AI Agents For Mental Health: Different Therapeutic Styles and Outcomes

Yasmina El Fassi • February 19, 2026

Share this article

What do Woebot, Wysa and Youper have in common? These are all AI agents that use therapeutic techniques to help users improve mental well-being, guide meditation and even help with managing anxiety.


In this article, AI mental‑health agents are goal‑directed conversational systems that sit with you in a chat or voice interface to support specific wellbeing tasks; for example, walking through CBT‑style exercises, practicing coping strategies, or checking in on mood over time.


In the broader AI literature, these would be considered agents because they are built around particular goals and workflows, whereas “agentic” AI usually refers to more autonomous systems that can independently plan multi‑step actions, call tools, and adapt their behaviour with relatively little human steering.

Translating that distinction into the mental‑health space, the tools we discuss here behave more like tightly scoped, therapeutically scripted companions than fully agentic systems that roam across apps and channels on your behalf.


The mental health industry has been completely disrupted by these agents, as AI no longer plays the role of a simple symptom checker or static content library. Millions of users make good use of these conversational users for a support that “feels” human.


When AI agents adopt therapeutic styles, from CBT‑inspired coaching to companion‑like reassurance, their interaction patterns start to shape how people think, feel and act over time. For teams building these products, the therapeutic style is therefore not an aesthetic choice; it is a safety decision with regulatory, reputational and ethical consequences

When an AI agent remembers past conversations, uses empathic language, or checks in unprompted, many people will treat it less like a feature and more like a therapist, coach or confidant.


Three dynamics are particularly important in this case:

If a system sounds like a therapist, users can infer that it must be clinically tested, supervised and suitable for “people like them”, even when it is positioned as non‑clinical.

Repeated late‑night chats, daily mood logs and personalised reflections create continuity and perceived caring, even when the system is driven by generic models and prompts.

As agents suggest coping strategies, reframe thoughts or encourage disclosure, they quietly move into roles traditionally held by therapists, peers or family, but without the training, accountability and boundaries those roles normally carry.

There is a plethora of psychotherapy styles, and since practitioners need to tailor the style to the individuality of the patient, there might even be as many styles as people on earth.

examples of

Psycotheraphy Styles

These systems focus on helping users understand what they are experiencing: “What is anxiety?”, “Why am I feeling this way?”, “What can people do in this situation?”. They echo the educational, rationale‑providing components that appear across many therapies; especially cognitive‑behavioural approaches, which often start by linking symptoms to understandable models and offering a coherent treatment rationale. Typical behaviours of these AI agents include: Short explanations, normalising statements, evidence‑aligned self‑help tips, clear signposting to human resources, limited emotional mirroring. The main risks associated with this approach is the generic or culturally misaligned content which can misinform or invalidate; developmental appropriateness is often an afterthought for children and adolescents.

Coaching‑style agents most visibly mirror cognitive‑behavioural therapies (CBT) and related approaches, which focus on the links between thoughts, feelings and behaviours and use structured tasks to support change. Usually, behaviours. of these agents encompass: Thought‑records, behavioural activation tasks, problem‑solving steps, goal‑setting, “homework” prompts and progress summaries. However, when stripped from case formulation and clinical judgment, CBT‑like tasks can become rigid, decontextualised and subtly blaming (“you just need to change your thinking”), especially for users facing structural constraints or trauma. This constitutes a major risk on users, especially the most vulnerable ones. 

Relational agents emphasise warmth, continuity and open‑ended conversation. They often borrow cues from humanistic and psychodynamic traditions that foreground the therapeutic relationship, empathy and the client’s subjective experience. These AI agents usually rely on small talk, empathic statements, following up on earlier conversations, “I‑language” (“I’m here for you”), apparent unconditional positive regard or a form of sycophancy. It appears that one of the reasons these agents were developed was to reduce loneliness, among other things. However, this creates risks of high dependency, blurred boundaries and unrealistic expectations from peers for example when it comes to emotional support. 

Once teams recognise how their systems echo psychotherapy traditions, the task is to design interaction patterns that preserve value and reduce harm.

Design Principles for Safer Therapeutic Styles

Role Clarity and Boundaries

Role and limits should be legible in the interaction, not only in the terms of service.


  • Use plain, repeated statements of what the system is and is not (“I’m a digital tool, not a therapist; I can’t diagnose or respond to emergencies”).
  • Avoid claims that overstate agency, care or expertise, especially those that mimic humanistic or psychodynamic language of deep understanding and unconditional acceptance.


For child‑reachable systems, boundaries need to be developmentally appropriate, visually reinforced and consistent across features.

Interaction Guardrails and Safe Exits

Guardrails should be visible and testable.


  • Define and test escalation patterns for risk cues, including when the agent stops generative conversation and surfaces crisis resources.
  • Provide obvious exit points: options to change topic, reduce intensity, contact a human or remove content, without penalty.


These patterns should be monitored and refined, not treated as one‑off compliance tasks.

Calibrated Empathy

Empathy scripts drawn from humanistic and integrative traditions need calibration when implemented in AI.


