A Credit Union Cut Chat Volume 67% With Agentic AI. Here's the Detail That Actually Matters
Case studies in AI vendor marketing tend to blur together after a while, so it's worth pausing on one that just came out of KeyPoint Credit Union, a member-owned institution with $1.6 billion in assets serving more than 55,000 members. The numbers are real and worth stating plainly: after replacing its older, intent-based chatbot with an agentic AI platform from Eltropy, KeyPoint is now resolving 94% of member inquiries through AI, cut chat volume by 67%, and lifted voice service levels from 75% to 87% in the platform's first month of expanded use.
Those are strong results. But the more useful part of this story for another credit union isn't the percentage improvements, it's the two decisions underneath them.
The tool understands context instead of following a script
The system KeyPoint replaced was intent-based, meaning it matched a member's question to a fixed set of pre-written responses. What it deployed instead pulls answers directly from KeyPoint's own knowledge sources in real time and can handle both authenticated and unauthenticated inquiries, including fairly involved requests like fraud reporting, not just simple ones like a balance check. That distinction, between matching a script and actually retrieving current, accurate information, is the difference between a chatbot that frustrates members into asking for a human and one that quietly closes the loop.
They didn't stop at the member-facing win
The detail worth paying closest attention to: KeyPoint didn't treat this as a single deployment and move on. It followed the voice and chat rollout with an internal knowledge base connecting to its existing SharePoint environment, letting staff pull up answers to internal questions (Zelle limits, ATM dispute steps, check hold policy) through natural language instead of pinging other departments over chat and waiting. According to KeyPoint's VP of operations experience and payments, Josh Herzog, that's meant to reduce the moments where a staff member has to put a member on hold while they track down an answer internally, which is a second-order improvement to member experience that doesn't show up in the headline chat and voice statistics at all.
The unglamorous part: someone is reviewing the misses
Herzog also described an ongoing process of reviewing the questions the AI couldn't answer, asking specifically whether the information simply wasn't available to it yet and fixing that gap. That's the least exciting part of this story and probably the most important one. An agentic AI deployment that goes live and gets left alone tends to degrade quietly as member needs shift. One that has someone assigned to regularly close the gaps it can't yet handle tends to keep improving instead.
What this means for your credit union
If you're evaluating a similar deployment, the vendor's resolution-rate claims are worth checking, but the more predictive questions are about your own side of the deployment: who is going to own reviewing what the AI couldn't answer, and are you planning to stop at the member-facing tool or also point the same technology at internal staff questions, where a lot of the quieter efficiency gains tend to live.
Sources