Rule-based therapy chatbots (Woebot, Wysa)

1. Introduction

Version followed: Woebot as described in its 2017 randomised trial (Fitzpatrick, Darcy & Vierhile) and in 2025 closure reporting; Wysa as described in the Wysa Mental Wellbeing AI Instructions for Use, Version 3.0 (effective 12 November 2025), the 2022 FDA Breakthrough announcement and NICE HTE30 (Wysa Digital Referral Assistant). Product status changes often; every status line below carries the date of its source.

What it is. Rule-based (scripted) therapy chatbots are smartphone apps that hold a text conversation with the user, but whose replies come from content written in advance by clinicians and writers rather than generated freely by a large language model. Woebot's 2017 version is described as a decision tree with suggested replies, using natural language processing only at certain points to route the conversation to the next scripted node; the tree was the same for every participant [source: Fitzpatrick et al. 2017]. STAT News reported that Woebot's responses were "pre-scripted", not generated by a language model [source: STAT, July 2025]. Wysa describes its chatbot as a messaging-style AI that asks structured questions to support self-reflection, alongside a library of exercises [source: Wysa IFU v3.0]. Whether every current Wysa feature is fully scripted, or whether generative components have been added, is [uncertain: not verified]. Other products of this kind (e.g. Tess) exist [uncertain: not verified]. Generative (LLM-based) tools and clients' use of general-purpose AI for emotional support are covered in 097 and 098.

Core idea of change. These tools package self-help elements of CBT and related approaches (mood tracking, psychoeducation, spotting thinking patterns, goal setting, mindfulness and breathing exercises) into short, frequent, conversational interactions available at any hour, so that skills practice happens between, or in the absence of, human sessions [source: Fitzpatrick et al. 2017; Wysa IFU v3.0]. For the psychotherapist the tool is at most an adjunct: it can carry homework, between-session practice and self-monitoring, but it is not therapy and is not a crisis service [source: Wysa IFU v3.0; Fitzpatrick et al. 2017].

Status (date-stamped).

Woebot (consumer app): closed. Woebot Health told users it was retiring its app as of 30 June 2025; conversation history could be downloaded until then, and account data was to be anonymised after 31 July 2025 (as of April 2025 per MobiHealthNews; closure on 30 June 2025 confirmed as of August 2025 per Telehealth.org). Founder Alison Darcy attributed the shutdown "largely" to the cost and challenge of FDA marketing-authorisation requirements, and to the arrival of large language models the FDA had not worked out how to regulate (as of July 2025 per STAT News). About 1.5 million people had used it (per STAT News and Telehealth.org). Woebot held an FDA Breakthrough Device designation (2021) for WB001, a planned prescription digital therapeutic for postpartum depression combining CBT and interpersonal psychotherapy elements over an eight-week programme (as of April 2025 per MobiHealthNews); Breakthrough designation is not marketing authorisation. Whether any Woebot enterprise or health-system products continue after 2025 is [uncertain: not verified]. Earlier partnerships included Akron Children's (adolescents aged 13 to 17) and Virtua Health (per MobiHealthNews).
Wysa: available. FDA Breakthrough Device designation (May 2022) for its AI-based conversational agent for adults 18+ with chronic musculoskeletal pain (over three months) and associated depression and anxiety (as of May 2022 per Business Wire / Wysa). Wysa's own note states the designation is not a blanket approval of all Wysa products. The consumer wellbeing app is described as a standalone self-help tool, not classified as Software as a Medical Device, not for diagnosis, prevention or treatment of any condition (as of November 2025 per Wysa IFU v3.0). In England, NICE early value assessment HTE30 recommended the Wysa Digital Referral Assistant (UKCA Class I AI-supported e-triage chatbot) and Limbic Access (UKCA Class IIa) as digital front doors that collect referral information and standard questionnaires for NHS Talking Therapies assessments (per NICE HTE30; date of the recommendation [uncertain: not verified]). Wysa also lists the NHS among its partners (per Business Wire, 2022).
US regulatory context. As of November 2025, the FDA had authorised over 1,200 AI-enabled devices but none for mental health uses, and fewer than 20 non-AI digital mental health devices (per Psychiatric Times report of the 6 November 2025 Digital Health Advisory Committee). Illinois's Wellness and Oversight for Psychological Resources Act (HB1806, signed August 2025) bans using AI to provide mental health and therapeutic decision-making; licensed professionals may use AI for administrative and supplementary support; fines up to $10,000 per violation, enforced by IDFPR (as of August 2025 per IDFPR press release). Check the current law in your own jurisdiction [synthesis].

