How we use AI

The Guide is deliberately narrow. It may use the direction you said you want, anything you told us you do not want, how you said you are feeling in this session, which practices you opened, completed or left, whether you said they helped, meanings you have explicitly confirmed, and how long it has been since your last visit.

It may not read free text from other sessions, anything you have not confirmed, anything belonging to another member, or anything from our internal business systems.

• Your reflections and Guide conversations. What you type or dictate to your Guide, your first experience answers, and the meanings you confirm.
• Voice recordings for dictation. The short audio clip captured when you press the microphone.
• Unconfirmed suggestions. A possible pattern your Guide has noticed but you have not agreed to yet.

Your name, email, billing details and session bookings are not sent to any AI provider.

Every AI call goes to one place, the Lovable AI gateway, using a single key we hold. We do not hold accounts with model vendors directly. The gateway then routes the call to a model: Google Gemini for language and one text to speech voice, and OpenAI for speech to text and an alternative voice.

That means our AI processor of record is the gateway, and the model vendors sit behind it as onward processors.

The microphone only records while you hold or toggle it. The clip is uploaded once, turned into text, and the text lands in the box so you can edit it before you send it. We do not store the audio in our database. Your Guide's spoken replies are generated from text each time and are not saved as files.

Whether something is treated as a safety concern is decided by fixed rules that run before the model is called. A model can raise a concern but can never clear one. When a concern is raised, the conversation stops and approved wording plus verified local crisis resources are shown by our own code. The model does not choose that wording and cannot invent a helpline.

Model output is also checked after generation. Diagnosis language, claims about cause, promises of an outcome, invented studies or numbers, and any suggestion that a human is reading your conversation are removed rather than shown to you.

Nothing the Guide does decides anything about your employment, your eligibility for care, your insurance or your money. It suggests what to do next, tells you why it suggested it, and always offers an alternative. You can ignore it.

We will not tell you your entries are never used to train a model until a vendor has told us that in writing. These are the open questions.

• How long our AI gateway retains prompts and outputs.Awaiting written confirmation from the vendor.
• Whether any model provider in the chain uses our prompts to train models.Awaiting written confirmation. Until we have it we do not claim that your entries are never used for training.
• The exact country where each model call is processed.Awaiting written confirmation. Assume the United States today.
• Independent security certification such as SOC 2 for our own platform.Not held. Formal assessment is planned and will be published here when complete.
• Data residency in the EU, UK or elsewhere outside the United States.Not available today. All processing described here happens in the United States.

When we have those answers this page is updated first, before any marketing material repeats them.