A single harness (Orchestrator · Claude) coordinates every tool, memory, guardrail and flow — build once, deploy across every channel.
Harness architecture
Tools + MCP
Resources / Connectors
🧰
⇅
Session
Deal memory
🗂️
⇄
Harness
Orchestrator · Claude
✳️
⇄
Sandbox · Guardrail
Review
🛡️
⇅
Orchestration
Multi-step flow
🔀
Customer-behavior library — accumulated data
The agent does not treat every customer the same. From 1M+ conversations it has learned many seller-behavior segments and a tailored play for each — applied tens of thousands of times.
🐢
Passive seller
2.310 applications
📱
Casual, unfamiliar with the process
1.665 applications
🔗
Has an external price reference
1.392 applications
📄
Legal issue / pledged car
1.390 applications
🏃
Urgent seller
1.229 applications
📋
Seller with a plan
1.096 applications
= Mr. Hiếu (demo deal)
🤔
Undecided, still weighing it
1.073 applications
🔍
Verifies carefully before closing
1.068 applications
💰
Expects above-market price
learning
😤
Firm price, hard negotiator
learning
🚗
Selling multiple cars at once
learning
🔄
Trade-in (upgrading car)
learning
👻
Re-engage after long ghosting
watching
🏔️
Remote province, outside inspection zone
watching
🛠️
Reveals car faults while negotiating
watching
💭
Doesn't know their car's value
watching
16 behavior segments, each with its own play learned from real deals — segments with a number are applied steadily; "learning/watching" segments are new patterns still being validated. Source: agent_insights.customer_type (db=9) · 3.93M log lines · 73,968 runs.
Results
AI Operations Leverage — treat the agent as a production line
Measured with operational metrics: Throughput · Cycle time · SOP. AI gives the biggest leverage in lead nurturing, photo collection and bidding; the first price touch and post-inspection close still need humans.
SOP conversion funnel — Milestone Transitions (AI vs human)
Lead →60%→76%→92%→20%→47%→13%→Win
T1
Lead → Quote
60.4%
3258/5398 · −15.6pp
Human: 76%
T2
Quote → 4+ photos
76.2%
2483/3258 · +32.1pp
Human: 44.1%
T3
Photos → Bid
91.8%
2280/2483 · +10.7pp
Human: 81.1%
T4
Bid → Gap ≤20M
20.4%
523/2561 · +0.4pp
Human: 20%
T5
Gap ≤20M → Inspection
46.8%
245/523 · +4.5pp
Human: 42.3%
T7
Inspection → Win
13.3%
125/937 · −21.3pp
Human: 34.6%
Reading the funnel: AI clearly beats humans at T2 (+32pp), T3 (+11pp), T5 (+4.5pp) — lead nurturing, photo collection, bidding ⇒ huge throughput leverage. It still trails at T1 (first price touch −15.6pp) & T7 (post-inspection close −21.3pp) ⇒ exactly the two bottlenecks at each end of the funnel, where human effort is concentrated.
Operational leverage — real numbers (06/2026)
Throughput
Volume one AI carries
1.07M
messages handled · 74K runs
94.8%
after-hours messages (10pm–7am) by AI
~4×
volume vs humans (20.2K vs 4.9K msgs/month)
Cycle time
Deal lifecycle speed
~3 min
response (median), 24/7 coverage
6 days
median Lead→Win · 45% close ≤5 days
3 days
Santafe demo deal — 2× faster than median
Source: Milestone Transitions from Vucar's internal dashboard (7 Transitions); ops metrics computed from CRM (db=2) + Zalo (db=11), 06/2026.
Customer trust
How do customers react to the agent?
These are customers' own words on Zalo. We include the praise, the criticism, and the customers who sensed "something was a machine". Their messages are kept in the original Vietnamese.