Evidence attributed (for orientation only). Woebot: a two-week RCT in 70 US college-age adults (18 to 28) found greater reduction in depression (PHQ-9) than an information-only ebook control, with no between-group difference in anxiety in the intention-to-treat analysis; authors note a small sample, short duration, no follow-up, unblinded design and that the second author founded the company [source: Fitzpatrick et al. 2017]. Wysa: an uncontrolled real-world study of 129 anonymous users found larger PHQ-9 improvement in "high" than "low" users, with authors noting no randomisation and possible regression to the mean [source: Inkster et al. 2018]. Wysa cites a JMIR trial in chronic pain as the basis of its FDA designation [source: Wysa blog 2022; trial itself [uncertain: not verified]]. Telehealth.org cautions that company-funded and short trials limit the evidence base [source: Telehealth.org 2025].

2. Framework

This is a category of tools, not a psychotherapy with its own phases. The app's own structure is brief, self-paced, conversational sessions (Woebot's trial: up to 20 sessions over about two weeks, daily or alternate-day prompts, weekly mood charts) [source: Fitzpatrick et al. 2017]; Wysa offers on-demand chats, a topic-grouped exercise library, reminders, and in some locations a paid asynchronous human coach [source: Wysa IFU v3.0]. The framework below is the psychotherapist's integration sequence [synthesis].

Phase	Goals	Typical format	Move on when
1. Ask and assess	Find out whether the client uses or wants such an app; check fit, age, risk, contraindications	Part of intake or any session, 5 to 15 minutes	Client is a suitable candidate, or already using one
2. Inform and agree	Psychoeducation on what the tool is and is not; privacy review; crisis plan that does not rely on the app; agree a specific purpose	One session segment; note in record	Client consents, purpose and limits are written into the plan
3. Trial	Client uses agreed features between sessions	2 to 4 weeks between reviews	Review shows benefit, no harm, workable engagement
4. Review and integrate	Bring app material into session; monitor outcomes, engagement, adverse effects	Brief check every session; formal review every 4 to 6 weeks	Continue, adjust use, or stop
5. Ending	Plan continued use or stopping; consolidate skills without the app; data decisions	One session segment	Client has a plan not dependent on the product

All timings in the table are [synthesis]. Wysa's wellbeing app is not intended for crisis, severe or enduring mental health problems, or significant cognitive impairment [source: Wysa IFU v3.0]; those features at phase 1 mean no trial, or a move straight to phase 5.

3. Guiding principles

Adjunct, never a replacement. The app supports, but does not substitute for, assessment, therapy, medication or crisis care; Woebot's trial told participants it should not replace therapy and Wysa's IFU says it does not replace a clinician [source: Fitzpatrick et al. 2017; Wysa IFU v3.0]. Practical meaning: keep the treatment plan built on human care; the app is one homework channel.
Not a crisis service. Wysa states it is "not registered for use as a suicide or self-harm detection, prevention or crisis support service" [source: Wysa IFU v3.0]. Practical meaning: every safety plan names human and emergency contacts first.
Know which kind of tool it is. Scripted and generative tools carry different risks (unexpected or fabricated content arises mainly with generative tools; see 097, and FDA committee concerns about confabulation and inappropriate content [source: Psychiatric Times 2025]). Practical meaning: ask the client which product and version; check the vendor's current description.
Date-stamp everything. Products close or change (Woebot, 2025) [source: MobiHealthNews; STAT]. Practical meaning: recheck status, terms and privacy policy before recommending, and at reviews.
Regulatory claims precisely stated. Breakthrough designation is not approval; a wellbeing app is not a medical device [source: Wysa blog 2022; Wysa IFU v3.0]. Practical meaning: describe status to clients in the vendor's and regulator's exact terms.
Respect age limits. Wysa: 13 and over, or 11 and over if the providing organisation agrees [source: Wysa IFU v3.0]. Practical meaning: for minors, involve guardians and services as law requires.
Informed choice and privacy. Clients decide whether to use the tool after knowing what data is collected and where it goes [synthesis; data details per Wysa IFU v3.0]. Practical meaning: read the privacy policy together.
Monitor like any intervention. Measure symptoms, engagement and harms; stop if not helping [synthesis].
Stay within your jurisdiction's law and professional code. Some places restrict AI in therapy (Illinois, 2025) [source: IDFPR 2025]. Practical meaning: you recommend a self-help tool; you do not delegate therapeutic decisions to software.
The relationship stays human. The chatbot has no feelings or empathy, as Wysa itself states [source: Wysa IFU v3.0]. Practical meaning: explore what the client experiences in the chat, including attachment to it, as therapy material.

