What we build
Every channel, every language, every hour.
Each capability runs in production against your live enquiry stream — wired to your CRM, calendar, and inventory on day one.
Multi-channel intake
Web widget, WhatsApp Business API, and phone IVR unified into one inbox — every enquiry scored and routed through the same pipeline regardless of entry point.
In-conversation intent detection
Is this a qualified enquiry or someone asking for a brochure PDF? The intent classifier reads the thread continuously and only escalates when there's real sales signal.
Live language detection + reply
EN, BM, ZH, TH, ID, VI auto-detected mid-conversation. Language switches are honoured turn-by-turn — buyers never feel forced back into English to make progress.
Appointment booking inside the chat
Calendar-aware slot pitching against agent availability and project gallery hours — the appointment is booked before the thread closes, not after a callback.
Brand-voice tone matching
Trained on your developer voice — same cadence, same positioning, same approved phrases. Replies sound like your brand, not a generic chatbot.
After-hours coverage + escalation
The agent qualifies and books out-of-hours; warm leads are handed off to a named human agent the next morning with the conversation history attached.
Who runs this
Every channel, same intake engine.
Whether the enquiry lands in a web widget on launch weekend or a voice call on a Sunday afternoon — the intent classifier, slot extractor, and brand-voice generator are the same. Only the channel adapter changes.
New-launch EOI funnels
Launch-weekend enquiry floods handled without human bottlenecks. Every conversation qualified and gallery-slot-booked before a single agent manually replies.
Resale + lettings intake
Multiple listings, same intake engine — the AI knows which listing each enquiry is about, references photos, and books viewings in the chat.
Multi-country property portals
SG, MY, ID, TH, VN coverage with per-market language, phone number formatting, and business-hours logic baked into the same platform.
WhatsApp-first developer sales
Where WhatsApp is the dominant channel (SEA-wide), the AI picks up inside the BSP thread, handles media responses, and escalates to agents on pattern triggers.
Agency co-broke intake
Agency-rep enquiries recognised and separated from direct buyers — routed to the co-broke desk with commission splits and protocol rules attached.
Voice IVR replacement
AI voice handles phone enquiries end-to-end — qualifies, books gallery visits, and drops transcripts into the CRM with the audio recording attached.
Model families we deploy
Four specialised models. One seamless conversation.
Each model owns a layer — intent, slots, voice, escalation. Composing them per turn gives you reliability a single end-to-end model can't match.
Multilingual transformer classifying enquiry intent (information · qualification · objection · booking). Runs per-turn so pivots are caught inside the thread, not after it.
Structured extraction model pulling slot values from every message — each slot versioned per turn so an agent joining late sees the latest qualified view at a glance.
Reply generation trained on your approved scripts, brand book, and top-performing human conversations. Outputs cleared by guardrails before send.
Learns when to hand off — from objection signals, high-intent triggers, or after-hours qualification completion. Escalation rules retrain against agent feedback weekly.
Data sources wired into every conversation
Every channel, integrated — every reply, grounded.
Channel adapters, calendar sync, and inventory lookup all live in the same conversation loop — no hallucinated availability, no stale unit status.
Reasoning, not replies
A reply alone isn't enough. The chain of intent, slots, and handoff reasoning is.
Every conversation logs the intent classifications, slot extractions, and escalation decisions that produced the agent experience. Managers audit threads at scale, not one-by-one.
- Per-turn intent + slot state with confidence
- Source-of-evidence quote for every extracted field
- Escalation rationale with policy version attached
- Full transcript + audio with retention per jurisdiction
Frameworks we align to
Why Axccelerate for conversational intake
Not a chatbot.
An intake system.
Off-the-shelf chatbots answer FAQs. Our stack qualifies, books, and hands off — with reasoning trails and retraining against your sales outcomes.
Pricing
Priced to your channel mix, not per-message.
Every deployment is scoped against your channels, languages, and enquiry volume before we quote — no per-message trap pricing.
Glossary
The vocabulary behind every conversation.
A reference for the terms your sales team, compliance team, and conversation platform will all share.
- BSP
- WhatsApp Business Solution Provider
Meta-approved infrastructure provider for WhatsApp Business API (e.g. 360dialog, Wati, Infobip). All enterprise WhatsApp goes through a BSP — direct-to-Meta isn't an option for most use-cases.
- IVR
- Interactive Voice Response
The phone-menu system that used to say 'press 1 for sales'. AI IVR replaces it with natural conversation — buyers describe what they want, the AI qualifies and books without buttons.
- ASR
- Automatic Speech Recognition
Speech-to-text. Modern ASR (Whisper-tier) handles SEA accents, code-switching, and noisy environments — the accuracy jump since 2022 is what made AI voice viable for property intake.
- NLU
- Natural Language Understanding
The layer above raw text — extracting intent, entities, and slots from what a buyer says. Separate from generation (NLG). Our stack uses NLU per turn to decide handoff timing.
- Intent
- Classification of user goal
Per-turn label like 'information-request', 'qualification-in-progress', 'objection', 'booking-intent'. Intent drives policy — only qualification-complete or booking-intent triggers escalation.
- Slot
- Structured extracted value
Discrete qualified fields — budget, timeline, unit-type, area preference. Slots get filled progressively across the conversation and are the handoff payload to human agents.
- Fallback
- When the AI can't answer confidently
A routing decision — to a human, to a clarifying question, or to a disambiguation prompt. Fallback rates are one of the health metrics we monitor per-project weekly.
- Handoff
- AI → human transition
The moment responsibility moves from AI to a human agent. Our handoffs carry transcript, slots, intent, and any promises made — the human agent never walks into a cold thread.
- Escalation
- Priority handoff trigger
Specific patterns that bypass normal queues — explicit buyer frustration, objection patterns, booking-intent with no slot conflict, etc. Escalation rules retrain against agent outcomes.
- Session
- WhatsApp 24h interactive window
After any user message, there's a 24h window for free-form agent replies. Outside it, only pre-approved templates can send. Session management is a core operational discipline in WhatsApp.
- Latency p50/p95
- Response-time percentiles
Median (p50) and 95th-percentile (p95) response times. We target p50 < 400ms and p95 < 900ms for chat replies — fast enough to feel like a real conversation, not polling.
- Turn-taking
- Who speaks next
In voice, the model must decide when the buyer has stopped speaking and it's the AI's turn. Getting this wrong is the #1 reason AI voice feels robotic.
- Barge-in
- Interrupting the AI mid-sentence
When a buyer starts speaking while the AI is still talking, the AI should stop and listen. Humans do this constantly — voice AI that can't barge-in feels like an IVR of the 2010s.
- Language auto-detection
- Per-turn language inference
The classifier that decides which language this particular message is in — critical in SEA where buyers code-switch mid-sentence (e.g. English with Bahasa words interspersed).
Every enquiry warmed. Every viewing booked.
30-minute scoping with a senior engineer and a conversation-ops operator. You'll leave with a channel plan, language matrix, and deployment timeline — not a sales pitch.