Every Monday morning, support teams at one of the largest credit unions in the western U.S., serving 750,000+ customers across four states, began their week digging out of a backlog of hundreds of unanswered weekend voicemails. A previous attempt at voice automation had failed, trapping callers in endless loops and leaving the team of 300 agents to deal with the fallout. This failure made executives highly skeptical of technology promises, but the constant operational strain meant they still had to find a secure, reliable solution. To eliminate operational bottlenecks and regain member trust, the credit union turned to Level AI to unify quality assurance, conversational intelligence, and virtual assistants on a single platform.
A broken legacy pilot left customers stranded & plunged operations into a technology standstill
The institution previously tried to automate voice operations with an external communications vendor, Eltropy. While the vendor's chat tool functioned, the voice software could not manage the weight of real customer conversations. The system failed to interpret natural speech, dropped active calls, and trapped callers in repetitive loops. Internally, leadership described the voice product as “an illusion built on smoke and mirrors”. The failure damaged the brand's reputation, leaving customers without after-hours support and forcing shift leaders to start every Monday morning facing backlogs of hundreds of unanswered voicemails.
On live calls, agents experienced high stress while switching between Five9 for phone controls, Salesforce for customer records, and a shared directory for policies. This manual search process drove up hold times and frustrated callers. Meanwhile, the quality assurance team could only monitor 2% of calls, leaving 98% of interactions entirely invisible to coaching or compliance audits.
The failed deployment left deep institutional scars. The executive team felt burned by the previous vendor, making them deeply skeptical of automated systems and reluctant to commit budget to another platform. This lack of trust stalled the procurement process for four years as leadership waited for reference customers, live deployments, and evidence that voice technology could operate securely in a regulated financial environment.
The executive team required proprietary models & mock integrations to restore faith in automation
To move forward, the organization required an architecture that integrated with its existing Five9 telephony system, Salesforce platform, and Pindrop security suite. Leadership explicitly rejected generic software vendors that wrapped third-party models.
According to the VP of Member Service Centers, “Generic software companies pass sensitive conversational data to external public servers, creating compliance and security vulnerabilities under NCUA guidelines.” They also struggle to recognize complex financial terms, resulting in high transaction error rates. The lender needed a partner to run proprietary, localized models that kept data secure and recognized specific banking intents with precision.
To prove the system could handle live customers, developers built a testing environment using simulated responses during the POC. Level AI called into these test environments to run mock conversations, verifying speech recognition and decision logic on fake balance checks and address updates before touching live systems.
This verification process restored institutional trust. The integration moved forward because the cooperative's leaders wanted a partner with capabilities across the entire call lifecycle, specifically seeking tools to assist the agent during the conversation and automate quality management after the call. The internal development team noted that “we knew we were moving forward with this project, so we built our endpoints ahead of schedule to be ready on day one.” Level AI's engineering team mocked those APIs to test live conversational flows, compressing the development cycle and ensuring the platform was ready to run.
Level AI’s unified architecture & domain-specific LLMs restored faith & removed manual friction
The credit union focused the POC on three functional areas: after-hours call routing, in-call agent support, and quality scoring. The deployment integrated natively with the credit union's existing infrastructure, which included Five9, Salesforce, Calabrio, and Pindrop for security and voice credential verification.
1. Connecting the phone system to database endpoints for automated calls
Level AI integrated with Five9 using a direct phone connection called a SIP transfer. When a call comes in outside normal business hours, the system routes it to Level AI's Virtual Agent. Virtual Agent resolves calls using a three-tiered approach:
Tier 1 (Public Information): For basic questions such as branch hours or loan rates, the system retrieves text directly from the credit union's public website. No customer login is required.
Tier 2 (Account Details): For balance checks or recent deposit history, the system asks for the customer's number and date of birth. Once verified via secure database queries, it pulls the data from Salesforce.
Tier 3 (Transactions): For actions such as mailing a replacement debit card or updating an address, the system sends a one-time text message code to the customer's phone. Once the customer speaks or types the code, the system writes the update straight to the credit union's core system.
To protect customer identity, Level AI does not store any personally identifiable information (PII). All verification checks pass through the credit union's secure APIs, which only return a simple pass or fail response to the virtual agent.
2. Providing direct policy answers on the agent's screen during live calls
For calls requiring human help, Level AI's Agent Assist runs inside an iframe directly in the Salesforce window. The software transcribes active phone conversations in real time, identifies the customer's needs, and displays matching policy documents from SharePoint on the agent's screen. If the caller asks about fee waivers, the correct waiver guidelines appear automatically. Level AI’s Agent Assist also has a checklist system that tracks whether the agent completed required identity checks and greeting scripts, checking them off as they speak.
3. Reviewing every single phone call, chat, and email transcript automatically
Rather than forcing managers to select and listen to random audio files, Level AI Auto-QA reviews 100% of customer interactions. It evaluates voice and chat transcripts against the credit union's rubric, grading agents on greetings, identity verification, issue resolution, and compliance. Level AI LLMs match written transcripts to recordings of the agent's desktop screen, showing supervisors exactly where agents struggled to find documents or enter account notes.
100% auditing and instant self-service deliver immediate operational relief without expanding headcount
Level AI eliminated the manual spot-checking bottleneck through this POC by automating quality management across 100% of voice, chat, and email channels. Supervisors didn’t have to spend hours hunting for random audio files or basing performance reviews on samples. Instead, they use automated scores to target coaching, address compliance gaps, and prevent regulatory violations across every conversation.
The virtual assistant managed 100K interactions during the POC within a month, answering common questions and processing card replacements in the evenings, on weekends, and during holidays.
As mentioned by one of the product owners, "We wanted to see proof that the voice technology actually worked under real pressure. Level AI validated their system in our test environment first, which gave us the confidence to automate our after-hours calls." This continuous coverage resolved the after-hours availability gap, keeping Monday morning voicemail queues empty and shielding staff from the weekly burnout of backlog recovery.
Inside the contact center, Level AI’s Agent Assist drives down average handle times by delivering policy answers directly to the agent's screen, eliminating manual search friction.
Following the successful proof of concept, the regional cooperative chose Level AI for its secure, proprietary architecture, a decision validated by the results. By grounding service automation in verified conversation data, the lender is eliminating the gap between leadership expectations and customer reality, establishing a new baseline for regional banking operations.
