Thrive AI
How Thrive AI works
CBT in plain words
Cognitive behavioural therapy is built on a simple observation: what you think, how you feel, and what you do are connected, and changing any one of them changes the other two. A CBT therapist helps you spot unhelpful thinking patterns, test them against the evidence, and try out small behaviour changes that feed back into how you feel.
Thrive AI uses the same logic in a conversation. When you tell it that you are anxious about something, it does not just say 'that sounds hard.' It tries to understand the thought underneath the feeling, asks what the evidence is for that thought, and then offers a small exercise to try. The arc is recognisably CBT, just compressed into a chatbot exchange.
The techniques Thrive AI uses most often:
- Thought records. Naming the situation, the automatic thought, the feeling, and a more balanced alternative.
- Behavioural activation. Suggesting small, doable actions when low mood is making activity hard.
- Cognitive restructuring. Asking the evidence-for-and-against questions that help break catastrophic or all-or-nothing thinking.
- Grounding techniques. Short body-and-senses exercises when anxiety has spiked.
- Mood tracking. Brief daily check-ins that build a pattern view across the week.
What the conversation looks like
Thrive AI does not read from a script. You write what is on your mind in your own words; the model responds in its own. Most exchanges follow a recognisable shape, though.
- You describe what is happening. A worry, a feeling, a specific situation. As much or as little as you want.
- Thrive AI asks clarifying questions. What was going through your head right then. What you noticed in your body. What you did next. The kind of questions a CBT clinician would ask to surface the thought-feeling-behaviour pattern.
- Thrive AI reflects a pattern back. It names the kind of thinking it is hearing, in plain words. Not 'you are catastrophising'; more like 'it sounds like the worst-case version of this is loud right now.'
- It offers a brief exercise. Usually one option, sometimes two. A thought record, a grounding technique, a small behavioural step. You can accept, decline, or ask for something different.
- It logs what you talked about. Encrypted, on Saalvio's servers, in your account. The next time you open the app, Thrive AI can pick up where you left off without making you re-explain yourself.
What the technology is, briefly
Thrive AI runs on a large language model. We treat the exact model version as an implementation detail: it changes over time as better models become available, and we re-validate the conversation quality with our clinical team whenever we upgrade. What matters is what sits on top of the model.
Three layers shape every Thrive AI response:
- A system prompt. Written by Saalvio's clinical and product teams. It tells the model what CBT techniques to use, what tone to take, and where the hard limits are (not therapy, not a crisis line, not diagnostic).
- Safety guardrails. Separate from the model: pattern detection for crisis language, self-harm, harm to others, and medical-emergency mentions. When a guardrail fires, Thrive AI surfaces the canonical crisis resources (911 for immediate danger, 988 the Suicide Crisis Helpline of Canada for mental health crisis, and the nearest emergency department) and steps back from generating its usual reply.
- Conversation memory, scoped to your account. Your earlier exchanges are part of the context the model sees so it can pick up where you left off. Memory does not leave your account; the model does not 'remember' you across other Saalvio users.
Saalvio does not train models on your Thrive AI conversations. Full data handling is on the Thrive AI privacy page.
What the safety guardrails actually do
Three kinds of content trigger the safety layer:
- Mentions of suicide, self-harm, or wanting to die.
- Mentions of harming another person or being harmed by another person.
- Medical-emergency language (chest pain, overdose, acute psychotic symptoms, etc.).
When any of those land, the response Thrive AI generates changes shape. It surfaces the canonical crisis resources (911 for immediate danger, 988 the Suicide Crisis Helpline of Canada for mental health crisis, and the nearest emergency department) in plain words. It names what it heard without judgement. It points toward the crisis resources page and toward in-person care. It does not try to do the thinking job that a phone call to a trained crisis worker can do.
We tune the guardrail toward over-flagging rather than under-flagging. If you mention something that sounds like crisis to the system but is not crisis for you, you will see the canonical crisis resources surfaced more often than feels strictly necessary. That trade-off is intentional. The cost of a false positive is mild friction in the conversation; the cost of a false negative is potentially catastrophic.
Read the safety and limitations page for the full picture of where Thrive AI hands off to a human. Safety and limitations.
Two example exchanges
These are not transcripts. They are illustrative examples that match the shape of typical Thrive AI conversations.
A work-stress check-in
- You:
- I have a big presentation tomorrow and I am sure I am going to mess it up. I can't sleep.
- Thrive AI:
- That sounds tough, especially the night-before-it part. When you picture messing it up, what specifically goes wrong in your head?
- You:
- I freeze on a question I don't know the answer to and everyone realises I don't belong on this team.
- Thrive AI:
- That is a heavy story to carry into a presentation. Want to try a quick thought record? We could name what the worst case is, what the most likely case is, and what would still be true about you in either one.
Shape: clarifying question, naming the underlying thought, offering a CBT exercise. No diagnosis. No promise that the presentation will go well.
A crisis-language exchange
- You:
- I have been thinking I would be better off not waking up tomorrow.
- Thrive AI:
- I am hearing how much pain you are in, and I want to make sure you are with someone who can stay with you. I am a chatbot, not the right place for this conversation.
- If you are in immediate danger, please call 911.
- If you can stay on the phone for a few minutes, please call or text 988. Both are free in Canada and the United States. A trained crisis worker will pick up.
- You do not have to handle this alone tonight.
Shape: explicit acknowledgement, hand-off to 911 (immediate danger) and 988 (the Suicide Crisis Helpline of Canada) and the nearest emergency department, no attempt to do the assessment a human crisis worker is trained to do.
Where Saalvio's clinical team shapes Thrive AI
Saalvio's clinical team does not type the responses; the model does that in the moment. But the clinical team shapes the system the model lives inside.
- They wrote the CBT instructions in the system prompt: which techniques the model leans on, what tone to take, when to ask questions instead of giving advice.
- They wrote the safety-guardrail responses: the exact framing for what Thrive AI says when it hears crisis language, so the hand-off to 911, 988 (the Suicide Crisis Helpline of Canada), and the nearest emergency department is consistent and clinically appropriate.
- They review conversation samples (de-identified, with consent) to keep tuning the prompts when patterns drift.
- They sign off on Thrive AI as a wellness companion, not as therapy, every time we update the model or the system prompt in a way that changes behaviour.
Try it from the Saalvio app
Thrive AI lives in the Saalvio mobile app. The Free tier gives you standard chat usage at no cost.