Most credit union AI conversations start in the wrong room. A vendor books a demo, walks the exec team through a slick chatbot, and the CEO leaves the meeting wondering whether to buy it. That is a sales cycle wearing a strategy costume.
The teams that get real value out of AI start somewhere else. They walk the back office. They look at what their operations, lending, and support staff already do every day, and they screen it against three signals. If a task hits at least two of the three, it belongs on a candidate list. The rest can wait.
This is the Three-Signals Test. Nothing about it is proprietary or new, and you do not need a consultant to run it. What it does require is discipline: a small team, an hour on the calendar, and a willingness to write down what your people actually spend their week doing.
The three signals
Signal 1: The task is done repeatedly on a schedule. Daily, weekly, monthly, or on a defined trigger. Not “sometimes.” Not “when a member happens to call in.” A task that runs 50 or 500 times a week is a task an AI system can be trained on, monitored against, and rolled back when it misbehaves. A task that runs twice a year is not.
Signal 2: Two people in different locations do the task nearly identically. If your Denver branch and your Boise branch handle member disputes with two completely different intake processes, this signal fails. There may be a good reason for the divergence, but it means the work is either genuinely non-standard or the standard has drifted. Either way, automation is premature. When two people you would trust to train a new hire produce nearly the same output from the same input, the task is standardized enough that a well-scoped AI or workflow tool can carry a first pass.
Signal 3: The end-to-end process takes under five minutes. This is the signal most CU leaders miss. AI handles short, well-bounded tasks far more reliably than long, judgment-heavy ones. If a process takes 45 minutes and includes three “it depends” branches and a phone call, the ROI on automating it is much lower than automating a two-minute task that runs 400 times a week. For the first two years of an AI program, pick the two-minute task.
Two-of-three is the threshold for putting a task on your candidate list. Three-of-three is a strong candidate for a pilot before the end of the quarter.
Three examples from credit union operations
Concrete beats abstract. Here are three back-office workflows that pass the test at most institutions in the $250M to $5B asset range. These are composite illustrations drawn from common operations patterns, not named case studies, so check the details against your own shop.
Member dispute intake
A member disputes a debit card transaction. A rep opens a case, captures the merchant, amount, date, dispute reason code, and any notes the member provides. The rep then routes the case to the disputes queue.
- Signal 1: yes. Every day; dozens of times at larger institutions.
- Signal 2: yes. The intake fields are the same in every branch and every call center. The Reg E clock does not care which employee filled the form.
- Signal 3: mostly yes. The intake itself is under five minutes. The downstream investigation is longer, but the intake step is short.
Score: three of three. Automation candidate: transcription plus structured extraction of the member’s account of the dispute, prefilled dispute forms, and automatic classification into merchant or fraud paths. A human still reviews and files. Good first-pilot material.
Lien release paperwork
An auto loan pays off. Someone in loan operations pulls the payoff report, verifies the balance is zero, and releases the lien: electronically through an ELT provider in electronic-title states, or by executing the release on the paper title and mailing it per that state’s titling process. Requirements vary widely by state, but every payoff carries similar data.
- Signal 1: yes. Runs on a schedule against the paid-off loan list.
- Signal 2: yes. If your ops team is doing this consistently, two people in two locations handle it identically.
- Signal 3: yes. Under five minutes per loan once the process is clean.
Score: three of three. Automation candidate: rules-based workflow with document generation, plus a light AI layer to handle exceptions (missing VIN, wrong state format, payoff with a hold flag). In ELT states, full elimination of the manual step is realistic; in paper-title states the win is prefilled documents and a clean exception queue. Either way, the human role becomes exception review. If you already run an ELT platform, point the automation at exceptions and the paper-state workflow instead.
ACH exception review
Nightly ACH exceptions land in a queue: NSF, closed account, invalid account, stop payment. A staff member walks the queue, resolves each item, and posts a return or a manual override.
- Signal 1: yes. Every business night.
- Signal 2: depends. Some CUs have one person in one office handling this end to end; if that is you, this signal is neutral. If two operators in two shops both handle it, they usually do it the same way.
- Signal 3: yes. Per item, well under five minutes.
Score: two or three of three. Automation candidate: rules-plus-classifier that clears the routine majority (matching patterns to prior decisions) and routes the ambiguous remainder to a human. A quiet efficiency win, and one that leaves a cleaner audit trail than the current queue.
Run this exercise next week
Block 60 minutes. Invite the head of operations, the head of member services, and one senior person from lending. Bring a whiteboard or a shared doc.
- List the top 20 tasks each of those functions handles in a normal week. Not projects. Tasks.
- For each task, mark yes or no on each of the three signals.
- Any task with two or more yeses moves to the candidate list.
- On the candidate list, rank by weekly volume multiplied by average time per instance. The top three are your first pilots.
- Stop the meeting. Do not evaluate vendors in this meeting. Vendor selection is a separate exercise once you know what you are actually solving.
A team running this for the first time usually walks out with more candidates than it can pilot, often a dozen or more. That is fine. Most credit unions do not have too few AI ideas; they have too many, ranked badly.
Why the signals matter more than the vendor
Every AI vendor demo will show you a workflow that looks impressive, and impressive tells you nothing about value to your institution. A vendor showing a call center summarization tool cannot tell you whether your call center’s after-call work meets the three signals in your operation. Only your team can tell you that. Once you know, the vendor conversation gets short and useful: “Here is the task, here is the volume, here is the current human time. What does your tool do with this exact scope?”
This is also the compliance leader’s fastest way into a productive AI conversation. NCUA’s existing supervisory framework covers third-party due diligence, information security, and internal controls whether the tool is AI-enabled or not. Starting with concrete, low-risk, short-cycle workflows means the third-party review and model risk conversations happen on a small scope first, rather than on a member-facing chatbot with everything to lose. A small, documented, low-risk scope is far easier to defend when examiners review your third-party risk and information security files, and it keeps early missteps from turning into exam findings. The NCUA compliance pillar covers the supervisory expectations in depth.
For a deeper walk through the pillar of back-office AI opportunities, and the sequencing rules that follow the three-signals screen, see the back-office automation pillar. This piece also seeded the weekly roundup with a live-example annex.
If the shortlist is longer than your team can execute
Most CU operations teams walk out of this exercise with more candidates than pilot capacity. That is the good problem. Prioritization is easier when the shortlist is honest.
If the exercise surfaces more candidates than your team can prioritize before Q4, Advisor Labs runs a 45-minute AI readiness audit that ranks candidates by volume, complexity, third-party risk, and time-to-value, and produces a written shortlist you can hand to your ops leader on Monday morning. Book the audit here.
For a broader view of how Advisor Labs works with credit unions, see Advisor Labs’ credit union practice.
The vendors will keep calling either way. Over the next 18 months, the credit unions best positioned to get real ROI out of AI are the ones that did the boring 60-minute meeting before taking the demos.