Picture a member calling her credit union at nine on a Saturday night. There is a charge on her debit card she does not recognize. The new AI agent answers on the first ring, confirms who she is, reads her the last five transactions, and asks whether there is anything else it can help with. She says she did not make that charge. The agent offers to read the transactions again. She hangs up, a little frightened and a lot annoyed, and calls back Monday morning to wait for a person. On the vendor’s dashboard, Saturday’s call counts as contained: the member was served without reaching a human.
On the vendor’s dashboard, Saturday’s call counts as contained: the member was served without reaching a human.
That gap, between a call the system counts as handled and a member who got what she needed, is the subject here. This article is for credit union and community bank leaders who are being shown AI call center platforms right now and want to know what to ask before they sign.
The timing is not an accident. Cornerstone Advisors’ 2026 survey of 416 bank and credit union executives found that 59 percent of credit unions had deployed generative AI in some form, against 49 percent of banks, and trade press coverage of the same survey reported that credit unions were nearly twice as likely as banks to have invested in chatbots or conversational AI. On September 11, 2026, the NCUA joined the FDIC, the Federal Reserve, and the OCC in proposing new interagency guidance on managing third-party relationships. Credit unions are moving faster than banks on member-facing AI, and the rules for overseeing the vendors who supply it are being rewritten at the same moment.
Contained Is Not the Same as Resolved
The number every AI call center vendor leads with is containment, the share of contacts the system handles without passing them to a person. It is a useful number, but it measures the vendor’s work, not your member’s outcome. A call can be contained because the member got an answer, or because the member gave up.
Vendor case studies for credit unions illustrate the problem. The published figures we found range from about a third of calls fully handled to more than four fifths of sessions contained, and the definitions differ from one to the next: some count calls, some count sessions, some count anything that did not reach an agent within the same contact. Most do not report a baseline, and none that we saw reported how many contained members called back within a week. These are vendor marketing figures, and we cite them only to show how loosely the word is used.
So the first question to ask is what the vendor means by contained, and the second is what happens to those members afterward. Ask for repeat contacts within seven days, complaints, and dispute or fraud claims opened through another channel, all measured on the same members the system counted as served. Our view is that resolution, the member’s issue closed with no repeat contact, belongs in the contract, and containment belongs on the cost report.
Ask Where the Handoff Goes
The moment that matters most in an AI call center is the moment the AI should step aside. In a 2023 research report on chatbots in consumer finance, which is still posted on its website, the CFPB warned that automated systems may fail to recognize when a consumer is invoking federal rights, noted that sometimes only specific words or phrasing trigger the recognition of a dispute, and cautioned against bots that leave customers without an offramp to a human representative. That report is research, not a rule, but it names the exact failure in our opening picture.
Ask the vendor to show you, live, how the system recognizes a dispute, a fraud report, a hardship request, a complaint, and a member who is simply upset, in the words members actually use rather than the words in the demo script. Ask how quickly a member can reach a person by asking for one, and whether the system ever makes the member ask twice. Then ask what the human agent sees when the call arrives. A handoff that drops the member into a queue to repeat everything from the start is a containment strategy, not a service strategy.
The Diligence Is Yours
A community bank has one advantage here that a credit union does not. Federal banking regulators can examine the technology service providers that banks rely on. The NCUA’s authority to examine credit union vendors and CUSOs expired at the end of 2001, and it has not been restored. The Government Accountability Office repeated its recommendation in 2025 that Congress grant that authority, and a bill to do so was introduced in September 2026, but as of this writing it has not moved.
What that means in practice is that the examiner looks at the vendor through you. The NCUA’s own AI resource page, updated this spring, says existing regulations are technology neutral and apply to AI, and that a credit union using a vendor’s AI must conduct appropriate due diligence. The Government Accountability Office found in May 2025 that NCUA’s own model risk guidance does not cover the range of AI models credit unions now use, and recommended NCUA broaden it. If the regulator finds vendor AI hard to see into, assume your examiner will want to know how you saw into it.
Two other obligations travel with the contract. The NCUA’s cyber incident rule requires a federally insured credit union to report a reportable incident within 72 hours, and it covers incidents at a third-party provider, with the clock able to start when the vendor tells you. A platform that records, transcribes, and verifies members holds exactly the data that rule exists for. And the proposed interagency guidance, which the agencies describe as principles based and non-binding, will replace the current third-party guidance when it is finalized. Its direction is to tailor oversight to the risk of each relationship, and a vendor that speaks to your members on your behalf sits near the top of that scale.
States are moving as well. In September 2026, Connecticut’s Department of Banking issued a bulletin on AI governance for the banks and credit unions it charters, including customer service uses, which law firm summaries describe as expecting human override and vendor due diligence. If you hold a state charter, check what your own regulator has said.