4. Intervention inventory

Part A describes what the apps themselves deliver, as their developers and published studies describe it; it is descriptive, so that you can talk about it accurately with clients. Part B gives the psychotherapist's own interventions. The example language throughout is illustrative.

Part A: what the apps deliver

A1. Conversational check-in and mood tracking

Purpose / when: Daily self-monitoring and engagement. In Woebot's 2017 version each interaction began with a context question and a mood check answered with words or emoji; the bot sent one personalised prompt daily or every other day and a weekly mood chart with a short automated summary [source: Fitzpatrick et al. 2017].
Steps (as the app runs it): Prompt arrives; user reports mood; bot replies in a way tailored to the stated mood and offers content to match (e.g. help with a specific anxious event) [source: Fitzpatrick et al. 2017].
Example language: "If you are using the daily check-ins, could you bring this week's mood chart? I'm curious what it shows you that you didn't notice at the time." (illustrative)
Watch for: Check-ins becoming a ritual of reporting with no change in behaviour; reassurance-seeking through repeated check-ins; distress on low-mood days with no human follow-up [synthesis]. Working: client uses the chart to spot patterns and brings them to session [synthesis].

A2. Scripted CBT psychoeducation and thinking-pattern exercises

Purpose / when: Teaching CBT concepts in short pieces. Woebot's psychoeducation was adapted from CBT self-help books and delivered through short videos and word games about cognitive distortions [source: Fitzpatrick et al. 2017]. Wysa's care library contains interactive exercises drawn from CBT, mindfulness and positive psychology, grouped by topic such as low mood, worry and sleep [source: Wysa IFU v3.0]; the 2018 study also listed DBT, motivational interviewing and behavioural reinforcement among the practices used [source: Inkster et al. 2018].
Steps: User chooses or is offered a topic; the bot walks through a structured exercise (e.g. naming a thought, identifying a pattern, reframing) [source: Fitzpatrick et al. 2017; Wysa IFU v3.0 for exercise library]. Exact exercise scripts vary by version [uncertain: not verified].
Example language: "The app called that thought 'all-or-nothing'. Does that label fit for you, or does it miss something about what was going on?" (illustrative)
Watch for: Labels used to dismiss valid concerns (e.g. trauma-related fear labelled as a distortion); intellectual agreement without emotional shift; scripts not fitting the client's culture, first language or reading level (Wysa notes reduced effectiveness when English is not the first language) [source: Wysa IFU v3.0 for language; rest synthesis].

A3. Goal setting and follow-up

Purpose / when: Accountability for small behavioural steps. Woebot asked about a personal two-week goal and followed up on it [source: Fitzpatrick et al. 2017]; Wysa describes helping users set personal goals and take practical steps, including "micro-actions" [source: Wysa IFU v3.0; Business Wire 2022].
Steps: User names a goal; app prompts later; user reports progress [source: as above].
Example language: "Let's make the goal you put into the app match the one we agreed here, so the reminders work for our plan, not a separate one." (illustrative)
Watch for: Goals set too large; self-criticism when the app records missed goals; competing goal lists between app and therapy [synthesis].

A4. Mindfulness, breathing, grounding and audio exercises

Purpose / when: Short regulation exercises on demand. Wysa offers meditation, breathing and mindfulness exercises and audio exercises; its Crisis Help button opens a grounding exercise [source: Wysa IFU v3.0; Business Wire 2022]. Wysa warns not to attempt physical exercises if medical advice says they may cause harm [source: Wysa IFU v3.0].
Steps: User opens an exercise and follows the guided text or audio [source: Wysa IFU v3.0].
Example language: "Which of the app's breathing or grounding exercises have you tried? Let's practise one here so you know how it feels when you're calm, before you need it." (illustrative)
Watch for: Increased anxiety or dissociation during breath- or body-focused exercises (relevant in panic, trauma, some health conditions); using exercises to avoid feared situations agreed for exposure [synthesis].

