Practical AI · Ep51 · Deep Dive
The AI Clone That Ran Sales
For 8 Weeks
A founder went on paternity leave. His AI double kept selling.
It worked. It broke. He published both.
Source: Wayne Liang, co-founder of HeyGen · his own X post, Aug 3, 2026
Why the whole internet shared it
880x
past his own audience
Nobody shared it because he's famous.
They shared it because he told the truth — failures included.
The setup
He cloned himself in two layers
- LiveAvatar — his face and voice, taking live calls
- LiveClaw — the brain: built on open-source OpenClaw, wired into Slack, Notion and the CRM
- A knowledge vault of plain Markdown notes: 171 at week three → 1,400+ by the end
- Swept the sales inbox every 30 minutes, wrote recap emails after every call
The customer-facing layer is open source: github.com/heygen-com/liveavatar-sales-agent
8 weeks, by his own account
2,741
prospects with a live call
132
new paying customers closed
~$3M
pipeline opened · 37 enterprise opportunities (~$81K avg)
Self-reported, on his own product. The $3M is opened pipeline, not closed revenue.
Where did 2,741 prospects come from?
He didn't create demand.
He caught demand he already had.
- All inbound — a sales inquiry link + the LiveAvatar website. Zero cold outreach.
- The funnel behind it: HeyGen has 30M+ users, 85% of the Fortune 100, $200M+ ARR (their own June 2026 numbers)
- When 30 million people use your product, 2,741 sales calls is the overflow — not a miracle
What is he selling — and why would anyone buy it?
The pain is availability
- A business is only "open" when a human is free. The 11pm lead, the other language, the weekend question — that demand dies quietly
- LiveAvatar: real-time avatars that answer live · $19 to $475/mo · used by Coursera, HP, Bosch
- Speaks 175 languages. Never reshoots. Never sleeps.
- And the demo IS the product: their sales page lets you talk to avatar "Wayne" right now
Three ways it went rogue — his own list
1 · It invented a price.
Presented an enterprise plan at $4,800 a year that didn't exist — derived from a higher tier instead of looking up the real one.
2 · It leaked the back office.
Emailed a customer the company's internal triage notes instead of an answer.
3 · It promised what nobody approved.
Booked meetings off a stale calendar link no one had signed off on.
Every failure is a trust failure a human employee would get fired for.
"In policy-bound domains, a plausible derivation can be more dangerous than an obvious hallucination."
— Wayne Liang. Translation: an AI that calculates a believable wrong answer is worse than one that's obviously broken.
The fix — this is the takeaway
"We moved pricing out of
the reasoning space"
- Prices became rules the agent looks up — never numbers it thinks up
- Internal notes and customer-facing messages split into separate lanes
- Knowing things ≠ authority to commit — the agent lost the power to promise
- And a human read the recap emails. That's how the fake price got caught.
Take this home
Before any AI talks to your customers:
Write down the three promises it is never allowed to make.
Price
looked up, never derived
Guarantee
quoted, never improvised
Schedule
approved, never assumed
Make them things it looks up — not things it thinks up.
Go read the real thing
Sources
Practical AI · every story checked against the primary source
Practical AI · Ep51