  • Use validating language that acknowledges experience without promising outcomes, unconditional availability or personal care.
  • Avoid self‑disclosure and “best friend” positioning, especially with young users; keep the agent’s status as a tool visible.


The goal is to support users while keeping expectations in line with what the system can reliably and ethically deliver.

Preserving User Agency

Interaction should support, not replace, users’ own reasoning and decision‑making.


  • Offer options and branching paths (“Would you like to understand what might be happening, explore coping strategies, or talk about support from people around you?”) rather than single, pre‑selected routes.
  • Use prompts that invite reflection (“How does that suggestion fit your situation?”) instead of treating outputs as prescriptions.


In CBT‑like and embedded styles, this is central to avoiding a dynamic where the agent quietly becomes the main decision‑maker.

This is the text area for this paragraph. To change it, simply click and start typing. Once you've added your content, you can customize its design by using different colors and fonts. 

Governance:

From Styles to Accountability

Therapeutic styles emerge from content, product and UX decisions, but they require explicit governance if they are to remain safe over time.


Organisations need clear ownership for role boundaries, escalation thresholds, data practices and acceptable trade‑offs between engagement and risk.


These responsibilities should span the lifecycle: initial design, deployment, model or prompt updates, and decommissioning. 


The most significant risks often appear only after launch. It is of utmost importance to involve clinicians, psychotherapy researchers and people with lived experience when reviewing logs and making changes to styles and guardrails.


Work on common factors and “what works for whom” shows that different therapeutic approaches need to be adapted continuously to the people using them. For AI systems, that adaptation has to be engineered and governed rather than assumed.

Practical Checklist

Here is a brief checklist for teams who would like to apply what has been discussed so far:

Which psychotherapy‑inspired elements (e.g., CBT‑like structure, humanistic‑like empathy, psychodynamic‑style exploration, mindfulness practices) are present in your agent?

How might users reasonably overestimate the level of care, expertise or crisis support these elements imply?

Where and how do you state the system’s role and limits (e.g., inside the interaction itself)?

How are you monitoring real‑world use for dependency, misuse, misfit and unintended behavioural effects across different user groups?

Who has the authority to change styles, guardrails and escalation patterns when new risks emerge, and how often do you revisit those decisions?

If you’re building (or scaling) an AI mental-health agent and want to pressure-test your therapeutic style, boundaries, interaction patterns, and escalation design, we’re here to help.


Reach out for a short review session to identify risk hotspots and translate them into an implementable safety & governance pack, your team can ship and maintain.

References


Biswas, M., & Murray, J. (2024). “Incomplete Without Tech”: Emotional Responses and the Psychology of AI Reliance. In Lecture notes in computer science (pp. 119–131). https://doi.org/10.1007/978-3-031-72059-8_11


Chandra, S., Shirish, A., & Srivastava, S. C. (2022). To be or not to be . . .Human? Theorizing the role of Human-Like competencies in conversational artificial intelligence agents. Journal of Management Information Systems, 39(4), 969–1005. https://doi.org/10.1080/07421222.2022.2127441


Darcy, A., Beaudette, A., Chiauzzi, E., Daniels, J., Goodwin, K., Mariano, T. Y., Wicks, P., & Robinson, A. (2023). Anatomy of a Woebot® (WB001): agent guided CBT for women with postpartum depression. Expert Review of Medical Devices, 20(12), 1035–1049. https://doi.org/10.1080/17434440.2023.2280686


Dwivedi, Y. K., Helal, M. Y. I., Elgendy, I. A., Alahmad, R., Walton, P., Suh, A., Singh, V., & Jeon, I. (2025). Agentic AI Systems: What it is and isn’t. Global Business and Organizational Excellence, 45(3), 253–263. https://doi.org/10.1002/joe.70018


Johnson, C., Egan, S. J., Carlbring, P., Shafran, R., & Wade, T. D. (2024). Artificial intelligence as a virtual coach in a cognitive behavioural intervention for perfectionism in young people: A randomised feasibility trial. Internet Interventions, 38, 100795. https://doi.org/10.1016/j.invent.2024.100795


Jones, B., Stemmler, K., Su, E., Kim, Y., & Kuzminykh, A. (2025). Users’ expectations and practices with agent memory (pp. 1–8). https://doi.org/10.1145/3706599.3720158


Khawaja, Z., & Bélisle-Pipon, J. (2023). Your robot therapist is not your therapist: understanding the role of AI-powered mental health chatbots. Frontiers in Digital Health, 5, 1278186. https://doi.org/10.3389/fdgth.2023.1278186


Kouros, T., & Papa, V. (2024). Digital Mirrors: AI companions and the self. Societies, 14(10), 200. https://doi.org/10.3390/soc14100200

Malmqvist, L. (2025). Sycophancy in Large Language Models: Causes and mitigations. In Lecture notes in networks and systems(pp. 61–74). https://doi.org/10.1007/978-3-031-92611-2_5


Maurya, R. K., Montesinos, S., Bogomaz, M., & DeDiego, A. C. (2024). Assessing the use of ChatGPT as a psychoeducational tool for mental health practice. Counselling and Psychotherapy Research, 25(1). https://doi.org/10.1002/capr.12759

