Vendor announcements are easy to find. Deployments with a named institution, a named workflow, and a number a reader can check against a public source are rarer, so when our screening turned up three strong ones this quarter we wrote up three rather than padding to five. Same format as our last case study review: who, what workflow, what tool, what changed, and the caveat to apply before repeating it.

1. Rockland FCU: indirect auto loan post-closing QC

Institution: $3.8 billion, headquartered in Rockland, Massachusetts, rebranding as Arise Financial. Workflow: post-closing quality control on indirect auto loans. Tool: Kintera AI, which converted the credit union’s existing standard operating procedures into an automated workflow with no core integration or custom code. Outcome: per the deployment announcement, live in 10 days, per-file review time down from 20 minutes to 2, every file now fully cross-referenced instead of sampled, and $250,000 in documented annual savings. Caveat: it is a vendor press release describing the vendor’s first production deployment, so the numbers are self-reported and the platform’s track record is exactly one credit union long.

2. KeyPoint: member chat and voice

Institution: $1.6 billion in assets, 55,000 members, Silicon Valley. Workflow: member inquiry resolution across chat and phone. Tool: Eltropy’s agentic AI platform, replacing the scripted intent-based chatbot KeyPoint had run for years. Outcome: per the announced results, 94 percent of member inquiries resolved through AI, chat volume down 67 percent, and voice service levels up from 75 to 87 percent in the first month. Caveat: first-month figures announced by the vendor, and a six-year existing relationship means this was an expansion with unusual integration depth, not a cold start.

3. PenFed: contact center documentation

Institution: PenFed Credit Union. Workflow: member conversation documentation and after-call work, via a real-time assist the credit union names Agent Wingman: the system transcribes the conversation, a human agent approves the content, and the record updates before the next interaction. Tool: Salesforce Agentforce, part of a consolidation from roughly 400 platforms to 12 strategic partners. Outcome: per CEO James Schenck’s remarks on Salesforce’s earnings call, $1.6 million in annual savings, handle time down 10 percent, after-call work down 50 percent, and held volume down 40 percent, with 76 AI agents now running across operations, mortgages, IT, and HR. Caveat: the numbers came from the customer’s CEO on the vendor’s earnings call, which is about as friendly a venue as a statistic can have.

The pattern

None of these three automated the judgment call. Rockland automated the checking around a loan that a human already approved, KeyPoint automated answers that pull from its own knowledge base with escalation to staff, and PenFed automated the documentation around a conversation while keeping a human approval step before anything enters the member record. That placement, at the work surrounding the decision rather than the decision itself, is what kept each deployment in the lighter control tiers we mapped in this week’s model risk piece, and it is why all three shipped in weeks instead of quarters. Run your own candidates through the Three-Signals Test and the same shape falls out.

Get this roundup in your inbox every Thursday: subscribe to the AiForCU newsletter.