A5. Risk-language detection, Crisis Help and safety plan

Purpose / when: Wysa recognises language that may indicate risk of suicide, self-harm or abuse; it first asks the user to confirm whether urgent support is needed, then, if confirmed, gives local or national crisis numbers with direct dial or text, offers a safety-plan conversation and tools for distressing thoughts. A Crisis Help button on the home screen shows emergency contacts and opens the safety plan and a grounding exercise [source: Wysa IFU v3.0]. Wysa does not publish in this document how detection works or how accurate it is, and states it is not a crisis or suicide-detection service [source: Wysa IFU v3.0]. In Woebot's trial, participants who reported long-standing depression, suicidality or self-harm were automatically given helpline and crisis text numbers [source: Fitzpatrick et al. 2017].
Steps (therapist side): Look at the app's safety plan with the client; make sure it matches the safety plan made in therapy; put human and emergency contacts first (see B8) [synthesis].
Example language: "The app may notice some worrying words, but it can miss things and it can't call anyone. So our plan is that if things get to that point, you contact [named person / service] first." (illustrative)
Watch for: Client disclosing risk or abuse only to the app; false reassurance that "the app would tell someone"; the detection prompt being dismissed; a safety plan in the app that conflicts with the clinical one [synthesis].

A6. Paid human coaching (Wysa)

Purpose / when: In some locations Wysa offers asynchronous human coaching for an extra fee [source: Wysa IFU v3.0].
Steps: Client chooses to add coaching within the app [source: Wysa IFU v3.0]. Coach qualifications and supervision arrangements: [uncertain: not verified].
Example language: "If you're thinking about the app's coaching option, it's worth us talking about how that would sit alongside our work, so you're not getting two different sets of advice." (illustrative)
Watch for: Overlapping or contradictory guidance; confusion about which helper to contact in a crisis; cost burden [synthesis].

A7. Digital front door and e-triage in NHS Talking Therapies (England)

Purpose / when: Not self-help but referral intake. NICE HTE30 describes Wysa Digital Referral Assistant (UKCA Class I) as collecting referral and demographic data and, if the referral is accepted, PHQ-9, GAD-7, WSAS and minimum data set information; signposting people who fail age or GP-location criteria; using mindfulness exercises during triage; producing a report for review and transferring data into the service record; it flags people for professional review on service-specific criteria. Limbic Access (UKCA Class IIa) collects similar information, produces a clinical report with risk levels and is used by about 40% of NHS Talking Therapies services [source: NICE HTE30].
Steps (therapist side): Read the generated report as referral information, not as an assessment; verify risk and presenting problems in your own assessment [synthesis].
Example language: "I've seen what you put into the online referral. I'd like to hear it in your own words too, because a form never captures everything." (illustrative)
Watch for: Risk information missed or understated in self-report; clients who disengage because the chatbot felt impersonal; people signposted away who still need help [synthesis].

Part B: the psychotherapist's interventions

Unless a source is tagged, the procedures in Part B are [synthesis]: good practice built from the product limits documented above and from standard clinical practice.

B1. Asking routinely about chatbot and app use

Purpose / when: At intake and at reviews, because clients may already be using an app, or may have lost one when it closed (Woebot, June 2025) [source for closure: MobiHealthNews; Telehealth.org].
Steps: (1) Ask openly about apps, chatbots and AI tools used for mood or stress. (2) Name the product and whether it is scripted or generative, and whether it was recommended by a service. (3) Ask what it is used for, how often, and what happens in hard moments. (4) Ask whether they have told it things they have not told anyone else. (5) Record it in the notes.
Example language: "Some people use apps or chatbots to help with their mood. Do you use anything like that? What's it like, and when do you tend to open it?" (illustrative)
Watch for: Shame or secrecy about use; the app as the main or only confidant; general-purpose AI being used instead (see 098). Working: the client talks about the app freely and it becomes ordinary therapy material.

B2. Assessing fit and contraindications

Purpose / when: Before recommending or endorsing use.
Steps: (1) Check the product's own stated limits: Wysa's wellbeing app is not for crisis, severe or enduring mental illness or significant cognitive impairment, and works less well when English is not the first language or with hearing or vision difficulties [source: Wysa IFU v3.0]. (2) Assess current risk with your usual risk assessment. (3) Consider literacy, digital access and cost. (4) Consider whether the client's difficulties (e.g. psychosis, mania, eating disorder, complex trauma) fall outside what a self-help CBT tool addresses. (5) Decide: recommend, accept existing use with monitoring, or advise against and explain why.
Example language: "I don't think an app is the right fit for what you're going through right now, because it isn't designed for this level of distress. Let's focus on getting you the right support first." (illustrative)
Watch for: Pressure (from waiting lists or cost) to use an app as a substitute for needed care; client feeling rejected if advised against.