Norcross, J. C., & Wampold, B. E. (2010). What works for whom: Tailoring psychotherapy to the person. Journal of Clinical Psychology, 67(2), 127–132. https://doi.org/10.1002/jclp.20764


Portell, S. (2026). Building AI Responsibly for Children: A Practical framework. HCRAI. https://www.hcrai.com/building-ai-responsibly-for-children-a-practical-framework


Sapkota, R., Roumeliotis, K. I., & Karkee, M. (2025). AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges. Information Fusion, 126, 103599. https://doi.org/10.1016/j.inffus.2025.103599


Short, F., & Thomas, P. (2014). Core approaches in counselling and psychotherapy. https://doi.org/10.4324/9780203773390


Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2025). The rise of AI Companions: How Human-Chatbot Relationships Influence Well-Being. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2506.12605

Author

Yasmina El-Fassi

Human-AI Interaction | PhD Researcher, Psychology & AI

Research Collaborator, HCRAI



< Older Post

Recent Posts

The Design System As The Operational Layer for Responsible Human-AI Interaction

By Sara Portell • February 6, 2026
Design systems were built to scale consistency, efficiency and quality in user-centric applications: reusable components, shared patterns and practices, and a common language across design and engineering , promoting collaboration. They improve velocity because teams stop solving the same interface problems repeatedly, providing measurable ROI . AI introduces both immense opportunities and complex (technical, legal and social) challenges, and it is reshaping the operating conditions traditional design systems were built for. User-facing outputs are adaptive and can vary by input, model behaviour can shift over time and responses that sound credible can still be wrong . These systems can also reproduce or amplify bias, creating unequal outcomes across users. In high-confidence, relational interactions, they can shape user judgment and behaviour . These shifts raise the bar for accountability, transparency, and governance across the full product lifecycle. The challenge is not only consistency and quality. It is ensuring consistency and quality safely, fairly and responsibly as both system behaviour and human behaviour evolve. At the same time, AI-powered copilots and no-code tools are increasingly used in the design process to support ideation, prototyping, and delivery, but their adoption also raises concerns about transparency, bias, privacy, and the need to preserve human judgment and oversight . Fast, polished design outputs often look complete even when the underlying logic is incomplete or flawed. As a result, familiar UX failures, misalignment with real user needs, hidden edge cases and context breakdowns, become harder to detect and more costly to correct later. Design systems can take on a bigger operational role in AI-enabled product development by codifying user-centric foundations, rules and infrastructure that guide consistent, safe, ethical and scalable human-AI experiences.
When AI Enters the Learning Process: Design Failures, Regulatory Risk and Guardrails for EdTech

When AI Enters the Learning Process: Design Failures, Regulatory Risk and Guardrails for EdTech

By Sara Portell • January 21, 2026
Generative AI (GenAI) and emerging agentic systems are moving AI into the learning process itself. These systems don’t stop at delivering content. They explain, adapt, remember and guide learners through tasks. In doing so, they change where cognitive effort sits. I.e., what learners do themselves and what gets delegated to machines. This shift unlocks significant opportunities. GenAI can provide on-demand explanations, examples and feedback at a scale. It can diversify learning resources through multimodal content, support learners working in a second language and reduce friction when students get stuck, lowering barriers to engagement and persistence. For some learners, AI-mediated feedback can feel psychologically safer, encouraging experimentation (trial and error), revision and assistance without fear of judgement . But these gains come with important risks. The same design choices that improve short-term performance, confidence, or engagement can weaken i ndependent reasoning, distort social development or introduce hidden dependencies over time . In educational contexts, especially those involving children and teens, we are talking about learning, safeguarding, regulatory and reputational risks. If the “Google effect” (digital amnesia) raised concerns about outsourcing memory to search engines , LLMs can be even more powerful in practice.
Designing AI Mental Health and Wellbeing Tools: Risks, Interaction Patterns and Governance

Designing AI Mental Health and Wellbeing Tools: Risks, Interaction Patterns and Governance

By Sara Portell • January 13, 2026
Designing AI Mental Health and Wellbeing Tools: Risks, Interaction Patterns and Governance
Building AI Responsibly for Children: A Practical Framework

Building AI Responsibly for Children: A Practical Framework

By Sara Portell • January 4, 2026
AI is alread y a core part of children’s and teens’ digital lives. In the UK, 67% of teenagers now use AI , and in the US 64% of teens report using AI chatbots . Even among younger children, adoption is significant: 39% of elementary school children in the US use AI for learning, and 37% of children aged 9-11 in Argentina report using ChatGPT to seek information, as stated in the latest Unicef Guidance on AI and Children. In parallel, child-facing AI products are expanding: more than 1,500 AI toy companies w ere reportedly operating in China as of October 2025. Adoption is accelerating across age groups and regions, often surpassing the development of child-specific ethical standards, safeguards and governance mechanisms.

Contact Us

Tel:  +351 915 679 908

Tel: +44 79 38 52 1514

contact@hcrai.com