Boards approve the AI project that cuts costs, because a cost saving comes with a number attached. Growth usually does not. So the decision gets made on whatever can be measured, well before anyone argues about which use is actually better.
A staffing cut produces a number by the next board meeting. Proving AI brought in new business takes data most institutions cannot pull together. So the cut gets approved. Build the data first. And know that using AI to grow puts it near decisions about real customers, which carry rules that internal tools never trigger.
Tim Pranger, founder and CEO of the AI vendor Appli, wrote in The Financial Brand on July 17, 2026 that credit unions have split into two camps. One uses AI to justify cutting positions, because lower headcount is easy to show a board. The other uses it to reach more members and close more loans, then hires to keep up with the growth it created.
He is right about the split, and we see the same thing in community banks. His fix is to reframe the internal conversation around growth instead of savings. That advice skips the reason boards keep picking savings.
Why do boards choose AI cost cutting over growth?
Because you can put a number on a cut and not on growth. Remove a role and payroll drops by a known amount on a known date. Proving AI won you new business is harder. You have to show what a member was about to need, what your system did about it, and whether any of it worked.
Those three facts usually live in three different places: the core system, the loan origination system, and the marketing and web tools layered on top of them. None of it was built to connect. Getting one clean answer means exporting from each and stitching the pieces together by hand, which takes weeks nobody has before a budget meeting.
So the CFO asks what the AI returned. The cut has a real number behind it. The growth case has a story. The cut wins on evidence rather than on merit.
Is using AI for growth more regulated than using it to cut costs?
Yes. Cost cutting AI mostly works on your own paperwork: summarize a policy, draft a memo, find a document from 2019. Growth AI works on people. It helps decide who you approach about a loan, and it helps decide who gets approved for one. Both are regulated, in different ways.
Keeping those two straight matters. Choosing who to approach is a fair lending question. If the model ends up targeting in a way that tracks race, age, or another protected class, you have a problem even though nobody was ever formally declined.
Choosing who gets approved is a separate rule with a separate obligation. Turn an applicant down, and you owe that person a specific reason for it.
When an internal summarizing tool goes wrong, you have a bad memo. When a model that scores applicants goes wrong, you have someone who was treated differently and a file that cannot explain why.
“The statement of reasons for adverse action required by paragraph (a)(2)(i) of this section must be specific and indicate the principal reason(s) for the adverse action.”
The same subsection says that telling an applicant they “failed to achieve a qualifying score on the creditor’s credit scoring system” is insufficient.
None of that argues against growth. It means growth carries obligations the cost cutting path never triggers. No examiner has ever asked a bank about a summarized memo. They ask who you approached, who you approved, and why the other applicant was declined.
Did the CFPB withdrawing Circular 2022-03 remove the AI adverse action rule?
No. On May 12, 2025 the CFPB withdrew its guidance, not the regulation. Circular 2022-03 explained how adverse action rules apply to complex models. Regulation B was never touched, and 12 CFR 1002.9(b)(2) still requires a specific principal reason. The obligation stands and the explanation is gone.
In 2022 the Bureau published Circular 2022-03, which said a creditor “cannot justify noncompliance with ECOA and Regulation B’s requirements based on the mere fact that the technology it employs to evaluate applications is too complicated or opaque to understand.” It was the clearest official word that a black box model is not an excuse for a vague denial.
Then it was pulled. Circular 2022-03 was item 14 on a list of 67 guidance documents the CFPB withdrew in a single notice. Circular 2023-03, which covered adverse action sample forms, was item 7 on the same list.
The Bureau called the withdrawal “not necessarily final” and said it intends “to continue reviewing all guidance documents to determine whether they should ultimately be retained.” Read that carefully. Guidance came off the table. The rule it described did not move.
If a model helps decide who gets approved, you still owe a specific reason to anyone you turn down. You now have less official help working out what counts as specific, which is a good argument for building the explanation into the system instead of reconstructing it a year later. Community banks are working through the same problem under SR 26-2 model risk guidance.
What does an AI data foundation need to do before you chase growth?
Four things, and only one is analytics. It has to connect a member to an outcome in a single query. It has to keep a record a person can still follow a year later. It has to keep member data inside your control. And on judgment calls, it has to show its reasoning instead of just handing back a verdict.
| What it has to do | Why it decides the outcome |
|---|---|
| Connect a member to an outcome in one query | Which member, what you did, what came of it. If that answer takes a week of exports to assemble, it never reaches the board in time to matter. |
| Keep a decision record | Not a dashboard. What the system recommended, what a person decided, and when. Written plainly enough that someone can still follow it a year later. |
| Keep member data in your control | Sending member financial records to an outside vendor for scoring raises questions your examiner will ask. Running the same work inside your own network does not. |
| Show reasoning on judgment calls | Whether a member is a good credit risk is not a lookup. A system that answers it in the same confident tone it uses for a rate question has made a decision that belonged to a person. |
That last row is the one most platforms skip. We wrote about the difference between a question your documents settle and a question that calls for judgment in there is no right answer, and your AI should say so. For credit unions specifically, private member service AI covers what keeping that work in-network looks like day to day.
Before the next budget cycle
The window Pranger describes is real. So is the reason most institutions are not walking through it. Boards are not short on imagination. Nobody can show them the growth number, so they approve the cut instead.
Build the thing that produces the number, and build it so it can also answer a regulator. The strategy conversation gets much shorter after that.
Sources: Tim Pranger, “Is AI Crisis or Opportunity? For Credit Unions, the Answer is Both … And Neither,” The Financial Brand, July 17, 2026. Pranger is founder and CEO of Appli, an AI vendor serving banks and credit unions. Filene Research Institute Report #639, “The AI Adoption Journey: A Survey of Credit Union Leaders,” January 30, 2025, by Lamont Black, Jessica Gamache, and Addison Davis, based on 110 participants across 78 organizations. Consumer Financial Protection Bureau, “Interpretive Rules, Policy Statements, and Advisory Opinions; Withdrawal,” 90 FR 20084, applicable May 12, 2025. Consumer Financial Protection Circular 2022-03, 87 FR 35864, June 14, 2022. Regulation B, 12 CFR 1002.9.
See what an auditable answer looks like
Book a short demo and watch a private model work inside your network. Every answer shows its evidence, and every decision leaves a record a reviewer can open later.
Request a Demo