The OCC made bunq prove it understood its own product, and an AI agent needs only a consent screen

The Office of the Comptroller of the Currency has published the letter refusing bunq a United States national bank charter, and it runs to six pages of specifics [1].

Most denials are a paragraph. This one shows its working.

What the letter actually says

Corporate Decision #1384 is dated 4 August 2026 and addressed to bunq's counsel at Morrison & Foerster, on OCC control number 2026-NE-Charter-344545 [1]. FinTech Futures carried it in its 14 August roundup, and American Banker reported it under the OCC's own word for the application, insufficient [3][2]. bunq is the Amsterdam neobank supervised by De Nederlandsche Bank; the proposed US entity would have sold deposit accounts and unsecured credit cards on four subscription tiers, earning from monthly fees and interchange [1].

The OCC denied it under 12 CFR 5.13(b)(1) and (2), and named three deficiencies [1].

On capital, the application proposed $50 million. The letter says bunq gave several different accounts of where it would come from, personal holdings at one point and a dividend from the Dutch bank at another, and supported none of them. Revised projections later put the figure at $58.3 million with no explanation of what changed [1].

On management, the proposed directors had no experience in unsecured credit cards, which was to be the bank's principal lending product. The proposed president and chief executive would have worked part time, sat on other boards, and spent most of the year outside the United States [1][2].

On profitability, the delinquency rate came from bunq's European book, and the proposed allowance for credit losses sat below other credit card banks the OCC supervises. The regulator's word for the assumptions underneath it was "not credible based on peer analysis" [1].

I have watched teams build a launch plan on a European loss curve because it was the only curve they had. It is a very easy thing to do and there is no cheap way to find out you are wrong.

The other direction of travel

The day before that letter was signed, Plaid published a partnership with Sierra that puts Plaid Link inside Sierra's customer service agents [4][5]. Plaid dated its post 3 August 2026; Sierra carries the same announcement [4][5]. The customer connects an account without leaving the chat, and the agent reads income and spending from there, or moves money through Plaid Transfer [4]. The named use cases are loan applications where a credit score falls short, insurance claims, fraud resolution and collecting overdue payments [4][5].

The engineering is the boring, correct kind. Plaid Link already handles consent, credential capture and the institution mapping, and putting it inside a chat surface is a placement decision rather than a new trust model. Nobody has invented a way for an agent to reach a bank account without the customer clicking the thing.

Two tests, and only one of them is written down

bunq had to satisfy an examiner that its board understood credit risk in a market it had not operated in, and the record of it failing runs to six pages with the reasoning attached [1]. An agent that assesses the same borrower's cash flow and recommends a refinance faces a different test, which is whether the customer tapped continue.

Plaid's product does what it says it does, and my interest here is in where the competence check sits. The chartering rule at 12 CFR 5.20(f) asks whether management has ability and experience relevant to the services provided [1]. No equivalent question is put to the model doing the assessing, because the model is not the institution. The institution is whoever deployed it, and they were already chartered.

If you are shipping an agent into a regulated workflow, Decision #1384 reads usefully as a specification. It wants a capital source evidenced rather than asserted, loss assumptions that survive a peer comparison instead of being carried over from another market, and a named person who is present and understands the product. All three transfer onto a model deployment about as well as they transfer onto a charter application, and in my experience the third one is the only one anybody puts in the launch checklist.

Read from the rails

A thin weekend. I went across agentic commerce, the card networks, AI payments and stablecoin rulemaking for the last 48 hours, and the honest answer is that nothing new and checkable landed on Friday or Saturday. Both items above are from early August and I have dated them rather than dressed them up. I would rather run a short brief than manufacture a third story.

One date does need putting in a calendar. The joint proposal from FinCEN, the OCC, the Federal Reserve, the FDIC and the NCUA on customer identification for permitted payment stablecoin issuers closes to comment on 21 August 2026, five days from now, at Federal Register document 2026-12460 [6]. If you issue, hold or plan to route a payment stablecoin, that comment window is the last cheap opportunity to say something about the identification rules you will spend the next several years implementing.

bunq can reapply. The letter says the OCC would expect any subsequent application to address the reasons for this one [1]. That is the whole of what it says about what happens next.