Every AI lending vendor raising money right now sells the same reassurance: our agents are compliant, our agents are auditable, our agents will not put you in front of a regulator. Kastle, which closed a 24 million dollar Series A this month to build what it calls an AI workforce for consumer lending, is the latest and most explicit version of that pitch. I do not doubt the engineering. I doubt the premise that there is a settled rulebook these agents are being built to satisfy. There is not, and the industry is not being honest about why.
The pitch leans on a rule that no longer exists
Kastle’s own funding announcement quotes Rebecca Liu-Doyle of Insight Partners, the fund that led the round, making the case for why banks should trust an AI agent with lending operations: “Financial institutions do not need another layer of software that creates more work for their teams. They need AI that can reliably complete the work while ensuring compliance.” Kastle’s co-founder Rishi Choudhary makes a similar claim, that the platform can navigate “regulatory rigor” without the multi-year risk of ripping out core systems.
That framing assumes there is a clear standard an AI-driven lending decision has to meet. For years there was one. The Consumer Financial Protection Bureau’s Circular 2022-03, and its 2023 successor, Circular 2023-03, told lenders in direct terms that a complex algorithm does not excuse them from giving an applicant a specific, accurate reason for a credit denial. Then-Director Rohit Chopra put it as plainly as a regulator ever does: “There is no special exemption for artificial intelligence.” Both circulars were withdrawn on May 12, 2025, along with dozens of other pieces of CFPB interpretive guidance, according to the Bureau’s own withdrawn-guidance record. The underlying requirement in Regulation B, that a creditor must give specific reasons for adverse action, was not repealed. Only the agency’s explanation of how to satisfy it for a black-box model was taken off the table.
Advertisement
300 × 250
The counter-argument, and why it is wrong
The obvious rebuttal is that this is good news for vendors like Kastle: fewer explicit federal rules should mean faster deployment and lower compliance overhead. I would take that argument seriously if guidance withdrawal actually reduced liability. It does not. Regulation B’s specific-reasons requirement is a statutory obligation, not a discretionary preference the Bureau can waive by silence. What changed is not the standard, but who gets to interpret it, and how. With the agency’s own explanatory circular gone, and CFPB examinations reduced, the practical work of deciding whether an AI-driven denial explanation is “specific” and “accurate” enough shifts to case-by-case litigation and to whichever state regulator is still active in this space. That is a less predictable environment for a lender to operate in, not a more permissive one, and it is a worse one for a vendor whose product is the decision engine itself.
What it means for the lending leader
A bank or lender evaluating an “AI workforce” platform should be asking a narrower question than whether the agent is fast or accurate. It should be asking who owns the adverse-action explanation when a denied applicant asks why, and whether that explanation was generated in a form defensible without a federal circular to point to. Kastle’s platform, by its own description, is built to operate inside a lender’s existing systems of record and existing controls, which is the right structural answer: the explainability obligation stays with the institution, not the vendor, no matter how autonomous the agent’s day-to-day work looks. Any AI lending platform, Kastle’s included, is only as compliant as the explanation it can produce for a single denied applicant, defended alone, without appeal to how the model behaves on average.
This is the same accountability question sitting underneath the broader push to give AI agents standing inside regulated finance, from the charter race now underway for AI-native banks to state examiners building their first formal AI supervisory playbooks because the federal one is thinner than it was two years ago. The industry keeps building the automation first and asking who is accountable for it second.
What to watch
The real compliance benchmark for AI lending platforms will not be set by a vendor’s marketing copy or by an agency circular that no longer exists. It will be set by the first Equal Credit Opportunity Act case where a lender has to defend, in front of a judge, a specific denial reason its AI agent generated with no CFPB template to lean on. Until that happens, “ensuring compliance” is a claim, not a demonstrated capability, and lenders buying these platforms should price that uncertainty in now, not after the first lawsuit tells them what the real standard was all along.
