When Compass Working Capital’s clients call, they are usually in the middle of a problem that has been unfolding for months or years. A missed rent payment. A credit score that will not move. A tax refund that never arrived. So when a client is routed to a new financial coach, something common in a program that lasts three to five years, the conversation begins with repetition as the client explains the rent notice again, or a job change again. All the knowledge and understanding built by the last coach doesn’t always make the handover.

That knowledge gap is the real challenge facing many customer-service AI projects, even when all the attention focuses on chatbots and voice assistants. Compass, a nonprofit that provides long-term financial coaching, knows the most important feature of an AI contact center is not necessarily the agentic voice assistant on the front end. It is the knowledge transfer to that assistant, so when the client calls next, the agent already knows their situation.

When Knowledge Doesn’t Transfer

Compass clients enroll for what could be years of savings and credit coaching, and staff turnover means they often switch coaches along the way. Each change forces the client to rebuild context with someone new. The frustration is relational, not technological. A client who has explained a medical debt or a custody expense to one coach must explain it all again.

This is not a problem unique to nonprofits. Customer-experience firm Genesys found that 95% of consumers expect information to be remembered across channels, while 48% of organizations fail to pass information automatically between virtual and human agents. Deloitte’s banking research from June 2026 reported the same pattern: fragmented case histories and customers forced to repeat explanations after transfers.

For Compass, the risk of an AI project was that the technology would answer the question and then switch the client into another cold handoff, adding automation to an already impersonal transition.

Keeping the Human at the Center

Compass began looking for technology that would expand the work its coaches do, not replace them, according to George Reuter, the nonprofit’s managing director of impact and innovation. “We were looking for a way to expand our reach without compromising the individualized support that makes our model effective.”

The organization chose Agentforce Contact Center, Salesforce’s AI contact-center platform, implemented with TTEC Digital, a customer-experience technology and services company. Compass was the first Salesforce customer to go live on the platform, according to TTEC Digital.

The client-facing part of the system is a voice assistant grounded in Compass’s existing knowledge base, designed to handle inbound questions and schedule appointments. Compass also loaded more than 100 resources on tax preparation, credit improvement and other topics into Salesforce, allowing coaches to access relevant information during calls. The back end processes voice transcripts and drafts structured data for coaches to review, edit and submit. Compass estimates the deployment will save 6,000 staff hours annually.

But Reuter emphasized that the goal is not automation for its own sake.

“We’re not using Agentforce to replace coaches or to rush clients through the system,” Reuter explained in a write-up describing the implementation. Instead, the intent is to remove friction so coaches can be more present and clients can experience smoother transitions.

The Handoff Is the Magic

The design decision that most often trips up AI contact-center projects, said Marc Hutchinson, vice president of solution architecture at TTEC Digital, is the handoff from the voice assistant to a human when the AI receives a question it cannot or should not answer.

Many deployments treat that handoff as an afterthought, something to configure later or buy as an add-on. Hutchinson contends that is a mistake.

“Organizations shouldn’t enter into an engagement that views the handoff as an optional add-on, or as an additional cost that needs to be addressed separately,” he said. “That slows you down and it limits how smart your AI can become long term.”

The handoff is not a feature bolted onto the side of a voice assistant. It is the architecture. If the AI cannot move a client smoothly into the hands of a coach who already understands the case, the project fails regardless of how well the bot speaks.

The speed of the Compass rollout, six weeks from kickoff to production, also came in part from that design-first approach, Hutchinson said. Compass was already a longtime Salesforce customer, so integration with existing business data and processes was faster than building a standalone voice system and trying to connect it afterward.

Most of the work, Hutchinson said, is in the details. That includes curating knowledge articles, tuning prompts, teaching the system to read phone numbers the way a human would and activating employee-facing automation.

The system connects caller information, a trained voice assistant that can take actions inside Salesforce and the underlying client records. In minutes, a caller can be matched to a profile. The harder part is making sure the next human who picks up the call knows what to do with it.

A Knowledge Base That Learns

Compass built its knowledge base from more than 100 internal resources, but regulations, processes and edge cases continue to evolve.

Voice transcripts offer one way to keep that knowledge base current. A conversation might reveal a changed local regulation or a process step that was never documented. Such information could be flagged for human review and, where appropriate, added to the knowledge base.

“That’s information the AI can suddenly capture without anybody lifting a finger,” Hutchinson said. “It just picks it up out of the transcript and flags it.”

The aim is a system that improves while it runs, instead of a static knowledge base that goes stale.

The early lesson is the design choice: use software to remove the administrative weight from human relationships, not to remove the people. Too often, contact-center AI projects prioritize automation over the continuity of those relationships.

Compass went from kickoff to production in six weeks. Its clients, who kept losing coaches and restarting conversations, did not need a better chatbot. They needed the next human who answered the phone to know their stories.