Gary Jeter had more than 600 policies that nobody trusted and a freshly merged credit union that needed them yesterday.
Jeter is the CTO of TruStone Financial, which in 2021 completed what was then the largest credit union merger of equals in the country when it combined with Firefly Credit Union. The merged institution has roughly $3.4 billion in assets across 23 locations in Minnesota and Wisconsin. The tech integration alone required mapping 53 applications to a single core banking system, with a team of 52 people running six rounds of manual data validation before the systems went live.
The systems got sorted. The operational mess did not. Two organizations' worth of internal policies and procedures, some dating back to 2019, sat scattered across an intranet that frontline staff had mostly stopped using. When a teller wasn't sure how to handle a transaction, they'd send a question to the back office and wait. Average response time was 15 minutes. Sometimes the answer never came back at all.
The standard playbook here is well established. Audit every document, reconcile the contradictions between the two organizations, update what's outdated, and then think about deploying AI on top of it. Jeter skipped all of it.
In October 2024, his team loaded every one of those policies into Senso.ai, a Toronto-based platform built for credit unions that had come out of Y Combinator's winter 2024 batch earlier that year. It works like a ChatGPT for internal knowledge, where employees type a question and get an answer pulled from whatever documents you've fed the system. Jeter fed it everything.
"We took a novel approach," Jeter told CIO. "We're not going to worry about data accuracy."
The logic was that bad answers would surface bad documents faster than any formal audit could. When an employee got a wrong answer, they'd hit thumbs down. That flagged the source document, and a dedicated staffer would work with the relevant department to update it. The corrected version went back in and the AI started pulling from it almost immediately. Every bad answer was a free audit finding.
Jeter started with 10 employees. Then 50. Then a full region of about 200. During that phased rollout, the system's accuracy climbed from 23% to roughly 75% based on employee feedback. By the time it went company-wide, it hit 92%. Frontline response time dropped from 15 minutes to under three.
Getting people to actually use it was its own challenge. VP of IT Mayka Thao, who oversaw the rollout, said her team ran focus groups and a phased proof-of-value process specifically to build trust before going company-wide. A generative AI training course they built around the tool became the most popular class in the entire organization.
TruStone isn't alone here. Six of the 10 largest credit unions in the country now run on Clutch's AI platforms, and 81% of financial firms globally are bringing AI into operations in some capacity. But regulators are watching closely. The UK's Financial Conduct Authority recently warned that widespread AI adoption across financial services risks creating "correlated behavior" and "common points of failure" if too many institutions end up relying on the same handful of providers.
Jeter's team is already pushing further. They're building an AI-powered chatbot for TruStone's website that can pull updated information like interest rates in real time, replacing a rules-based system. They're exploring AI agents for fraud detection outside business hours since their fraud team only works during the day. And at a recent hackathon, one of their developers used the Gemini API to build a fraud detection prototype in two days.
The enterprise AI consulting industry has spent years selling data readiness as step one. Clean your data, build your taxonomy, hire a data governance lead, then deploy. Jeter flipped the order and let 350 employees do the cleanup for him without even realizing it, just by asking the questions they were already asking every day. The 23% to 92% jump happened faster than most companies finish their assessment phase. If you're sitting on a messy internal wiki or a Confluence graveyard nobody trusts, the lesson from a Minnesota credit union is pretty straightforward. Your best QA team is the one already asking questions. You just need to give them a thumbs-down button.