B3. Psychoeducation: what the tool is and is not

Purpose / when: Before a trial, and whenever the client over- or under-estimates the tool.
Steps: (1) Explain scripted versus generative: a rule-based bot gives prewritten replies chosen from a script [source: STAT 2025; Fitzpatrick et al. 2017]; generative bots compose new text (see 097). (2) State it is self-help, not therapy, not a medical device unless the vendor and regulator say so, and not a crisis service [source: Wysa IFU v3.0]. (3) Explain that the bot does not have feelings or empathy [source: Wysa IFU v3.0]. (4) State the regulatory status in exact terms, e.g. that Breakthrough designation is not approval [source: Wysa blog 2022]. (5) Mention that products can close [source: STAT 2025].
Example language: "Think of it as an interactive workbook. It can remind you of skills and help you practise, but it can't understand you the way a person does, and it isn't watching out for your safety." (illustrative)
Watch for: Anthropomorphising ("it understands me better than anyone"); belief that the app is supervised by clinicians in real time (Woebot's trial explicitly said monitoring was not in real time [source: Fitzpatrick et al. 2017]).

B4. Privacy and data review with the client

Purpose / when: Before use, and when terms change or a product closes.
Steps: (1) Read the product's current privacy policy together or ask the client to. (2) Note what is collected: Wysa says it does not ask for email or phone, uses a nickname, encrypts data in transit and at rest, and lists ISO 27001, ISO 27701, SOC 2 Type 1 and Cyber Essentials; storage locations and international transfers are in its privacy policy [source: Wysa IFU v3.0]. Access via referral or QR codes may link use to a service [source: Wysa IFU v3.0]. (3) Discuss using the app in private, as Wysa advises [source: Wysa IFU v3.0], and phone lock / passcode. (4) For closures, note export and deletion windows (Woebot: download until 30 June 2025, anonymisation after 31 July 2025) [source: MobiHealthNews]. (5) Discuss what the client chooses to share with the app, especially about third parties.
Example language: "Before you start, let's look at where what you type actually goes and who could see it, so you can decide what you're comfortable putting in." (illustrative)
Watch for: Shared or monitored phones (including in coercive relationships); employer-provided access where the client fears visibility; health data entered into general-purpose AI (see 098).

B5. Age limits and young people

Purpose / when: With any client under 18, or families asking about apps.
Steps: (1) Check the product's age limit: Wysa 13 and over, or 11 and over with the providing organisation's agreement [source: Wysa IFU v3.0]; Wysa's FDA Breakthrough indication was for adults 18+ [source: Business Wire 2022]; Woebot's trial included adults 18 to 28 [source: Fitzpatrick et al. 2017], while Woebot had an adolescent partnership for ages 13 to 17 [source: MobiHealthNews]. (2) Follow local consent, guardian-involvement and safeguarding rules. (3) Agree with the young person what will be shared with parents. (4) Revisit safeguarding: the app's abuse-language prompts are not a safeguarding referral [synthesis, based on Wysa IFU v3.0 limits].
Example language: "This app is meant for people your age and older, but it's not a person you can tell about something unsafe happening to you. If that ever happens, I want you to tell me or another adult you trust." (illustrative)
Watch for: Under-age use; disclosure of abuse to the app only; parents using the app as a substitute for care.

B6. Agreeing a purpose and building it into the treatment plan

Purpose / when: Once the client chooses to use the tool.
Steps: (1) Pick one or two specific uses (e.g. daily mood check-in; practising thought-challenging between sessions; a breathing exercise before a feared task). (2) Match app goals to therapy goals (A3). (3) Agree a review date. (4) Document the agreement, the information given and the client's choice. (5) Where the tool is offered while waiting for treatment, say clearly that it does not replace the assessment or treatment the client is waiting for.
Example language: "So the plan is: use the evening check-in, and when you notice a big drop, try the thought exercise and screenshot or jot down what came up. We'll look at it together next week." (illustrative)
Watch for: Unfocused use; the app introducing techniques that clash with the current phase of therapy (e.g. cognitive challenging while you are doing acceptance-based work).