The customer in the demo deal — Mr. Hiếu (Santafe)
💬 Mr. Hiếu's messages, verbatim — the whole conversation handled by AI
“A có bớt 1 chút để ra lộc cho khách mà e”
“Xe a đi ít lắm, thay dàn lốp Michelin rồi, đèn pha thay sáng hơn… A bớt thêm 5tr cho khách”
“Ok e” → closed at 889M VND (13/06)
Mr. Hiếu speaks casually as if messaging a real salesperson, volunteers a price and closes in 3 days — with no idea he was talking to an AI. (This deal wasn't separately star-surveyed; the trust signal is his own reaction in the chat.)
What customers say about the service
“cảm ơn em, anh sẽ nhắn cho em khi thấy muốn bán xe, anh thấy dịch vụ này rất tuyệt”
— Trần Văn Cảnh · 30/06
“cảm ơn em đã nhiệt tình hỗ trợ anh ạ”
— Vũ Đình Lưu · 30/06
“Cảm ơn a hỗ trợ nhiệt tình quá”
— Lý Thiên Sinh · 30/06
“Cảm ơn em nhiều, chúc công việc của em luôn thuận lợi, thành công!”
— A Đức · 30/06
“Cảm ơn em. Xe anh khách mà thấy trực tiếp là ưng ý ngay. Tks em”
— Nguyễn Văn Phúc · 30/06
“Tôi thấy Vucar rất tốt. Nhân viên nhiệt tình. Coi xe kỹ và cẩn thận”
— in-app survey (sell_experience) · 5★
Straight talk — what do customers criticise, and do they notice the AI?
Pushback is mostly about PRICE, not the service
Hundreds of messages like "giá thấp quá / rẻ quá anh không bán" (price too low, won't sell) — that's the dealer bid not being high enough, the nature of market negotiation, not an agent fault. It's exactly why the agent must price sharply to keep sellers at the table.
Real negative experiences (rare but real)
"Do hôm trước đợi lâu quá nên chị cho thuê rồi" (waited too long, so I leased it out) — slow response lost the deal. "Lặp như robot" (repeats like a robot) — the agent sometimes replies mechanically. "e để cho a hy vọng, cuối cùng thất vọng tràn trề" (raised my hopes, then let me down) — mismatched price expectations. These are being tightened with SOPs + guardrails.
Do customers realise they're messaging an AI?
Almost never. Most address it casually as a real person. In a whole month only ~9 customers asked "bot à?" (is this a bot?) — e.g. "tổng đài bên em là bot tự động hả em, sao gọi nói gì lạ vậy @@" (is your hotline an auto-bot? the call said odd things). The more common trust question is about identity / fees rather than human-or-machine: "vì sao đảm bảo e có phải nhân viên chính thức của cty Vucar k" (how do I know you're a real Vucar employee?).
Source: inbound customer messages on Zalo (db=11), filtered to real customer threads (excluding internal Vucar), 01–30/06/2026 + in-app user_feedback survey (db=2). Quoted verbatim, only sensitive info redacted.
See it for yourself
Leave a lead and watch the agent work — live
Enter a phone number that has a real lead and watch the actual conversation and the agent's runtime thinking, side by side.
Real deal
A real transaction — end to end
Hyundai Santafe 2021 · Hanoi · closed at 889M VND in 3 days. Click each step to see the agent work — the right column is the harness's real trace.
Anh Hiếu 📋 Seller with a planHyundai Santafe 2.5L HTRAC 2021 · Hanoi · 035****888
Vucar's proposed price
889.000.000 ₫
Anh HiếuCar owner · Santafe 2021
ZALO
Type a message…
➤
Behind the scenes — how the agent produces its output
Harness running
The Orchestrator (Claude) moves through each component — real trace of this lead
🧰
Tools + MCP
⇅
🗂️
Memory & Skills
⇄
✳️
Harness
⇄
🛡️
Guardrail
⇅
🔀
Orchestration
</> View raw log — agent_pipeline_logs (db=9)
Real trace from agent_pipeline_logs (db=9) for deal car_id 1dd7716f… — Hyundai Santafe 2.5L HTRAC 2021, Hanoi, 10–13/06/2026.
Conversation from Zalo (kept in Vietnamese). Tool names & params unchanged. · Move steps: ←→
