The quality of a wellbeing application or protocol is not shown only when everybody follows the anticipated route. It is shown when a person cannot use the form, does not understand a question, needs to speak with someone, declines to share data, or is living through something a scale cannot describe. An exception is not a marginal fault. It is information about the system's limits and about the regard given to the person left outside its pattern.

A digital tool can organise reminders, offer general information, and make an initial route easier to find. It does not diagnose, interpret an experience on its own, or decide what care a person needs. In immediate danger, with thoughts of self-harm or harming another person, violence, abuse, severe confusion, or an inability to stay safe, the appropriate step is to contact local emergency, health, or crisis services. In Spain, 112 handles emergencies and 024 supports people affected by suicidal behaviour.

The exception is a design test#

Many services are built around an imagined person: they have a private phone, reliable connection, time, fluent reading, a dominant language, privacy, and confidence to answer. That picture removes ordinary circumstances. There may be shift work, family care, disability, grief, violence, migration, financial hardship, or fear that somebody else will see the screen. If the only route requires those conditions, the service is not neutral; it distributes support and burdens unequally.

Many services are built around an imagined person: they have a private phone, reliable connection, time, fluent reading, a dominant language, privacy, and confidence to answer.

A responsible protocol records exceptions without turning them into personal defects. It asks which part of the process does not work, permits a pause, and offers equivalent routes. It also examines patterns. If many people leave after a question, request help in another language, or cannot complete a video call, the answer is not to press them. It is to revisit the design, resources, and institutional decision that created the barrier.

Scenario one: the answer that will not fit the menu#

A person opens a municipal app after a week of anxiety and finds closed choices: sleep better, move more, practise breathing. Their main concern is that a former partner controls their phone and they are afraid to go home. No category lets them express this without leaving a detailed trace. Selecting a sleep option would be misleading simplification; leaving the app may be the safer decision.

The appropriate response is not to request further detail through a bot. The screen should provide a discreet exit, plain privacy explanations, and contacts for human or specialist support in the local setting. A quick-exit button can reduce exposure, but it does not replace a safety assessment or guarantee protection. The team must understand that an incomplete enquiry can be prudent rather than evidence of disinterest.

Scenario two: the inaccessible video appointment#

A workplace programme offers a wellbeing session by video. A worker with low vision uses a screen reader, but the calendar lacks labels and the platform requires images to be identified. Another person cannot join because they share a room and have no headphones. Both are classified as absent. The report celebrates attendance by those able to enter and does not measure exclusion created by the channel itself.

Accessibility means planning alternative formats before a complaint arrives: telephone, readable text, captions, interpretation where needed, material compatible with assistive technology, and reasonable times. It also requires a person who can resolve barriers without forcing disclosure of a condition to a manager. Equivalence matters. An alternative must not provide less information, less confidentiality, or poorer follow-up choices.

Scenario three: an automated alert with consequences#

A weekly questionnaire assigns a high alert to a student who selects several answers about distress. The system sends the result to a school administrator without plainly explaining who will see it. The alert might reflect a difficult night, a literal response, an important situation, or an error. Automatically turning it into a disciplinary label, a public call, or a broad family notification can deepen harm and close later conversations.

Automated signals may help prioritise a carefully defined human review, never make sensitive decisions without context. There should be trained staff, an understandable explanation, a way to correct data, and a proportionate procedure. Review should ask respectfully, not interrogate in order to confirm a model. If urgent risk appears, the protocol should activate appropriate help; otherwise it should avoid surveillance, punishment, and unnecessary disclosure.

Equity is not identical treatment#

Giving everybody the same screen can look impartial while ignoring prior inequality. An interface only in one language, a call only during working hours, or advice that assumes money and private space favours some people over others. Equity requires identifying what resources the service demands and who can supply them. It also requires participation by affected people, disability organisations, and community mediators when the product is assessed.

Data collection needs the same caution. Asking for sensitive information to tailor an experience can create risk if there is no specific purpose, appropriate basis, limited access, and defined retention period. Minimisation is not a loss of precision; it is a way to reduce foreseeable harm. An organisation should be able to explain why each datum is needed, who may see it, how it is protected, and how a person can correct or withdraw it where it is optional.

Referral means supporting a transition#

Simply saying “seek professional help” can be correct yet insufficient. A person may not know which service fits, fear the cost, need interpretation, live far away, or be unable to make a call safely. Prudent referral offers verifiable general information, options appropriate to the place, and a route onward without requiring more disclosure than necessary. It does not promise availability, outcomes, or absolute confidentiality that the receiving service does not control.

