Case studies/Thriving Center of Psychology

Giving the support team a policy base they could actually retrieve from

Company:
Thriving Center of Psychology
Years:
2024
Role:
Head of Product
Stack:
Claude · AWS · Vector store
Faster response drafting
coordinators drafted policy-grounded responses in minutes instead of hours
Reduced escalations
consistent policy retrieval reduced the rate of inconsistent answers that triggered callbacks

Thriving Center of Psychology 's customer service team handled hundreds of patient and client navigator inquiries per week. A meaningful share required policy lookup — billing questions, coverage rules, cancellation policies, provider credentialing requirements. The team answered them from memory, which meant inconsistent answers, frequent escalations, and coordinators spending hours tracking down the right document.

The Challenge

A support team answering policy questions from memory

TCP had grown fast, and the policy documentation had grown with it — spread across Notion pages, Google Docs, and internal wikis that nobody had time to reconcile. When a patient navigator needed to answer a billing question, they'd ask a colleague, search Notion, or guess. The inconsistency was visible in the escalation rate.

The answer wasn't better documentation. The team had seen too many documentation initiatives fade. The answer was making the existing documentation retrievable in the moment a question appeared.

The solution

A RAG system grounded in TCP's actual policy corpus

We built a retrieval-augmented generation system that indexed TCP's policy documents, knowledge base articles, and internal SOPs — and gave the CS team a way to query it directly inside their ticketing workflow.

  1. 01

    Ingest what exists — don't rewrite it first.

    The temptation was to clean up the docs before ingesting them. We skipped that. We ingested the existing corpus as-is, which meant the system was useful on day one. Curation happened over time as the team flagged gaps and inconsistencies the retrieval surfaced.

  2. 02

    Return source context alongside the answer.

    Coordinators didn't trust responses without a citation. We surfaced the source document and passage with every retrieval — so the coordinator could verify, and so the system built credibility gradually rather than asking for blind trust upfront.

  3. 03

    Integrate into the existing ticket workflow, not a new tool.

    A new tool requires a behavior change. We integrated the retrieval interface into the workflow the CS team already used, so using it was the path of least resistance rather than an extra step.

The outcome

Policy-grounded responses, faster drafting, fewer escalations

The CS team went from searching Notion and asking colleagues to getting a retrieved, cited response draft in the time it took to type a question. Response consistency improved. The escalation rate for policy-related tickets dropped as coordinators started answering from the same source instead of from memory.

An unexpected outcome: the retrieval gaps the team flagged became the most useful documentation backlog TCP had ever had. The system surfaced what was missing more efficiently than any documentation audit.

One thing I'd do differently

Source quality matters more than retrieval quality. We spent time tuning the retrieval and not enough time on the underlying docs. When the policy document was well-written and specific, the retrieved answer was good. When it was vague or outdated, no amount of retrieval tuning helped. I'd build a doc quality review step into the next ingestion process.