Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter curates critical developments in agentic AI and explains why they matter for community financial services.

At bunq, the customer-service agent can now log in as the customer. It reaches the same accounts that person sees, and it can act in their name

Moving an agent from demo to everyday use happens in stages. This week shows three. Eltropy built one way in: a marketplace that certifies outside agents and distributes them to 750-plus credit unions and community banks.

bunq shows the integration stage. Its first version ran outside the bank’s core systems and handled about a third of the work. After three rebuilds, Finn now runs inside them. Blend shows the control stage: the lender names who can use the agent, decides which actions require approval, and can shut it off in five minutes.

This is bigger than three companies. BNY and UBS give agents their own logins and make a person approve any money that moves. Canada's regulator told banks to give agents only the access they need, an off switch, and a full log.

Taken together, the week reads like a green light. Agents can now carry a bank's support, work a mortgage file, and act in a customer's name. The ones doing it well put agents on serious work, then built the limits, logs, and off switches to keep it safe.

Sources: MD|DC Credit Union Association, Jul 2 · Finopotamus, Jul 8 · FFNews, Jul 9

Eltropy opened its AI agent platform to outside fintechs. They can now use Eltropy’s integrations, test environments, and certification process, then offer approved agents to more than 750 community financial institutions. Eltropy says accepted developers get test environments modeled on a live credit union, compliance documentation, and a path to distribute their agents across the network once certified. The company calls it the first phase of a governed marketplace.

Credit unions and community banks can also build their own agents on the platform, which Eltropy says connects to more than 50 banking systems. A separate 90-minute course walks staff through building one.

Shared integrations could make specialized agents practical for smaller institutions. Instead of connecting separately to every credit union or community bank, a fintech could build once on Eltropy’s platform and reach hundreds of potential customers.

Source: The Stack (sponsored by AWS), Jun 30

Benjamin Kleppe, bunq's machine-learning engineering lead, wrote Finn's first lines of code. The first version sat beside the bank's systems and stalled at roughly 30% automation.

Kleppe's team rebuilt Finn from scratch three times. A later version could reach the bank's core systems. But bunq blocked anything it judged unsafe for customers, so Finn still could not finish the harder requests.

Finn now runs inside bunq's core systems. It can sign in as the customer and see everything that person can.

A main model passes each request to a smaller, cheaper agent built for that job. bunq says the setup gave better answers, cost less, and ran faster.

bunq says Finn now handles 97% of user interactions and resolves 87% of customer questions on its own. At a bank with 20 million customers, that is nearly the whole support job. bunq also logs every action. The record shows who set it off, what tools it used, and what it changed. Legal and compliance joined at the start.

Sources: National Mortgage Professional, Jul 9 · Blend Help Center

Five mortgage lenders moved Blend Autopilot into commercial use after a four-month preview across more than 25,500 loans. The agent reviews borrower documents and checks them against lending guidelines. It calculates the income that qualifies for the loan and asks for anything missing before underwriting.

Lenders decide who can use Autopilot and what needs a person's sign-off. They can limit it to named loan officers, require approval before it sends a follow-up or changes an application, and set the rules it follows for each product. Those rules can be Fannie Mae's, their own, or custom. A running log shows the rule behind every action.

Autopilot does not approve or deny loans. A loan officer reviews its work, and an underwriter or the automated underwriting system makes the credit decision. A lender can switch it off from the same settings screen in under five minutes.

Controls are why a lender can hand an agent 25,500 loans and still know what it is doing at every step. The lender picks who uses it, signs off on the actions that matter, sets the rules for each product, and keeps a five-minute off switch. A lending team can start a pilot tomorrow and point to where a person stays in charge.