How one message becomes a reply
marriageAI answers every customer message on WhatsApp. Each message makes one pass through the flow below: the agent loads everything it knows about you, picks the right "brain" for the job, optionally uses tools, and every reply leaves through one door — in one consistent voice.
flowchart TD
IN["📱 Customer message arrives"] --> CTX["Load context
12 sources gathered at once"]
CTX --> SHORT["Quick checks
(no AI needed)"]
SHORT --> ROUTE["The router
classifies the message"]
ROUTE -->|"greeting or FAQ"| CHEAP["Cheap brain
fast + inexpensive"]
ROUTE -->|"building a profile"| ONB["Onboarding brain"]
ROUTE -->|"complex / sensitive / photo"| DEEP["Deep brain
most capable model"]
CHEAP -.->|"not confident"| DEEP
DEEP --> TOOLS["Tools
look up proposals, save fields,
send payment details… (max 10 rounds)"]
TOOLS --> DEEP
ONB --> R["Reply drafted
📼 episode recorded"]
DEEP --> R
CHEAP --> R
DEEP -->|"needs a human"| HUM["🙋 Hand over to staff"]
R --> VOICE["🎭 The formatter
ONE persona for every reply"]
HUM --> VOICE
VOICE --> OUT["Sent via whatsappGateway
the single door"]
classDef hi fill:#e7f2ec,stroke:#166b4e,color:#182420
class CTX,ROUTE,VOICE,OUT hi
One customer turn, end to end. The green boxes are the four load-bearing stages.
What the agent knows about you
Before thinking at all, the agent gathers 12 sources in parallel — one central "context builder", so every brain sees the same picture:
Who you are — account and profile, your journey stage,
engagement history, recent phone calls.
What was said — the recent conversation, including the
bot's own earlier messages, with one-time codes deliberately excluded.
What we've learned — a rolling summary updated every turn,
per-customer behavior notes, and three knowledge searches: company
policies, approved canned answers, and past human corrections.
What you sent — photos are classified automatically.
Five brains, one budget rule
Not every message deserves the most expensive model. The router assigns each message a tier — the simplest capable brain wins:
| Tier | Handles | Cost profile |
|---|---|---|
| DEEP | complex, sensitive, photos, safety topics | most capable model |
| DEEP-LITE | simple general questions | same brain, lighter model |
| CHEAP CHAT | greetings | fastest, cheapest |
| FAQ | common questions, grounded on the knowledge base | fastest, cheapest |
| ONBOARDING | profile-building conversations | mid-range |
A cheap brain that isn't confident escalates itself to the deep one — the customer never sees the handoff.
One voice, enforced in one place
Every reply path — deep, cheap, onboarding, even the message sent when handing over to a human — converges on the formatter before sending. It applies the persona: tone, language mix, spelling rules, message splitting. Change the persona once, every reply changes. The persona text itself lives in the database and is editable from the Dashboard — never hard-coded.
The flight recorder
Every reply writes an episode: the message, the reply, the agent's reasoning, which tools ran and what they returned, which brain handled it, and what it cost. Staff can open any bot message in the Agent Inbox and see exactly why the AI said what it said. Episodes also feed the learning pipelines — scored nightly for outcomes like "did the customer go silent?" and "did they complete a profile afterwards?".
When the bot steps aside
Two independent switches silence the bot for a customer: an escalation by the agent itself, or a staff member pressing "Take over from AI" in the Dashboard. Either one is enough. Handover reaches the Agent Inbox instantly through the shared database, and the bot stays silent until staff resolve the conversation.