B7. Bringing app material into session

Purpose / when: Each session while the app is in use.
Steps: (1) Ask what the client did in the app and what they noticed. (2) Review mood charts or exercise outputs the client chooses to show. (3) Link app work to session themes; correct misunderstandings of concepts. (4) Reinforce skills that transfer away from the phone.
Example language: "You did the worry exercise three times this week. What did you notice about the worries that came up? Was any of it new?" (illustrative)
Watch for: Clients reading app content aloud instead of reflecting; using the app's framing to avoid emotional material in session.

B8. Safety and crisis planning that does not depend on the app

Purpose / when: For any client with past or current risk who uses a chatbot.
Steps: (1) Do your standard risk assessment and safety plan. (2) Make clear the app is not a crisis service [source: Wysa IFU v3.0]. (3) List personal contacts, your service's out-of-hours route and emergency services first; the app's crisis numbers are a backup. (4) If the client uses the app's safety plan, check it matches yours. (5) Agree that if the app flags risk or the client confirms urgent need, they also contact a human. (6) Include lethal-means safety in the usual way.
Example language: "If you ever find yourself telling the app you don't feel safe, that's the moment to use the first steps on this plan and reach a person, even if the app has already shown you numbers." (illustrative)
Watch for: Rising risk disclosed between sessions only in-app; reliance on the bot to "talk me down"; nights spent in long chats in place of sleep or human contact.

B9. Monitoring outcomes, engagement and harms

Purpose / when: Every session briefly; formal review every 4 to 6 weeks.
Steps: (1) Use your usual outcome measures (e.g. PHQ-9, GAD-7) rather than the app's own scores. (2) Ask about frequency and pattern of use. (3) Ask directly about any upsetting, unhelpful or confusing replies. (4) Note any adverse effects; Wysa states none have been identified in research or feedback [source: Wysa IFU v3.0], but you should still ask. (5) Decide with the client to continue, change or stop.
Example language: "Has the app ever said anything that felt wrong, dismissive or upsetting? I want to know about that as much as about the parts that help." (illustrative)
Watch for: Plateau or worsening; compulsive use; withdrawal from people. Working: symptom scores improve and the client uses skills without prompting.

B10. Collaboration with services and referrers

Purpose / when: When an NHS or employer service has supplied the app or used a chatbot front door (A7), or when another clinician supervises coaching (A6).
Steps: (1) Clarify who is responsible for the client's care and for responding to flagged risk. (2) With consent, share relevant information. (3) Report product safety problems through the vendor and, where available, the regulator's reporting system (e.g. MHRA Yellow Card for devices in the UK) [uncertain: not verified for specific products].
Example language: "Who at the service sees what you put into the app? Let's check, so we know whether anyone would notice if you were struggling." (illustrative)
Watch for: Assumed monitoring that does not exist; gaps in responsibility between services.

B11. Managing product closure or change

Purpose / when: When a product closes, changes features or terms, or adds generative AI.
Steps: (1) Tell the client what is known, with dates. (2) Help them export data if they want it and if the vendor allows [source: MobiHealthNews for Woebot export window]. (3) Explore the loss; some clients feel real grief. (4) Consolidate skills learned into paper or session-based tools. (5) Reassess whether any replacement is needed and whether it is scripted or generative.
Example language: "It makes sense to feel a bit lost now the app has gone; it was part of your routine. Let's write down the things it helped you with so you keep them." (illustrative)
Watch for: Switching to a general-purpose or generative chatbot without discussion (see 097, 098).

B12. Ending

Purpose / when: At the end of therapy or when stopping use.
Steps: (1) Decide with the client whether to keep using the app after therapy. (2) If continuing, restate limits and crisis routes. (3) If stopping, discuss data deletion. (4) Write a relapse-prevention plan that does not depend on the product.
Example language: "If you keep the app after we finish, use it for practice, and remember the plan we wrote: if things slide, that's a sign to get back in touch with a person." (illustrative)
Watch for: The app being treated as "my therapist now"; ending feeling less important because "I still have the app".

