For most people in therapy, care is concentrated in a single weekly hour. What happens in the other 167 hours is largely left to the client. Increasingly, they are filling that space with AI. Among U.S. adults who use large language model (LLM) chatbots, nearly half report using them for mental health needs, and more than a third say they find them more helpful than traditional therapy [1]. Most are doing so outside of any clinical relationship, without provider knowledge or oversight.

But most of these tools were not designed for clinical use. They lack access to client history, operate without clinical guardrails, and are not connected to a broader system of care. When clients use them as a substitute for treatment, or as a between-session resource without provider involvement, there is no mechanism to ensure safety, accuracy, or alignment with their care.

Grow Therapy (Grow) takes a different approach: creating purpose-built AI tools that operate within the clinical relationship, with providers at the center. The model is guided by three core pillars: improving client outcomes, ensuring strong oversight and safety, and protecting the therapeutic relationship. It is grounded in clinical science and established AI governance frameworks. In practice, this includes tools like “coach,” Grow’s conversational AI tool that helps clients reflect, process challenges, and apply evidence-based skills between sessions while remaining connected to their provider. Early data show strong adoption and engagement, with nearly 80,000 clients using coach. Grow’s AI also supports providers through tools like “Session Insights,” which generates structured post-session summaries that providers can share directly with clients to reinforce continuity between visits, with more than 1.3 million generated to date.

Grow’s approach demonstrates how AI can be designed to operate within care, extending support beyond the session while maintaining clinical integrity, safety, and human connection.

More than one in five U.S. adults are in need of mental health support, but only about 50% receive care in a given year and over 120 million people live in mental health provider shortage areas [1]. That supply-demand mismatch helps explain why so many people are turning to general-purpose AI tools for support. But the fact that people find these tools accessible does not mean they are adequate. Most were not designed for clinical use, yet they are increasingly being used as substitutes for care.

This is not a hypothetical shift. People are already using AI to reflect, process emotions, and navigate challenges in real time, often without any clinical oversight. Whether these tools help or harm may depend less on the technology itself and more on how it is designed, who oversees it, and whether it is connected to a broader system of care.

AI can expand access to mental health support by offering immediate, accessible engagement and a space for reflection, particularly between therapy sessions. Research suggests that people may share more openly with conversational systems they perceive as nonjudgmental, and some studies have shown the potential for AI tools to support elements of the therapeutic alliance [2]. But these benefits depend on how AI is designed and used.

When AI is used for mental health outside of a clinical context, it introduces real safety risks. For example, recent reporting has documented severe adverse events tied to the use of general-purpose chatbots for mental health needs. In one widely covered case, a wrongful-death lawsuit alleges that a 14-year-old became intensely attached to a chatbot that reinforced suicidal ideation prior to his death [3].

Safety concerns extend beyond individual cases. Mental health experts have raised warnings about chatbot-induced delusions, reinforcement of harmful thinking patterns, and failure to recognize or appropriately respond to crisis situations [4, 5]. Policymakers have taken action, with several states moving to restrict or ban certain types of AI chatbots in mental health care [6, 7], yet many of the most widely used general-purpose chatbots remain largely outside the scope of existing regulation.

These risks reflect how the tools are built. General-purpose AI systems were not designed for clinical mental health care. They lack access to client history, operate without accountability structures, and are not embedded within a broader system of care. How well AI performs in clinical settings depends inlarge part on whether it has access to that context and whether a provider is positioned to act on what it surfaces.

It is worth noting that the landscape is evolving. Some general-purpose tools now allow users to upload medical records, and many companies have begun building internal accountability structures. However, Grow’s position is that AI is most safely and effectively applied when it operates within a care delivery network where a licensed provider can curate its use, stay informed by its interactions, and remain in the driver’s seat of the clinical experience.

This does not mean AI cannot play a role in mental health care, but rather that its role must be intentionally designed and connected to existing structures that deliver quality and effective care.

Grow was founded to offer therapy that’s accessible, effective, and grounded in meaningful human connection. Our AI strategy is central to achieving these goals, developed in close collaboration with clinicians, providers, and advisors rather than in isolation.The approach is grounded in both clinical science and established AI governance frameworks (see Table 1; [8, 9]). It draws on research showing the importance of therapeutic alliance and between-session engagement in driving meaningful client outcomes, and aligns with guidance from organizations such as the World Health Organization (WHO), the National Institute of Standards and Technology (NIST), and the Office of the National Coordinator for Health IT (ONC). These frameworks consistently emphasize the importance of human oversight, safety, and continuous evaluation.

Alignment of Grow Therapy’s AI approach with established frameworks

Our core pillars

Grow’s AI principles are guided by three core pillars: 1) Client Outcomes, 2) Strong Oversight, and 3) Protecting the Therapeutic Relationship. 

At Grow, AI is embedded into workflows that support both clients and providers across the course of care.

Extending Care Between Sessions: Coach in Practice

Coach is Grow’s conversational AI tool designed to address that gap by helping clients reflect, process challenges, and practice skills between sessions. It is not a standalone product but designed to complement the clinical relationship. Clients can use coach to work through situations in real time, reinforce evidence-based skills introduced in therapy, and capture topics to bring back to their provider. Providers can review coach interactions, use them to inform session priorities, and decide whether coach is appropriate for a given client. Clinical judgment and decision-making authority remain with the treating provider throughout.

