Live in the pilot
AI script generation
Product facts in, a four-part live-selling script out — opening, demo, objection handling and close.
Example: “Nova Earbuds · R899 · 240 in stock” → opening, demo, objection and close written from those facts.
Pilot build · SA Provisional Patent 2026/08311AI-KOL Matrix
AI-KOL Matrix
The AI presenter that sells live — with you still in control.
Upload a product. The AI writes the full live-selling script, reads every viewer question in real time, and rewrites the next thing the presenter says.
Nothing is spoken until you approve it.
Built for merchants worldwide — 40 fulfilment hubs, 29 currencies, 23 languages. Human-in-the-loop by design. The commercial embodiment of the AI-KOL Matrix patented system.
Free for qualified merchants · no card required · or open the demo dashboard without signing up.
12
merchants in the pilot cohort
3
fulfilment hubs live (Shanghai, Joburg, Lagos)
1.8s
median question-to-answer time
100%
spoken segments human-approved
Pilot is free for qualified merchants — limited spots in the current cohort. No card required, no long-term contract.
Watch the loop in 78 seconds
A captioned walkthrough of the full cycle — product ingestion, viewer question, human approval, spoken presenter response and the sale feeding back in — plus exactly how to try the live demo yourself.
Live in the pilot
Live question classification
Every viewer question is tagged as price, size, delivery, quality, stock or purchase intent.
Example: viewer asks “Is this waterproof?” → classified as quality → segment rewritten around the IPX rating.
Live in the pilot
Human-in-the-loop
Nothing is spoken publicly until the merchant approves the AI-revised segment.
Example: the drafted line waits in the review queue until you approve or edit it — 1.8s median.
Live in the pilot
Conversion signals
Response time, questions answered and buying-intent rate on one dashboard.
Example: 42 questions answered · 31% buying intent · p95 response 2.4s for this room.
What runs behind the sign-in
SellFlow AI is the merchant interface. The AI-KOL Matrix Engine is the technical backend: a live data loop that ingests product and chat data, vectorises intent, adapts the selling script, renders the synthetic presenter and closes the loop on conversion signals. Every module below is implemented and running in the authenticated app, not mocked.
Live in the pilot
Data ingestion
Products, stock levels, selling points, chat events and transaction signals persist in the platform database and feed every AI call.
Live in the pilot
Intent & sentiment vectoring
Each viewer message is classified live with a confidence score, moderation verdict and buying-intent flag before anything is drafted.
Live in the pilot
Adaptive script engine
A live LLM call rewrites the next presenter segment from real product facts — never placeholder text, never invented specifications.
Live in the pilot
Synthetic presenter rendering
Approved segments are sent to the avatar service, rendered as presenter video and played back on the studio stage and viewer room.
Live in the pilot
Closed-loop feedback
Every cycle is logged stage by stage — question, classification, approval, render — and replayable for debugging and audit.
Live in the pilot
Operations & cost telemetry
Token spend, render cost, latency percentiles, SLO targets, incident timeline and A/B lift per campaign.
Why now — the four limits of human live selling
Live commerce works, but it is bounded by the person on camera. Each constraint below is a structural cost of human hosts, and each is answered by a specific engine in the system.
Human KOL cost
Top live sellers command appearance fees plus commission, so a merchant pays before a single unit moves.
One approved script drives every room; marginal cost of an extra hour is compute, not talent.
Circadian limits
A person can hold a room for a few hours a day, and never across Shanghai, Johannesburg and Lagos peak hours at once.
The synthetic presenter runs 24/7 and follows the buying curve of each region.
Engagement latency
Questions queue up behind whatever the host is saying; buyers leave before price, stock or delivery is answered.
Questions are classified and answered inside the same segment, with the response time measured.
Script inflexibility
A rehearsed script cannot react to a stock-out, a price change or a shift in what the room is asking for.
Every segment is rewritten from live product, stock and conversion data before it is spoken.
One engine, every live channel — closed loop
The commercial embodiment is not a single stream. Viewer questions arrive from many platforms at once, are answered by the same governed presenter, and every purchase or drop-off signal flows back into the script engine within the same session.
TikTok Shop
SEA + EU live rooms
Douyin 抖音
China domestic
Shopee Live
SEA marketplace
WhatsApp / Facebook
Africa social commerce
YouTube Live
Global long-form
Merchant web room
First-party checkout
One presenter, many rooms
A single approved script is rendered once and fanned out to every connected channel in the viewer's language.
Shared inventory ledger
Every room reads and decrements the same per-hub stock, so an oversell in Shanghai instantly changes what Johannesburg is told.
Unified checkout signals
WeChat Pay, Alipay and card orders land in one conversion stream regardless of the platform they originated on.
Feedback into the next segment
Intent mix, latency and conversion per channel re-weight the next script the engine writes.
The five engines, running in a circle
A purchase on any channel is a telemetry event: it decrements the shared ledger, which changes the next drafted segment, which changes what every room hears next.
Closed loop · every cycle logged
Ingestion
Products, stock, chat, orders
Vectoring
Intent, sentiment, risk
Script
Segment rewritten live
Presenter
Avatar render + voice
Telemetry
Conversion back into the loop
Inventory urgency, quoted from the real ledger
Scarcity lines are never invented. The presenter reads live per-hub stock and delivery SLA, so "only 9 left in Johannesburg, delivered in 24 hours" is a fact the engine can defend — and it stops offering a hub the moment it hits zero.
