The signals therapist software was never built to see

Every therapy session produces clinical signals about alliance, symptoms, and progress. The entire industry builds tools that ignore them. Here's what happened when I designed a system to surface what notes were never built to capture.

June 2, 2026·6 min read

A therapist told me something I didn't expect during a product feedback call. She wasn't talking about time savings or compliance formatting. She was describing what it felt like to read an AI-generated account of her own session.

"It named an intervention I'd been using for years without knowing what it was called."

She meant it literally. The AI analyzed the session from a third-person vantage point and surfaced clinical observations she hadn't written down, because she couldn't see them from inside the conversation. She was learning from her own documentation.

I kept hearing versions of this from other clinicians. And it raised a question I hadn't planned for: if a note generated from a therapy transcript could teach a therapist something about their own session, what else was hiding in the conversation?

The documentation assumption

The mental health tech industry treats therapy sessions as documentation events. The dominant product category is AI scribes: transcribe, summarize, generate a note. Practice management platforms handle scheduling and billing. The value proposition across the category is the same: reduce the admin burden.

It's a real problem. Therapists spend 12 to 15 minutes per note on formats borrowed from primary care. Administrative work drives 82% of behavioral health clinician burnout. Faster notes are genuinely valuable, and at Mentalyc, therapists told me the AI-generated notes were transformative because they finally got documentation formatted the way insurers needed it, without spending their evenings writing.

But the assumption underneath all of this is that the useful output of a therapy session is the note. This is the same kind of comfortable fiction I encountered mapping 60+ pain points on a regulated platform: the team's mental model of the product hides an entire layer of problems. And once you question that assumption, the product opportunity changes completely.

Three signals no one was extracting

When I started researching progress tracking at Mentalyc, I found that therapy transcripts contained far richer clinical signals than what was being used for note generation. Working with our UX research partner, we interviewed therapists about how they evaluate client progress. Three signals consistently emerged.

Alliance. The therapeutic relationship between clinician and client. Research consistently identifies it as the strongest predictor of therapeutic outcomes. Therapists track it intuitively, but no tool captured it.

Symptoms. Not diagnoses, which are static labels. Symptoms: the evolving patterns of distress, avoidance, mood shifts, and behavioral change that actually move during treatment. Clinicians told us that symptom signals were more meaningful than diagnostic categories because they reflected how clients were actually changing.

Goals. Therapy goals are frequently symptom-based or ability-based: reducing anxiety, improving emotional regulation, increasing social engagement. By linking symptom signals to treatment goals, you could track therapy progress across sessions rather than session by session.

These three signals formed the foundation of clinical reasoning in therapy. And almost no tool in the market was designed to surface them.

Building the detection layer

The hard part wasn't identifying the signals. It was building a system that could detect them reliably inside real therapy conversations.

On the symptom side, we needed a structured taxonomy. Therapists don't think in diagnostic codes during sessions. They notice recurring anxiety language, expressions of hopelessness, mood shifts, mentions of risk. We extracted a detection framework from DSM-5-TR diagnostic criteria, mapping the specific conversational markers that could be identified in therapy transcripts. This let the system track how symptoms evolved across sessions rather than flagging them once and moving on.

The diagnosis-to-symptom shift was a deliberate product decision. Diagnoses are assigned at intake and rarely revisited. They're static. Symptoms are what actually move during treatment, and they're what therapists use to gauge whether therapy is working. Designing around symptoms instead of diagnoses meant the system tracked what clinicians actually cared about, not what the administrative layer required.

Connecting symptom tracking to treatment goals created something none of the existing tools offered: a dual-value system. Therapists got clinical intelligence about how therapy was progressing, grounded in how progress actually works rather than how insurance companies model it. Insurance got what they already needed: evidence of movement toward diagnosis. Same clinical data, structured to serve both directions.

Every insight the system surfaced was anchored to a specific moment in the transcript, with verbatim quotes. This wasn't optional. In clinical software, trust scales with transparency. If a therapist can't see exactly where an observation came from, they won't act on it. The evidence model had to be traceable back to the conversation or it was worthless.

The gap in the landscape

Across the landscape of therapist tools, I found a consistent pattern. Practice management platforms like SimplePractice and TherapyNotes treat sessions as billing events. AI documentation tools like Freed and Heidi reduce admin work but provide no clinical reasoning support. Even hybrid platforms like Blueprint and Upheal offer limited visibility into what's actually happening inside therapy conversations.

The missing layer is clinical insight: tools that analyze the conversational dynamics of therapy sessions and surface the signals therapists already use in clinical reasoning, but have never had captured systematically.

When we presented the concept at the Evolution of Psychotherapy Conference in Los Angeles, the response confirmed the gap was real. Goal tracking became one of the most highly rated features during validation. Clinicians weren't asking for faster notes. They were asking for a system that helped them see what was happening across their caseload over time.

Mentalyc later shipped the framework as two public products: Alliance Genie, an AI tool that measures therapeutic alliance by analyzing over 30 psychosocial markers from session recordings, and the AI Progress Tracker, which surfaces symptom trends and goal attainment across sessions without additional clinician work. The insight layer I'd designed as a research question became two shipped features.

What the note was hiding

Every therapy session produces a conversation rich with clinical signal. Alliance dynamics, symptom patterns, progress toward goals. These signals are the substance of clinical reasoning, and they're sitting inside transcripts that most tools process only for documentation.

The architecture required to surface them (symptom detection from DSM-5-TR criteria, goal-linked progress tracking, transcript-anchored evidence) isn't a feature set you bolt onto a documentation product. It's a different product category. The industry mapped therapist software along a single axis, from manual documentation to automated documentation. The higher-value axis was always clinical intelligence.

The most valuable thing a therapy session produces isn't the note. It's what the note was never built to capture.