Demonstrating Safe, Clinically‑Governed AI Triage: An STCC - VStorm Case Study

Executive Summary

Schmitt-Thompson Clinical Content (STCC) and Thompson Adult Guidelines (TAG) have partnered with VStorm to build and test a Proof of Concept (PoC) AI‑powered triage chatbot. The goal is to demonstrate that STCC’s existing telehealth triage guidelines as currently structured can be used safely and effectively by generative AI to support efficient patient intake and generate a correct disposition.

STCC understands and actively promotes the value of human triage nurses but also believes that generative AI can be used to make nurse triage safer and more efficient. This project has focused on building workflows that integrate smoothly with nurse triage while giving patients a clear and easy‑to‑navigate experience.

The PoC uses nurse-authored, nurse-validated standardized clinical scenarios and a dual-track test design: a full-scenario mode for controlled benchmarking, and a conversational mode for realistic information gathering. Across both tracks, the project measures metrics that clinical leaders value most: disposition accuracy against the validated scenarios, time to disposition, question burden, rationale transparency, and exportable audit logs.

Problem Statement

Telehealth nurse triage teams face increasing demand, growing documentation burden, and staffing pressures. In STCC’s 2023 benchmarking survey, over 40% of respondents reported average triage call time of 11 to 16 minutes, and 15% reported calls taking 20 minutes or longer. Many respondents reported that call duration rose over the prior three years.

Healthcare leaders are exploring generative AI to address these challenges, but triage is a high‑stakes workflow that demands proven safety and robust clinical governance. Health systems and software companies need evidence that AI tools to support nurse triage can stay inside a governed clinical structure, support safe performance without under- or over-referral, and remain auditable for quality and risk review.

STCC Objectives

  1. Prove the system is safe and reliable: Demonstrate with a controlled PoC that the AI can follow the STCC triage structure without reducing patient safety.

  2. Follow governance rules: Build a validation process that ensures safe, repeatable testing results.

  3. Measure real operational impact: Track how long it takes to reach a disposition, how many questions the system asks, and how accurate the dispositions are across different scenarios. Reduce redundant steps and streamline the triage process.

  4. Protect Intellectual Property (IP): Keep STCC triage guidelines and encounter data out of AI training.

  5. Capture lessons learned:Use an iterative approach for development and testing. Document what worked and what did not work. Identify what should change going forward.

VStorm Core Delivery Objectives

  1. Deliver a fast, reliable PoC: Build a working PoC that runs reliably, follows sound engineering practices, and produces results that can be measured and trusted.

  2. Enable clear audit trails: Log all decision points so deviation from expected guideline logic can be explained.

  3. Support controlled comparisons: Include configuration options that make it easy to run meaningful A/B (side‑by‑side) tests across models, prompts, or settings.

  4. Create a reusable playbook: Produce a repeatable approach that can be applied to future clinical AI projects.

Approach

The teams have approached this PoC as a safety and evidence project, not just a demo.

  • STCC is providing product direction and has set the clinical and governance requirements.

  • Nurse triage subject‑matter experts have established quantitative validation benchmarks using curated clinical scenarios. Nurse triage experts confirm whether the system meets these benchmarks by assigning safe and correct dispositions.

  • VStorm is building the workflow, connecting the AI model, and creating the reports and dashboards.

Dual-Track Testing

  • Track 1 - Full Scenario Input: A complete patient scenario is provided in one pass. The system must identify the primary reason for visit, select the correct guideline, identify positive and negative Triage Assessment Questions (TAQs), and assign a correct disposition.

  • Track 2 - Conversational Intake: The system must collect enough structured information through back-and-forth chat to avoid “insufficient information”. The system must identify the reason for visit, select the correct guideline, identify positive and negative (TAQs), limit total number of questions, and assign a correct disposition. The conversational flow should be fluid, natural, and efficient for the user.

Closing Thoughts

Many AI triage projects run into predictable problems: unclear inputs, decisions that are hard to explain, and weak documentation jeopardizing governance and quality review. This ongoing collaboration directly addresses those issues by pairing clinically respected content and validation with an engineering team focused on efficiency, transparency, and disciplined configuration.

Authors

Konrad Budek, Growth Manager
VStorm

This partnership works because roles are clear. Clinicians owned the gold standard scenarios; Vstorm owned delivery; STCC / TAG owned product governance.

Laurie O’Bryan, RN, Nurse Editor
Schmitt-Thompson Clinical Content, Thompson Adult Guidelines (TAG)

Nothing can replace telehealth triage nurses. But we can reduce friction in intake and documentation so nurses can focus on judgment and patient safety.

Matthew Thompson, MBA, CSPO, Product Manager
Schmitt-Thompson Clinical Content and Thompson Adult Guidelines (TAG)

A clinical AI pilot needs an audit trail, not just a demo. We built the PoC to produce exportable evidence on every run.

Contact Information

This is an ongoing research project. If you would like hear about future updates on this project, please follow STCC on LinkedIn.

‍