What the Early Adopters Learned
The cautionary evidence comes mostly from outside the credit union world, and it is worth knowing. In early 2024, the payments company Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, the work of about 700 full-time agents. By the spring of 2025, according to press reports, its chief executive was saying publicly that the focus on cost had produced lower quality and that the company was investing in human support again. In Australia, the Commonwealth Bank reversed the elimination of 45 customer service roles in 2025 after a union challenge, conceding, as reported, that its redundancy assessment had not adequately considered all relevant factors before it introduced a voice bot to replace them.
Gartner predicted in June 2025 that by 2027, half of the organizations that had planned to significantly reduce their customer service workforce because of AI would abandon those plans. That is a forecast based on a poll of 163 service leaders, not a finding. Gartner’s 2024 survey of 5,728 customers is more direct: 64 percent said they would prefer companies not use AI for customer service, and 53 percent said they would consider switching to a competitor if they found out a company was going to. Those were general consumers, not credit union members, but a cooperative that sells itself on knowing its members should take the number seriously.
Those were general consumers, not credit union members, but a cooperative that sells itself on knowing its members should take the number seriously.
Where AI Helps Most Today
None of this means AI does not belong in the contact center. The strongest evidence points to a different place to start. The largest field study of generative AI at work, by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond in the Quarterly Journal of Economics in 2025, followed more than 5,000 support agents and found that an AI assistant working beside them raised issues resolved per hour by 15 percent on average and by about 30 percent for newer agents. Customer sentiment improved, and the gains were largest for the people who needed help most. That was a software company, not a financial institution, but the design carries: the AI drafted, suggested, and looked things up, and a person stayed on the call.
Our judgment, for most credit unions and community banks, is to sequence the work. Start with AI that supports your member service staff, such as call summaries, suggested answers drawn from your own policies, and faster account lookups. Add self-service for the narrow, low-risk requests members already prefer to handle themselves, such as balances, payment dates, and branch hours. Let the AI speak for you on anything involving money in motion, disputes, or a member in difficulty only after it has proven, on your members, that it knows when to hand off.
Questions to Ask Before You Sign
We suggest working through a short set of questions with every vendor, and these are our judgment rather than a regulatory checklist. How do you define containment, and what is your repeat-contact rate on contained calls at institutions like ours? Show us how the system recognizes a dispute, a fraud report, or a request for a person, and let us test it with our own members’ language. What does our human agent see at handoff? Where are our call recordings and transcripts stored, who at your company can reach them, and are they used to train models that serve other clients? Which subcontractors touch our member data, and how will you notify us of an incident in time for us to meet a 72-hour reporting clock? What evidence can you give us that our examiner will accept about how the model behaves and how it is changed? And what happens to our data, and to our members’ service, if we leave?
A vendor that answers those questions with specifics is a vendor you can supervise. A vendor that answers them with the demo again is telling you something too.
Pilot It Against a Comparison Group
The best answer to most of these questions is your own data. Before any rollout, run the AI on a share of calls, or for one call type, while a comparable group of calls goes to your current process, and agree in advance on the numbers that decide the outcome: resolution, repeat contacts, complaints, disputes caught, and member satisfaction, alongside cost. That is the discipline we described in The Control Group Is Gone, and it is the Prove step of Project Vista.
We would also treat a member-facing AI agent as Production AI from the first day, not as an experiment that graduates later. Our Citizen AI and Production AI pipeline lets staff try low-risk tools quickly on internal work, and holds anything that speaks to members, touches their data, or can affect their money to a higher gate: a named owner, tested handoffs, monitoring, and a documented way to change it. A credit union that can show its examiner that gate, and the comparison group behind it, is in a far better position than one holding only the vendor’s slide.
What This Will Not Do
The opening scenario is a picture, not a case. The credit union figures we cite come from vendor case studies, the Klarna and Commonwealth Bank accounts come largely from press reports, the Gartner prediction is a forecast, and the customer preference survey and the field study were not of credit union members. The CFPB report is research, not a rule, and the interagency guidance is a proposal that may change before it is final. Our question list and sequencing advice are our judgment, built on that evidence. This is not legal or compliance advice, and your examiner and counsel should weigh in on any specific contract. What the evidence supports is a posture: buy the outcome, not the containment rate, and prove it on your members before you scale it.
Buy the outcome, not the containment rate, and prove it on your members before you scale it.
Where to Start
Pick one call type, write down how you would know the AI served those members well, and take that definition into your next vendor meeting. Our credit union technology practice works with institutions on exactly these decisions, and our independent assessment can review a vendor proposal and the pilot design before you commit. For the longer view of where member service is heading, see Financial Services in 2040. If you would rather talk it through first, you can reach us through the contact form on our About page.