B13. Clinician due diligence and professional limits

Purpose / when: Before recommending any product.
Steps: (1) Check current status, ownership and terms with the date. (2) Check regulatory status in your jurisdiction (FDA, UKCA/CE, NICE, DiGA) as the regulator states it. (3) Read the published evidence and note who funded it [source: Telehealth.org 2025 for funding caution]. (4) Check local law; e.g. Illinois bans AI from providing therapeutic decision-making [source: IDFPR 2025]. (5) Check your professional body's guidance and your indemnity. (6) Document the recommendation and its basis.
Example language (to a colleague): "Before we put this on the service leaflet, can we check its current status and terms, and who responds if it flags risk?" (illustrative)
Watch for: Marketing language ("AI therapist", "virtual therapy") that overstates the product; Telehealth.org flags such terms [source: Telehealth.org 2025].

5. Handling problems

The sources describe product limits but give almost no guidance for these situations; all responses below are [synthesis] unless tagged.

Client stops using the app (non-adherence). Treat it as information, not failure. Ask what got in the way (boredom with repetitive scripts, notifications felt intrusive, it felt impersonal). Either adjust to one simple use, switch to paper or session-based homework, or drop the app. Example: "It's fine if the app didn't suit you. What part of it, if any, was worth keeping?" (illustrative)
Over-use or dependence. Long or night-time chats, checking many times a day, distress when unable to use it. Explore what need the app meets (company, reassurance, avoidance). Set limits together; build human contact and offline coping; consider whether reassurance-seeking is part of an anxiety or OCD cycle to be treated directly. Wysa itself recommends regular screen breaks [source: Wysa IFU v3.0].
"The app understands me better than you do" (rupture). Take it seriously as relationship feedback; ask what the app does that feels better (always available, no judgement, no waiting). Acknowledge real limits of your availability; work on what can be repaired in the therapy relationship. Avoid competing with or disparaging the app.
Scripted replies that miss the point. Rule-based bots follow a fixed tree [source: Fitzpatrick et al. 2017], so they can respond to the wrong thing or repeat themselves. Clients may feel unheard or invalidated, or blame themselves. Normalise: "It's following a script, so it can miss what you actually meant." (illustrative) Bring the missed material into session.
Risk disclosed to the app, not to you. Ask routinely whether the app has shown crisis numbers or a safety plan recently. If yes, do a full risk assessment and update the safety plan (B8). Remind the client the app cannot act on what it detects and is not a crisis service [source: Wysa IFU v3.0].
Overwhelm from exercises. Breathing, body-scan or trauma-related reflection prompts can raise anxiety, panic or dissociation in some clients. Pause the exercise type, teach grounding in session, and choose alternatives. Wysa advises against physical exercises where medical advice says they may harm [source: Wysa IFU v3.0].
Conflict between app techniques and therapy approach. E.g. app pushes cognitive challenging during acceptance-based or trauma-processing work. Explain the difference; agree which features to use and which to leave for now.
Product closes or changes mid-treatment. See B11. With Woebot, users had about two months' notice and a short download window [source: MobiHealthNews; Telehealth.org]. Process the loss; consolidate skills; check any replacement carefully.
Client moves to a general-purpose or generative chatbot. Ask what it is, what it is used for, and whether it has said anything harmful; explain the specific risks the FDA advisory committee discussed, such as confabulated or inappropriate content and suicidal-ideation monitoring [source: Psychiatric Times 2025]; see 097 and 098 for detailed guidance.
Privacy fears or a data incident. Review what was shared; help the client delete or export data and change access settings; if a coercive partner may have seen content, carry out a safety assessment.
Waiting-list substitution. If a service offers an app instead of timely care, support the client in requesting the care they need; document unmet need; do not present the app as treatment [source for "not a treatment": Wysa IFU v3.0].
Young people using a product below its age limit, or a parent pushing an app. Follow safeguarding and consent rules; explain the age limit and the limits of the tool to the family (B5) [age limit source: Wysa IFU v3.0].
Your jurisdiction restricts AI in therapy. Recommending a self-help app may still be allowed, but you must not use AI for therapeutic decision-making where the law forbids it (e.g. Illinois) [source: IDFPR 2025]; seek advice from your licensing body if unsure.

Sources

Web sources:

Further reference: the JMIR Formative Research 2022 chronic-pain trial of Wysa that Wysa cites as the basis of its FDA designation (formative.jmir.org/2022/5/e36203) [uncertain: not verified].

Related guides: 097 and 098 (generative AI chatbots; clients' use of general-purpose AI for emotional support).

← Previous · IX-095Mental-health and wellbeing apps Next · IX-097 →Generative-AI therapy chatbots (Therabot; experimental)