A member services rep takes a call. The voice is familiar. The account details check out. The request is urgent and slightly unusual, and the caller is apologetic about that.
Everything about the call is right except the person on the other end of it.
The short answer: before deploying any AI tool that speaks to members or makes decisions about them, a credit union should be able to answer four questions. What member data the tool touches, who reviews its output, how a member can tell it apart from a person, and what happens when it gets something wrong.
Those questions matter more this year than last, because your members have started to distrust the exact channels you are being sold AI for.
What Deepfake Calls Have Already Cost Credit Unions
This is no longer a projection. Michigan State University Federal Credit Union deployed AI-powered call screening in 2024 and identified $2.57 million in fraud exposure from deepfake calls in a single year, close to a quarter of a million dollars a month that might otherwise have passed as ordinary member calls.
The detail that matters is how they found it. The deepfakes were caught by AI. Without a tool listening for what human ears can no longer catch, those calls were simply members having a slightly odd day.
Voice authentication is the specific casualty. Speaking at a Federal Reserve event, OpenAI’s Sam Altman said it is now crazy to rely on voiceprint authentication. Deloitte projects that generative-AI-enabled fraud losses in the US will reach $40 billion by 2027, up from $12.3 billion in 2023.
The part that lands hardest for credit unions is cultural. Knowing members by name and voice has always been the advantage over a national bank, and it is now the surface being attacked. When a cloned voice says it needs a transfer authorized today, the instinct to help works against the procedure.
Why That Changes Your Own AI Plans
Here is the connection most vendor pitches skip.
Your members are being trained, by their own experience and by the news, to be suspicious of voices and faces on a screen. At the same time, the AI products being sold to credit unions are voice receptionists, chat agents and automated member service.
You are being asked to introduce synthetic voices into member conversations at the moment members have learned that synthetic voices are how they get robbed.
Staying out is no longer the safe option either. The MSUFCU figure exists because they deployed AI, and a credit union relying on staff alone to detect cloned voices is defending a position that has already fallen. What matters is which uses you choose and how openly you tell members about them. A member who cannot tell whether they are talking to a person, a bot or a fraudster trusts the channel less every time they use it.
If You Currently Use Voiceprint Authentication
Three things worth doing before your next deployment decision, in order of how quickly they can be done.
- Demote voice to one signal rather than proof. It can still contribute to a confidence score. It should no longer be sufficient on its own to authorize movement of money.
- Set a dollar threshold above which out-of-band confirmation is mandatory. A callback to the number on file defeats a cloned voice at almost no cost, and gives staff a procedure to point at when a caller applies pressure.
- Ask your core and contact-center vendors what deepfake detection they already offer. Several have added it, so you may be paying for a control you have not switched on.
Six Questions to Ask Before You Deploy
- What member data does this tool see, and where does it go? Ask for the data flow in writing.
- Can a member tell immediately that they are talking to AI? Disclosure is a trust decision before it is a compliance one.
- What happens when it is wrong? Not whether it will be. A chatbot giving incorrect account or lending information is your liability.
- Who reviews the output, by role? Name a job title, not a department.
- Does this tool weaken any verification we currently rely on? Anything that makes it faster for a caller to prove identity makes it faster for a cloned voice too.
- What does the vendor do with our members’ interactions? Ask directly whether member conversations train the vendor’s models, and get the answer in the contract rather than the sales deck.
Question one is where most credit unions discover a problem they did not know they had. In every AIStack Challenge we have run, we have found at least one AI account in use that leadership did not know about. For a credit union, that account is an undocumented third party with access to member information. Our guide to what not to put into AI covers the categories that should never leave your control.
Does NCUA Have Rules for AI?
Not a standalone rulebook. NCUA evaluates AI through the frameworks it already has: vendor due diligence, fair lending, IT risk assessment and model governance.
In December, NCUA consolidated its AI guidance into a single resource page, framed explicitly around performing due diligence on third-party AI vendors, and tied it back to existing letters 07-CU-13 on evaluating third-party relationships and 01-CU-20 on due diligence over third-party service providers.
The practical translation is that you will not be cited for using AI. You will be cited for not governing it the way you govern every other vendor relationship, and the finding will reference a rule that already existed.
One structural point deserves more attention than it gets. The OCC and FDIC can examine third-party service providers directly. NCUA cannot, a gap the GAO flagged in its 2025 report. Your vendor due diligence therefore carries more weight than a bank’s would, because your regulator has less ability to independently verify what your AI vendor is actually doing. The vendor handles it is not available as an answer.
If you do not yet have anything written down, our guide to the one-page AI policy your team will actually follow covers what that document needs to settle.
What to Tell Members
Member education is the cheapest control available. Tell members plainly that your staff will never ask them to authorize a transfer on an inbound call, and give them a callback number to use when anything feels off. A member who has been told what you will never do has a rule to fall back on when a familiar voice asks for something strange. The same principle applies to wire fraud and business email compromise.
Frequently Asked Questions
Can credit unions use AI under NCUA rules?
Yes. NCUA has no standalone AI rule and evaluates AI through existing frameworks including vendor due diligence, fair lending and IT risk assessment. The expectation is documented governance rather than avoidance.
Is voice authentication still safe for credit unions?
Voice alone is no longer considered a reliable authentication factor, and AI-generated speech can defeat voiceprint matching. Current practice points toward treating voice as one signal within multi-factor authentication, with out-of-band confirmation required for sensitive transactions.
Who is liable when an AI tool gives a member wrong information?
The credit union. Third-party involvement does not transfer the obligation, which is why contractual protections and documented due diligence matter before deployment rather than after.
Ready to Ask These Questions of Your Own Setup?
First Call has supported Montana credit unions since 1998, across more than a million closed tickets and 250-plus clients.
The AIStack Challenge is a 20 to 30 minute call with someone who does this for a living. You leave with a written report covering what surfaced, where your data is going, and what to do next.
