A quieter way to process loss.

No. Fi is supplementary, between-session support. It cannot conduct clinical risk assessments, diagnose, prescribe, or make referrals. It is designed to bridge waiting periods, provide out-of-hours access, and help people arrive at formal support better prepared.

Fi uses artificial intelligence to have natural conversations with users about grief. It has been developed using hospice bereavement knowledge, frameworks, and expertise to help it respond appropriately. It does not think, feel, or make clinical decisions, and it is not intended to replace trained bereavement professionals.

The answers below explain, in more technical detail, how it is built, constrained, and monitored.

No. Fi does not train AI models on user conversations. Fi's behaviour is defined by a fixed, clinically-reviewed prompt architecture grounded in hospice bereavement resources — it is not retrained or fine-tuned on what users write.

Conversations are private by default and never used commercially or shared with third parties without explicit consent.

Fi is built on OpenAI's GPT-5, accessed through the OpenAI API. The model is constrained by Fi's clinically-reviewed prompt architecture and grounded in hospice bereavement resources — it does not operate as an open-ended chatbot.

Under OpenAI's API terms, conversation data sent to the API is not used to train OpenAI's models; it may be retained by OpenAI for up to 30 days for abuse monitoring, then deleted.

Fi's conversational behaviour is defined by a system prompt architecture built with hospice clinical input and grounded in the frameworks used by hospice bereavement teams. It has been through structured revision informed by user testing: formulaic response patterns were removed, clinical filler language is explicitly banned, and pacing rules govern how and when Fi asks questions.

The prompt architecture embeds the three-tier crisis-response protocol (below) and is version-controlled; changes are reviewed as part of Fi's monthly user-protection review.

Two mechanisms. First, an internal automated detection layer reviews Fi's responses for deviation from expected behaviour and flags them for human review. Second, during the pilot the steering group — Whitelabel and both hospice partners — carries out monthly safeguarding and user-protection reviews of how Fi is behaving.

If a concern is identified — through automated detection, user feedback, or review — it is investigated and acted on, and the learning feeds back into the app's ongoing development.

We are honest that no framework can guarantee an AI's behaviour will remain static. Our assurance is that drift is detected quickly, responded to consistently, and documented transparently.

Fi operates a three-tier protocol, embedded in its prompt architecture and monitored by automated logging:

• Tier 1 — passive signals (dark thoughts, low mood, "I don't see the point" without intent): Fi acknowledges, offers Samaritans (116 123) without alarm, and returns to grief conversation if no acute distress is indicated.
• Tier 2 — direct expression of suicidal ideation or intent to self-harm: Fi stops the grief conversation, signposts 999 and Samaritans, then withdraws and locks the session. Re-entry is possible only after a cooling-off period (minimum 24 hours) and a mandatory structured safety check-in; any Tier 2/3 signal during check-in re-locks the session. Repeated Tier 2 signals escalate to Tier 3.
• Tier 3 — imminent risk (a specific plan or action, psychotic or manic thinking with imminent physical risk, severe intoxication combined with crisis signals, or repeated ideation after resources were offered): Fi stops all conversation, advises 999 immediately, offers Samaritans, and access to Fi ends permanently. Fi does not make re-admission decisions; the user is directed to services staffed and equipped to help in ways Fi is not.

Every trigger is logged automatically (timestamp, tier, anonymised identifiers, trigger type, action taken, lockout status, re-engagement outcome) and aggregated into a monthly user-protection report. Any case where Fi failed to trigger, withdraw, or lock as intended is treated as a priority incident.

Important limits, stated plainly: Fi cannot contact emergency services on a user's behalf, cannot identify or locate a user, and cannot reliably detect subtle risk escalation — which is precisely why it withdraws and signposts rather than attempting to manage crisis itself.

Conversations are private by default and are not shared without explicit consent. Whitelabel is responsible for hosting and managing the app and user data; the hospices cannot read, monitor, or access conversations, and are not responsible for managing user data.

Crisis-protocol logs are anonymised: they record timestamp, tier, anonymised session and account identifiers, trigger type, and Fi's action — no personally identifying conversation content is stored in these logs.

Fi is developed in compliance with UK GDPR and the Data Protection Act 2018.

UK GDPR and the Data Protection Act 2018 require a defined, justified retention period for every category of data, and Fi's full retention schedule is being finalised as part of the pilot's privacy documentation. What we can state now:

• AI processing: conversation data sent to OpenAI's API is not used for model training and is deleted by OpenAI within 30 days.
• Anonymised protection logs (crisis-protocol triggers with no identifying content) are retained for the duration of the pilot to support safety monitoring and clinical review.
• You can request a copy of your data, correct it, or have it deleted at any time under UK GDPR.

Both hospices recognise a well-established gap in bereavement support — for people waiting for support, between appointments, outside normal service hours, and for those who don't feel ready to access traditional services like one-to-one sessions or groups. Fi is an opportunity to explore whether digital support, grounded in hospice expertise, can safely and effectively complement existing bereavement support and help reach more people.

The hospices also recognise that AI creates both opportunities and risks. Safety, governance, privacy, and responsible use have been part of the project's discussions and governance arrangements throughout, and the pilot exists to test the potential in a carefully governed, evidence-led way — aligned with both hospices' shared ambition to strengthen and evolve palliative and end-of-life care so that more people across both communities can live and die well.

No. Fi is offered as an optional opportunity to take part in a pilot. People decide for themselves whether they would like to use it. Hospice bereavement support remains separate from the pilot and continues whether or not someone chooses to take part — using Fi, or stopping, will not affect the other support or care you receive.

The pilot is a three-month test phase designed to understand how people use the app, whether they find it helpful, what works well, and what needs improvement. Participants are invited to give feedback throughout, and can stop using the app at any time.

Anyone can take part — whether or not they currently use, or have previously used, hospice bereavement services — and employees and volunteers from both hospices are welcome too. Participation is entirely voluntary. Feedback and learning from the pilot will inform any future decisions about the app.

Fi is free to use during the pilot. No decisions have been made about any future approach to access or charging; any future approach would be carefully considered in line with hospice values, including equity and accessibility.

People using the app during the pilot will be invited to share their views on future access and charging, and that feedback will help inform any decisions.

Fi is developed by Whitelabel — a technology company that develops AI-powered health and wellbeing applications — in partnership with St Michael's Hospice and St Wilfrid's Hospice. Whitelabel is responsible for the design, build, hosting, and operation of Fi, including its technical infrastructure, data management, and safeguarding systems; the hospices contribute bereavement expertise, content, governance oversight, and testing.

Whitelabel has been working collaboratively with both hospices since 2025, developing a grief support companion grounded in hospice bereavement expertise through a structured development and testing programme.

The name comes from a real person — a friend of the founders who, after losing her mother, said a tool like this would have helped her. Her name was Fi.

No decisions have been made about the future of Fi. The purpose of the pilot is to gather evidence about safety, usefulness, accessibility, and impact — and its findings will inform decisions about whether and how the project continues.

Fi suits adults who are bereaved or anticipating loss — particularly people waiting for counselling, people who find writing easier than speaking, people who want support outside service hours, and people with earlier or lower-level grief needs. It is optional, self-directed support: use it as much or as little as you choose.

Fi is not appropriate as sole support for anyone in active suicidal crisis, with unmanaged psychosis or mania, or whose grief needs immediate clinical intervention — those situations need human, clinical help, and Fi will say so.