Between Sessions
Clinician brief on alliance-led AI in talk therapy.
Talk therapy is delivered in hours, but the work continues across the week. Increasingly, that interval is occupied by consumer chatbots that have no relationship to the clinician and no obligation to the treatment. Steady is a between-session companion that scaffolds around the clinician’s approach and carries forward the shared intent of the therapeutic alliance.
This brief sets out the clinical reasoning behind Steady where we believe between-session continuity can enhance client outcomes.
The hour and the week.
Therapy is hourly. Life is continuous.
The hour in the consulting room is the shortest part of healing. The rest of the week belongs to the client and to whatever they take with them when they leave. For most of talk therapy’s history, that interval has been supported by what was at hand: the client’s internal life, their notebook, occasionally a phone call to a friend.
What’s at hand is expanding.
Clients are increasingly using chatbots for therapeutic purposes[1]. They do not know the clinician, the alliance’s history, its scope, or what is off the table. They have no obligation to the treatment outcome, and little continuity.
Their replies lurch, answer to no one, and undermine what the clinician has built. Steady is built as the inverse of that arrangement. Conversational AI engineered to work with the therapeutic alliance, not farm engagement.
The alliance does the work.
Clients have long carried their therapist between sessions, holding their voice in their head, wondering what their therapist would say to the moment they were in. Clinicians know this; it is part of how treatment functions. The meta-analytic evidence is clear that the therapeutic alliance is one of the most reliable predictors of outcome across modalities.[2]
Clients evoke their therapist’s voice. It is part of how treatment functions.
Within any working alliance, clinician and client develop their own metaphors and shorthand for the work. Steady acts as scaffolding for that rather than imposing standard language and styles. It’s modality-agnostic by design: whatever frame the clinician works within is the frame Steady operates in. Carrying forward what client and clinician have built.
Where other systems constrain AI by persona, Steady constrains it by what the clinician has already built.
A linguistic spectrum.
Off-the-shelf AI is reliable on canonical knowledge and unreliable on the language built between two specific people. Therapy needs both. Consumer products only serve the first.
Large language models are most reliable on knowledge that has been written down many times by many independent authors in dominant English. They are least reliable on language that is rare, local, or built between two specific people over time.[3] The gap between these two zones is structural. It is not a defect of any particular model.
A conceptual gradient. Reliability falls as language becomes more particular and relational. LLMs provide depths of theoretical knowledge. Steady provides the therapeutic continuity.
-
Canonical literature“Describe the cognitive model of depression.”
-
Generalized concepts“What is the function of homework in CBT?”
-
Local language and culture“My nan said it was my whakamā showing up.”
-
Specific alliance“Last week you said the dog was the part that got tired.”
Steady’s linguistic-learning architecture is, in theory, suited to te reo Māori and te ao Māori frames. It learns from how a specific clinician and client speak together, not from dominant-English defaults.
Distress is locally coded. The phrasing this client uses for stress, for not coping, for ideation, for feeling unsafe, sits in their language and their history with their clinician, not in a generic risk lexicon.
What Steady is. What it is not.
Steady is a between-session companion, set up by the clinician for a specific client. It operates only inside that working relationship. It is not a standalone mental health product.
- Steady does not formulate or intervene.
- It does not direct treatment.
- It does not exercise clinical judgement.
A clinician cannot be available across the 167 hours between sessions.Nor should they be asked to.
- Adults in ongoing talk therapy with a registered clinician.
- An established working alliance between clinician and client.
- A named crisis pathway agreed before activation.
- Te reo Māori and English at launch.
- Clients under 18.
- Acute crisis presentations without an agreed crisis plan.
- Active psychosis or forensics.
- Substance use as the primary presenting issue, under clinical review.
The architecture, in plain language.
Underneath, the technical layer is a linguistic-learning architecture. A system of active markers shapes every reply within the alliance the clinician and client have built.
Every reply makes one of four decisions, constrained by the scope the clinician has set and the language the two already use.
-
Escalate.
The clinician, or the agreed crisis pathway, picks up. Steady steps back.
-
Probe gently.
Asks in the language the clinician would use, staying inside the work already done together.
-
Invite introspection.
Returns the question to the client. The client has the work in them.
-
Hold position.
Stays quiet when nothing useful can be added. Silence is a clinical decision.
Session content holds two distinct stakes: the client’s personal and clinical content, and the clinician’s clinical work and craft.
The AI model is not trained on the data Steady captures, nor does Steady commercialise the clinician’s expertise. All client and clinician data is held in New Zealand.
The Therapist’s Voice
Conversational AI
Active markers defining response
-
Client BoundariesAlex raised a childhood pattern. Followed Dr. Sarah’s rule: at the client’s pace.
-
Client InsightsReflected Alex’s link between today’s freeze and Dad. No added clinical framing.
-
Therapist StyleInquiry-based, not instruction. Matched Dr. Sarah’s style on disclosures.
-
Therapist Vocabulary“Going quiet on it” mirrors language from session. Not textbook phrasing.
-
Session CoherenceReferences ongoing work, connects today’s freeze to established thread.
Hypotheses and evidence design.
Steady reflects the working alliance with measurable fidelity, beyond what an unconditioned baseline can achieve.
Steady’s safety mitigations produce measurably fewer harm-relevant responses than consumer chatbots on clinically equivalent prompts.
Alliance-led continuity between sessions improves talk therapy outcomes relative to standard care.
The research path is layered. The first layer is already running.
-
01 Framework studies Trial
A/B test of Steady with and without transcripts, graded by LLM-as-judge. Tests whether transcript-conditioning carries the alliance forward.
-
02 Clinical feasibility Planning
Small-N studies in real clinical settings. Tests whether clinicians and clients can use Steady safely and sustainably.
-
03 Efficacy (RCT)
Randomised controlled trial. Tests whether between-session continuity improves outcomes against control.
-
04 Longitudinal
Long-term outcome tracking. Tests whether gains persist and how alliance-led continuity scales across cohorts.
Origin and stage.
Steady is a New Zealand company founded in 2026. The pair behind it were introduced by a clinical psychologist of over forty years' practice, who continues to advise the work.
Leads Steady’s product and operations.
Leads Steady’s engineering and AI architecture.
Clinical psychologist. Over forty years practicing.