Referral should state limits. The app does not replace clinical care, protection from violence, or an emergency response. When indications suggest immediate danger or inability to remain safe, the priority is facilitating urgent contact with relevant local services, not extending a chat or collecting more metrics. Outside such urgency, a person may prefer primary care, an accredited professional, social services, or a specialist organisation; the system should respect that choice.

Do not automate sensitive decisions#

A sensitive decision affects access to support, privacy, education, work, family, or safety. No mood score, transcript, usage pattern, or missing click alone contains enough meaning to decide these matters. Models can be wrong, reproduce bias, and mistake correlation for need. Moreover, knowing that a machine classifies somebody changes how they answer and may cause important information to be withheld.

Human oversight is real only if the reviewer has authority to depart from the recommendation, time to listen, and training to recognise uncertainty. Clicking accept on a suggestion does not meet that test. Criteria should be documented, disparate outcomes tested, secondary uses limited, and an understandable challenge route offered. Accountability is not solved by a clause. It requires identifiable owners, audit, and remedy when a decision causes harm.

Scenario four: silence read as disinterest#

A service sends daily reminders and labels a non-responder as having “low adherence.” Yet that person changed number, shares a device with a family member, and chose to disable prompts because they cause distress. The system treats an absent signal as behaviour to correct. That interpretation turns a legitimate boundary or material difficulty into a reason for pressure.

Careful design permits pausing, leaving, and returning without penalty. Prompts should be configurable and discreet; inactivity must not trigger intrusive contact or reports to others. When an organisation needs follow-up for a specific obligation, it should explain this beforehand, limit it, and provide human contact to review circumstances. Autonomy includes the right not to make every day into wellbeing data.

Privacy, safety, and trust#

Wellbeing information can reveal routines, relationships, beliefs, and vulnerable moments. Privacy is therefore not a legal screen at the end of enrolment, but a practical safety condition. Explanations should be short, accessible, and visible at the moment of choice. They should distinguish data needed for a function, optional uses, recipients, and confidentiality limits. Protective settings should be on by default whenever feasible.

Safety also includes the physical environment. An explicit message on a shared screen can expose someone; a notification can appear before a partner, family member, or employer. Neutral names, notification control, and quick exit are useful, but they should not imply that these measures are sufficient under coercion. Risk analysis must include technical threats, human mistakes, and power dynamics that encryption alone cannot solve.

Governance that learns#

Work begins before launch and continues afterwards. The organisation should define purpose, limits, owners, training, and resources for support routes. It should test the service with diverse people and examine access, abandonment, and complaint data without turning evaluation into new surveillance. Incidents must reach people able to change forms, contracts, messages, and protocols, rather than remain buried in an indicator dashboard.

A simple way to report a problem should be published, with responses within reasonable times. Affected people need to know what happens after a complaint and how to request human review. Providers must also be accountable: a contract does not remove responsibility from the organisation that invites use of the tool. Assessing providers includes accessibility, security, data processing, subcontracting, explanations, and real mechanisms for stopping harmful use.

Indicators that do not hide anybody#

Downloads, sessions, and completed questionnaires are easy to count, but they can hide pressure, abandonment, or exclusion. A more honest report combines quantitative measures with qualitative evidence: which barriers were reported, which alternatives were used, how long a human answer took, and what changes followed. It also separates voluntary participation from participation required by an institution. Success is not making everyone leave data; it is ensuring that a person needing support can find a safe, appropriate option.

Evaluation should look for unintended effects. Did an alert reach somebody who should not see it? Did a translation alter meaning? Did people with lower connectivity receive a worse response? Did declining participation have consequences? Such questions prevent an average from erasing important experiences. An exception becomes a learning signal when it can be reported without fear and has real power to change the system.

Digital help with clear limits#

A responsible platform states its limits in direct language. It can offer general education, voluntary tools, and contact routes; it cannot determine a clinical condition, guarantee safety, or replace a professional relationship. Nor should it use certainty that makes a person feel obliged to follow a recommendation. Clarity permits freer choice and helps distinguish general orientation from individualised care.

Quality appears in details: an accessible alternative exists before it is requested; sensitive data are not sought by habit; an alert reaches a competent person rather than a wide list; refusal brings no penalty; a possible emergency receives a clear route to immediate support. Those details reveal whether technology organises care or merely organises information.

Institutional and primary sources: World Health Organization, mental health; Spain's Ministry of Health, 024 line; Spanish Data Protection Agency, data-protection rights; United Nations, Convention on the Rights of the Child; UNESCO, Recommendation on the Ethics of AI; NIST, AI Risk Management Framework.