its autonomy.
For Tier 3, I described the Employees classification as “It has an identity, budget, and oversight.” and asked who conducts its performance audit and who is legally liable.
My take was that 2027 won't be won by the company with the best model, but by the company that can prove which of its agents did what, and on whose behalf.
I declared that Agent identity is the new KYC.
I identified the three authority questions as Permission, Capability, Liability.
I asked whether we are treating agents as tools, or whether it is time for a new class of digital identity.
I invited CTOs, Risk Officers, and AI builders to share how their teams address Permission, Capability, and Liability, and asked readers to comment with their tier—1, 2, or 3—and their view.