This design reflects a deliberate clinical philosophy. Coach does not diagnose, provide treatment, or position itself as a therapeutic substitute. It operates within boundaries that are explicit to users from the start and reinforced throughout. When potential safety concerns are detected, interactions are paused, crisis resources are surfaced, and providers are notified. Grow’s internal clinical team reviews a portion of all interactions, alongside an automated evaluation system that assesses clinical quality and safety at scale. These oversight mechanisms are integrated into how the product operates rather than layered on after the fact.

Early data suggest the model is working. In its first five months, nearly 80,000 clients have used coach. Language patterns suggest clients feel better by the end in a majority of interactions. These are early signals, not clinical proof, but they are consistent with the hypothesis that between-session support, when integrated into a care relationship rather than offered in isolation, can reinforce rather than replace the work clients do with their therapist.

What these numbers reflect, more than adoption, is that providers and clients are using coach the way it was designed: as part of overall care.

Supporting Providers: AI in the Care Workflow

Coach supports clients in the space between sessions. Grow’s provider-facing AI tools are designed to work alongside providers throughout the care workflow, not simply to reduce administrative work, but to surface clinical insight, support continuity of care, and keep providers focused on what only they can do.

Grow’s AI notetaker is designed to reduce that burden at multiple points in the care workflow. During sessions, it captures key themes and details so that providers can stay focused on the clinical interaction rather than splitting attention between listening and note-taking. After sessions, it generates structured summaries and next steps, referred to as “Session Insights,” which providers can review and share with clients to reinforce progress and maintain continuity between visits. Over time, these tools also help surface patterns in how clients are progressing, giving providers a more complete picture to inform clinical decision-making.

These tools are already integrated at scale across Grow’s provider network. More than 1.5M+ Session Insights have been delivered to clients, and providers report reductions of up to 40% in time spent on intake notes and up to 66% for progress note documentation. These tools reflect the same principle that guides Grow’s AI strategy more broadly: AI should reduce the work around care while keeping clinical judgment and decision-making authority with the provider.

AI is already part of how people manage their mental health. The question is not whether AI belongs in this space but how it should be built, governed, and integrated into care.

Grow’s approach offers a model: purpose-built AI that operates within the clinical relationship, with providers retaining decision-making authority and multiple layers of oversight working together to monitor safety and quality. Early data from coach and Grow’s provider-facing AI tools suggest this model is gaining traction with both clients and providers, though Grow treats these as early signals that warrant continued evaluation rather than proof of concept.

What generalizes beyond Grow is the underlying principle. AI in mental health care should be held to high standards of oversight, accountability, and continuous improvement. As the field evolves, the organizations that get this right will be the ones that treat responsible AI not as a set of principles on paper but as an operational commitment embedded in how their products work every day.

Grow believes independent experts serving as accountability partners are critical to setting and maintaining a high bar for responsible AI in mental health care. To support that commitment, the company established an external Advisory Panel to help inform Grow’s approach to responsible AI in mental health care more broadly. The panel brings together expertise spanning psychology, clinical AI research, mental health epidemiology, ethics, and consumer insights, providing interdisciplinary guidance grounded in both scientific rigor and compassion for the people Grow serves. Their counsel supports Grow’s approach to safety, evaluation, and continuous improvement as the AI coach evolves.

[1] Rousmaniere T, Zhang Y, Li X, et al. Large language models as mental health resources: Patterns of use in the United States. Pract Innov. Epub ahead of print 2025. DOI: 10.1037/pri0000292.

[2] Fitzpatrick KK, Darcy A, Vierhile M. Delivering Cognitive Behavior Therapy to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot): A Randomized Controlled Trial. JMIR Ment Health 2017; 4: e7785.

[3] Silicon Valley AI company sued over 14-year-old’s suicide. ABC7 San Francisco, https://abc7news.com/post/silicon-valley-based-company-character-ai-sued-florida-14-year-olds-suicide/15461471/ (2024, accessed 8 April 2026).

[4] Preliminary Report on Dangers of AI Chatbots | Psychiatric Times, https://www.psychiatrictimes.com/view/preliminary-report-on-dangers-of-ai-chatbots?utm_source=chatgpt.com (accessed 3 April 2026).

[5] “AI Psychosis”: How ChatGPT Amplifies Delusions & Triggers Psychosis. Psychiatry & Psychotherapy Podcast, https://www.psychiatrypodcast.com/psychiatry-psychotherapy-podcast/episode-253-ai-psychosis-emerging-cases-of-delusion-amplification-associated-with-chatgpt-and-llm-chatbot (accessed 3 April 2026).

[6] 10/03/2025: States struggle to regulate AI chatbots for mental health therapy amid rising need for care | News, https://www.healthpolicyohio.org/health-policy-news/2025/10/03/states-struggle-to-regulate-ai-chatbots-for-mental-health-therapy-amid-rising-need-for-care (accessed 3 April 2026).

[7] (5) States set new rules for AI chatbots in mental health care | LinkedIn, https://www.linkedin.com/pulse/states-set-new-rules-ai-chatbots-mental-health-care-mathematica–wpkie/ (accessed 3 April 2026).

[8] Mausbach BT, Moore R, Roesch S, et al. The Relationship Between Homework Compliance and Therapy Outcomes: An Updated Meta-Analysis. Cogn Ther Res 2010; 34: 429–438.

[9] Autio C, Schwartz R, Dunietz J, et al. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST, https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence (2024, accessed 3 April 2026).