
I’ve been nerding out on two things from the first week of November. The OCC published a framework for supervising AI agents, and Wells Fargo updated its online banking agreement to cover them. Both put a human in the loop on paper.
That’s a reasonable default. What I’m less sure of is whether anyone’s designing the actual human in the loop. Putting a person in the middle of an agent’s work is a policy call. What that person can actually see, approve and stop is a design call, and most of that is still open.
What the OCC wants from the human in the loop
The OCC’s framework (Bulletin 2026-54) is supervisory guidance for its examiners. It isn’t a new rule, and it’s careful to call its tiers and controls illustrative. There are a few parts in here worth taking in if you’re building digital products in the finance world.
It says human approval has to be “more than a nominal click.” It calls out alert fatigue and asks how many agents one person can credibly monitor. And it lists the controls you’d expect for anything that moves money: circuit breakers, rate limits, least-privilege permissions, kill switches.
Read as a design brief, it’s pretty specific, which is great. A human in the loop clicking approve on a steady stream of agent actions without the context to judge them isn’t really overseeing anything.
The framework also says revocation of outside agents “may be calibrated to the action requested.” That’s a control mechanism, but a person still has to know how to use it. At Grand we spend a lot of our time on calibrated trust: getting someone’s reliance on AI to match what the system can actually do.
A nominal click is what over-trust looks like from the inside. You approve because everything before it was fine. And if stopping an agent is hard to do and harder to show you did, a careful customer just won’t turn one on.
What the bank’s contract says about AI agents
Wells Fargo’s updated Online Access Agreement takes effect November 24, and a new Section 15 covers third-party AI agents. Instructions your agent gives the bank are “deemed to have been initiated and authorized directly by you, without further verification.” Using one is “at your own risk,” including when it acts on bad or manipulated inputs.
If you revoke an agent, you have to keep evidence that you did and hand it over if the bank asks. Until you notify the bank, you’re responsible for what the agent does, to the extent the law allows. The bank, for its part, can require you to register agents and can suspend or cut off their access.
Douglas Borthwick, who spent time at Morgan Stanley and Standard Chartered, told American Banker that telling a chatbot to stop “leaves nothing in the customer’s hands that anyone else can check.” In other words, there is no receipt. The contract asks the customer to keep evidence. Whether the product gives them any is still an open design question.
Two documents, one human in the loop
The regulator wants the human in the loop to be able to oversee for real, without drowning in alerts. The bank wants the customer responsible until they say stop, with evidence kept on hand. Both are fair, protective asks, and they picture the same person: someone who knows what the agent’s allowed to do and can shut it off with a record.
I don’t think most of the products I see are built for that type of person yet. The questions are just newer than the products.
I’d start with the concrete ones. An approval should transparently show the payee, the amount and why the agent picked them, so saying no takes as little effort as saying yes. An approval you can read in two seconds is also how one person keeps up with more than one agent. And a customer should be able to say in plain words what they’ve let an agent do, and have that match its real permissions.
Then there’s the receipt. It doesn’t need to be clever. A confirmation that names the agent, says what access ended and when, and matches the record the bank keeps would give a customer something to hand over. If stopping an agent means typing “stop” into a chat window, there’s nothing to hand over, which is Borthwick’s point.
The bank has screens to design here too. Registering an agent, reviewing it and cutting off its access are all moments someone on the bank’s side has to get right. A clean revocation record helps them as well, since if a dispute comes up, both sides start from the same set of facts.
Designing the loop
One thing I’ve noticed in our financial services work is that AI wins trust a little bit at a time, as long as the person retains ownership of the decision. On a trading-desk tool, the easy version would have handed traders a pricing window and told them to use it. We showed the reasoning in layers, with the boundaries behind each recommendation, and protected the trader’s own read of a fast-moving market. Most of the work was deciding which parts of the decision-making process belonged to the system and which stayed with the person.
On a wealth management engagement, we went in expecting to improve a self-service tool. What we found was that advisors and their clients didn’t want self-service at all. They wanted the relationship between them made stronger, more collaborative.
Agents raise the same question, except now the money moves without anyone touching it: which parts of a decision does the person hand over, and how do they take them back? In my experience you answer that by testing with real customers before the permissions model gets built, because it’s expensive to change once it’s in code.
The approve button, the scope screen (where the customer can review what an agent is allowed to do) and the off switch are what a customer remembers when something goes wrong. If you leave them to a disclaimer, you’ve pushed the risk onto your customers and called it consent. I think the answer is showing people the right thing at the right moment.
That trust gets decided by one person at one screen, choosing how much to let an agent do with their money. The regulator and the bank have both put that person in the loop. What’s left is designing agentic experiences they can actually steer with confidence.
If you’re working on how customers approve and revoke an agent, we’d like to hear what you’re building. Bring any in-progress ideas, and we’ll go through what we’d test before the permissions model gets built. Let’s connect.