Workshop length: Half-day

Presenters:

Eduardo Bunge Ph.D (Palo Alto University, Co-Founder of ParenteAI)

Juan Dellarroquelle (CTO and Co-Founder at Parente AI)

Description

Recent advances in Large Language Models (LLMs) have created new opportunities for digital mental health interventions. However, many clinicians, researchers, and developers remain uncertain about how to safely and effectively integrate LLM-powered conversational agents (CAs) into clinical practice.

This hands-on workshop is designed to bridge this gap by providing a practical introduction to the capabilities and limitations of LLMs in mental health. Participants will explore the evolution of conversational agents, comparing rule-based systems to generative artificial intelligence (GenAI) models, and reviewing existing research on their effectiveness, including treatment outcomes, therapeutic alliance, engagement, and attrition.

We will then shift to applied learning, introducing the fundamentals of LLM architecture and the key components necessary for building AI-driven chatbots tailored to digital mental health interventions. Topics will include prompt engineering, guardrails for safety, few-shot learning, multi-agent approaches, and fine-tuning for specific populations.

A focal point of the session will be the concept of an AI co-therapist developed for parent management training programs. Through interactive exercises, participants will gain hands-on experience in structuring CA dialogues, implementing ethical safeguards, and how to integrate them into clinical practice (see figure 1).

By the end of this workshop, attendees will have a foundational understanding of LLM-powered CA, along with practical strategies for safely integrating them into research and clinical workflows.

This session is ideal for clinicians, researchers, and digital mental health innovators seeking to harness AI for therapeutic applications.

*Registered participants are requested to bring their own laptop to be able to work on relevant materials during the workshop.

Key Learning objectives: 

1. Understand the evolution of conversational agents in mental health, distinguishing between rule-based and LLM-powered models.
2. Evaluate the evidence base for AI-driven mental health interventions, including treatment outcomes, engagement, and therapeutic alliance.
3. Learn how LLMs work and explore different techniques for using them for specific mental health applications.
4. Implement safety measures such as guardrails, and multi-agent architectures to ensure ethical and effective AI deployment.
5. Apply practical design strategies to develop or integrate AI chatbots into their research or clinical practice, using an AI co-therapist for parent management training as a case study.

Workshop Structure (3–3.5 hours):

1. Introduction to Conversational Agents (30 min)

• History and evolution of chatbots in mental health
• Comparing rule-based vs. LLM-based systems
• Research findings on engagement, alliance, and attrition

2. How LLMs Work (45 min)

• Basics of generative AI
• Strengths and limitations in mental health applications

3. Designing Safe and Effective AI Chatbots (45 min)

• Guardrails and ethical considerations
• Prompt engineering, Retrieval augmented generation, and Multi-agent approaches

4. Hands-on Session: AI Co-Therapist for Parent Training (60 min)

• Case study: An AI co-therapist supporting evidence-based parenting programs
• Interactive exercise: testing a structured dialogue flow
• Group discussion on implementation challenges and opportunities

5. Q&A and Next Steps (30 min)

• Discussion of practical applications for attendees’ work
• Resources for continued learning and collaboration.

Figure 1. Example of ParenteAI’s conceptual framework