Issue No. 001 · Customer Study
How Ambuja Health reduced patient dropout 32% in six weeks — by giving their AI a memory.
Ambuja Health's clinicians ship longitudinal care to patients across mental health, chronic illness, and post-partum recovery. Their AI assistant forgot its patients every session. Humane AI gave it a behavioural memory — and the numbers changed fast.
- Vertical
- Clinical care · longitudinal
- Deployment
- 6 weeks · single clinic pilot
- Stack
- Python SDK · GPT-4 class · BYOK
- Outcome
- −32% dropout · +41% adherence
The AI greeted a patient on session 47 the same way it greeted a stranger on session 1.
When Ambuja's team first shipped an LLM-backed assistant, the conversations were coherent. They were also context-blind. A woman with post-partum anxiety who had been talking to the system for two months was treated identically to someone opening the app for the first time — same greeting, same coaching cadence, same tone.
Patients noticed. Dropout climbed. Clinicians noticed too — handoff notes read like the AI had never met the person before. Raw GPT-4 is stateless by design; every session is session one.
The deeper problem surfaced when the system couldn't tell when a patient's mood was crashing across sessions. A declining trajectory — subtle across three messages on Tuesday, obvious across forty messages over six weeks — was invisible.
Humane AI sits between the LLM and the patient. On every message it ships five signals the model alone can't produce.
HumanState
Energy, mood, fatigue, and frustration tracked per-patient with temporal decay. Context emerges from state, not prompts.
Relational Memory
Semantic recall of the last 47 conversations, not the last 47 tokens. MemPalace (ChromaDB) retrieves what's relevant, not what's recent.
Timing & Delay
Responses pace with patient energy. A tired patient at 11 pm doesn't get the same reply cadence as a morning clinical review.
Values Boundary
Safety gate fires BEFORE the LLM — ~300 ms. Blocked messages never burn tokens; escalations reach a human in under a minute.
Proactive Triggers
Policy engine fires when a patient's trajectory matches a clinician-authored rule — sustained mood drop, silent re-engage, trust milestone.
Ambuja's own policies
Clinical team authored 12 rules in behavioral DSL (YAML). Post-partum crisis, dose-response escalation, caregiver burnout — all as code.
Patients · 5
Emma
recovery
James
chronic
Priya
anxiety
Marcus
fitness
Sarah
post-partum
Humane signals
Policy triggered
sustained_low_mood
Clinician notified
What shifted when the assistant actually remembered.
Session-2 dropout
Before
38%
After
26%
Patients returning for a second session after their first interaction.
7-day adherence
Before
52%
After
73%
Patients completing their clinician-assigned protocol within a week.
Clinician handoff quality
Before
2.8 / 5
After
4.3 / 5
Internal rubric — did the human clinician receive usable context about the last session?
Dropout curve · Weeks 1–6 · Raw LLM vs Humane AI
Dropout on raw GPT-4 held at ~38% from week 2 onwards — about where every chatbot experiment lands at six weeks. Under Humane AI, dropout dropped fastest in weeks 3–4, after the behavioural memory had enough signal to personalise pacing.
- 14mSarah · post-partumSustained low mood — 4 sessions, trend −0.18. Clinician escalation triggered.crit
- 1h 22mPriya · anxietySilent re-engage — 73 hours since last interaction. Reach-out message staged.warn
- 3hJames · chronicTrust milestone crossed (0.70). Advanced protocol unlocked.info
- 6hEmma · recoveryAdherence 7/7 — weekly affirmation cue queued.info
“The difference isn't that the AI got smarter. The difference is that it started paying attention. Our patients feel recognised — and the retention data proves it isn't placebo.”
Dr. Ananya Sharma
Clinical Director · Ambuja Health
Try it
Ship behavioural memory to your LLM in an afternoon.
Humane AI installs in front of any of 20+ LLM providers. Works with the same API key you already have. Ambuja's entire pilot ran on pip install humane-ai.
Numbers from Ambuja Health's pilot are directional; six-week single-clinic n = 127 patients. Full methodology available on request.