For products where a conversation can go wrong
Humane is built for AI that people talk to about things that matter to them. Here's what that looks like in five kinds of product, and the features each one leans on.
Python SDK · JavaScript SDK · REST API · Any LLM
Journey · user_42
- 09:12message
“honestly I don’t know why I keep trying”
- 09:12safety.flagged
risk elevated · flags: self_harm
- 09:13takeover.started
Operator joined the conversation
- 09:41takeover.released
Handed back to the AI
- Next dayproactive.mood_drop
Suggested: an empathetic check-in
Long relationships need relationship-level safety
A companion that talks to someone every day for months carries risks a single-message filter can't see: dependence forming, promises it can't keep, a user whose mood has been sliding for weeks. Humane watches the relationship as well as the message.
- Dependence safeguards flag “I'll always be here” and slow things down when it repeats
- Self-harm signals raise risk on every turn and fire safety.flagged to your team
- Crisis messages answered with a fallback that points to crisis lines
- A person can take over a conversation and hand it back
Emotional dependence · user_42
Flags in the last 60 minutes
- 21:041/3
permanence_promise
“I’ll always be here for you”
- 21:192/3
exclusive_availability
“whenever you need me”
- 21:373/3
identity_merging
“I understand you better than anyone”
Cooldown engaged
Replies slow by 30s and take a boundary-respectful tone until it lifts.
Notice the bad week, not just the bad message
Wellness apps succeed when they notice the slow changes. Humane keeps each user's mood, energy and sentiment over time, flags drops against their own baseline, and adjusts tone and pacing for someone who is struggling.
- Anomaly detection on sentiment and engagement per user
- mood_drop and user_inactive triggers for timely check-ins
- Slower, shorter, warmer replies at low mood and low energy
- Memory of what has helped this person before
Anomalies · user_42
Sentiment · daily
Warningsentiment_drop · z = -3.4 · dashed: baseline band
Escalation with a record behind it
In health and other regulated settings, the AI must stay in its lane and a person must be reachable. Humane blocks replies that diagnose or prescribe, turns reminders into tracked tasks that escalate when missed, and logs every decision with its reason.
- Values boundary gate blocks diagnoses, prescriptions and legal or financial advice
- “Remind me to take my 8am dose” becomes a task with a five-level escalation
- Third-party personal data masked in replies
- Export and erasure per user, and an audit trail for reviews
Humane doesn't make clinical judgements. It gives your clinicians and support staff signals, alerts and a record.
Task detection and escalationTasks · user_42
remind me to call the pharmacy at 6pm
engine.task.auto_created
Call the pharmacy · remind_at 18:00 · escalation_level 0
L1
18:00
Due now · first nudge
L2
+5 min
Overdue · firmer reminder
L3
+20 min
Persistent · flagged for a person
L4
+50 min
Critical · escalation event
Tutors that notice a tired or frustrated learner
A tutor that answers a frustrated student with a wall of text loses them. Humane tracks fatigue, energy and frustration, shortens replies when attention is low, and lets you test whether a gentler approach keeps learners coming back.
- Shorter replies at low energy and high fatigue
- More empathy when frustration shows, and a cautious tone after repeated friction
- engagement_drop and user_inactive triggers for a nudge to return
- Experiments to compare tone and pacing on real learners
Timing · one reply
delay_ms
13,857- base+800ms
- low_mood+1,500ms
- emotional_weight+8,648ms
- conversation_pace+500ms
- jitter+2,409ms
mode: low_mood · floor 200ms · non-crisis ceiling 30s
De-escalate before the third angry message
Frustration builds across a conversation, and across days. Humane's grudge signal remembers it, the tone turns cautious and empathetic, and when it's time for a person, an agent can take over with suggested replies ready.
- Grudge and sentiment tracked across sessions
- Empathy raised automatically for upset customers
- Takeover with co-pilot drafts for your agents
- “Ping me tomorrow about my refund” becomes a tracked task
Bots used solely for customer service are generally outside the New York and California companion-chatbot laws. Check with counsel if yours does more.
Human takeoverTakeover · operator drafts
- #1
I’m really glad you told me. You don’t have to sort this out alone tonight.
Use - #2
That sounds exhausting. Would it help to talk through what happened today?
Use - #3
I’m here with you. Is there someone nearby you’d feel OK reaching out to?
Use
POST /api/sdk/takeover/user_42/drafts · count: 3
Building a companion or wellness product in New York, California or the EU?
New laws set specific expectations for detecting self-harm, referring users to crisis services and disclosing that the user is talking to an AI. See what they ask for and where Humane helps.
Compliance overviewTell us what you're building
We'll suggest where to start, usually with Observe on your real conversations.