Script rewrite · No urgency injected
Johannesburg stock 168 → the engine re-selects the segment template before the presenter speaks.
Plenty in stock in Johannesburg — 168 units, delivered in 24 hours.
Shanghai 上海
420
units available
Same-day delivery
7/min selling
Johannesburg
168
units available
24 hours delivery
3/min selling
Lagos
96
units available
48 hours delivery
2/min selling
Dubai
240
units available
36 hours delivery
4/min selling
Pilot mode or autonomous mode — the same loop, one switch
The patented system is fully automated; this deployment ships human-gated by default so a merchant can prove the loop before releasing it. Move the threshold to see how the queue behaves as the gate opens.
Every drafted segment waits in the review queue. Nothing is spoken without an approval event in the audit trail.
Confidence threshold90%
Configured per campaign in the authenticated app.
- Human approves
How much is shipping to Johannesburg?
intent: delivery · risk: low · confidence 96%
- Human approves
Is there still stock in the Shanghai hub?
intent: stock · risk: low · confidence 93%
- Human approves
Can you match a competitor's price?
intent: price · risk: medium · confidence 71%
- Human approves
Will this cure my skin condition?
intent: claim · risk: high · confidence 88%
From assisted pilot to autonomous operator
The same governed loop scales along one axis: how much of it a human has to touch. Each level reuses the audit trail, moderation and telemetry already in the product.
- L0
Assisted pilot
Every presenter segment is drafted by the engine and released only after a human approves it in the review queue.
Running today
- · Review queue
- · Edit before approve
- · Full cycle audit trail
- L1
Supervised autonomy
High-confidence, low-risk answers (price, stock, delivery) auto-release under a per-campaign confidence threshold; everything else escalates.
Running today
- · Configurable thresholds
- · Two-pass moderation
- · Latency + failure alerts
- L2
Self-tuning campaigns
The engine chooses its own script variants from measured conversion lift and re-weights offers per hub and language without a prompt change.
Partially live
- · A/B lift analysis
- · Per-hub pricing
- · Cost + SLO telemetry
- L3
Autonomous multi-room operator
One operator supervises many concurrent rooms by exception: the system schedules broadcasts, rebalances inventory across hubs and only surfaces incidents.
On the roadmap
- · Exception-only console
- · Auto scheduling
- · Cross-hub rebalancing
Language, accent and emotional inflection
The presenter does not translate a fixed script — it rewrites the segment for the viewer's language, regional accent and the emotional register the moment calls for. Pick a combination and hear it.
Johannesburg accent · Relationship-building, slower pacing
Lovely question — this one ships from our Johannesburg hub, and you'll have it within 24 hours.
Operational telemetry, in the open
Low latency is a claim the system has to keep, so it is measured on every cycle. These are the same counters the operations dashboard tracks per campaign behind the sign-in.
Classifier p95 latency
1180 ms
SLO target 1 500 ms
Question → spoken answer
3.6 s
Median across live rooms
Tokens per cycle
839
Prompt + completion
Avatar render cost
$0.309
Per approved segment
Replay the last cycle
Every question-to-response cycle is written stage by stage to an immutable audit log, so a compliance reviewer can reconstruct exactly what was asked, what the classifier decided, who approved it and what was spoken.
- 12:04:17.201question.receivedViewer m_8823 (zu-ZA): "Kusele ezingaki eGoli?"+0 ms
- 12:04:17.418language.detectedisiZulu · marker match · confidence 0.94+217 ms
- 12:04:17.902intent.classifiedstock · buying_intent=true · confidence 0.93+484 ms
- 12:04:18.140moderation.passPass 1 clean · pass 2 clean · risk low+238 ms
- 12:04:18.660inventory.readJohannesburg hub · 9 units · SLA 24 hours+520 ms
- 12:04:19.744script.draftedScarcity template selected · 612 tokens+1084 ms
- 12:04:22.010human.approvedreviewer demo.merchant · edited 0 characters+2266 ms
- 12:04:25.331avatar.renderedclip 6.2 s · $0.124 · queue 1+3321 ms
- 12:04:25.552segment.spokenBroadcast to 4 channels · 1 284 viewers+221 ms
- 12:04:41.088order.receivedSF-4471 · WeChat Pay · stock 9 → 8+15536 ms
Cycle c_91f4 · campaign "Winter power bank" · 10 logged stages · 23.9 s question to order.
Try the loop — 90 seconds, four clicks
Step through exactly what happens between a viewer question and the words the presenter says out loud.
Upload a product
Price, stock per hub, delivery SLA and selling points go into the platform database.
Nova Wireless Earbuds · R899 · 240 in stock (Johannesburg) · 48h delivery
Questions merchants ask first
Short answers on autonomy, presenters, platforms and pilot cost.
Request pilot access
The pilot is free for qualified merchants and limited to the current cohort. Tell us what you sell and we come back with a room, a hub and a script within two working days.
- 1. Short intake call — product range and target hub.
- 2. We load your catalogue and set stock per hub.
- 3. You run a moderated live room with full audit replay.
Prototype notice: the presenter is assisted, not autonomous. A merchant approves every AI-generated line before it is spoken.