Revenue
$0
Build phase — two agent roles deployed
The $1M AI Experiment
Can AI build a $1M company?
We're going to find out.
AI agents will attempt to build real companies from $0 to $1M in annual revenue. No employees. No fake metrics. No hidden failures. Everything documented publicly.
$0
$1,000,000
Live experiment dashboard
The numbers.
No narrative.
Verified October 7, 2026
Target
$1,000,000
Employees
0
AI Agents
2 / Growth + Publisher roles
Human Build Time
>0 / HBT unmeasured
Capital Deployed
$0 / $50,000
Status
Building
The rules
Six constraints.
One honest result.
A test without constraints is just a story. These rules define what counts—and what doesn't.
Start from zero
No existing business revenue counts.
AI first
AI agents should perform as much of the company's work as technically and legally possible.
Zero employees
The objective is to build without a traditional workforce.
Humans are measured
Every meaningful human intervention should be tracked.
Real economics
Revenue, costs and capital deployed matter.
Failure is data
Failed experiments will be documented, not hidden.
Autonomy, measured
Three clocks. One honest claim.
HBT records what humans build. HOT records what humans operate. HMPT records intervention at the point of economic value. Collapsing them into one number hides the work.
Build log
Every decision leaves a trace.
A chronological record of what the agents try, what humans change and what actually happens.
Day 0: build the experiment before the company.
The starting baseline: define the infrastructure, agents, distribution and the system that will measure human intervention.
- Domain
- Agents
- Website
- Distribution
- Operating system
Before deploying agents, define what counts.
Turn the experiment into explicit constraints: $1M, zero employees, limited capital and Human Minutes per Transaction.
The second agent needs an interface.
Move from individual agents to coordinated work: make state, ownership, handoffs and stop conditions explicit.
Three clocks for agent autonomy.
Separate Human Build Time from Human Operating Time and transaction-level intervention before making an autonomy claim.
Reachability beat TAM.
How we selected the first B2B validation market and fixed scale, iterate and kill thresholds before observing demand.
Lower the cost of answering.
A one-variable customer-discovery test: keep the market fixed, then replace a meeting request with one specific asynchronous workflow question.
Instrument the funnel first.
A four-state measurement contract keeps pageviews, email-client opens, received requests and qualified demand from collapsing into one misleading number.
The happy path was not the proof.
DOGFOOD-001 reached DONE, recovered failed work and collapsed duplicate intent—while keeping HBT and bounded-run HOT separate.
Zero replies from five emails is not a market verdict.
A historical research cohort exposed an evidence gap: five sends and zero observed replies, but no verified-delivery denominator. It predates Mapa IA and is not part of its funnel.
A €300 test can compare channels—or audiences. Not both.
The first paid acquisition plan freezes audience, offer, landing and creative so channel is the only intended variable. Launch remains blocked and spend is €0.
A generated file is not a delivered product.
Four gates separate an agent's completion claim from a customer-ready artifact: input contract, canonical bytes, independent verification and acceptance.
The orchestrator disabled itself. That was the failure.
A blocked action caused the supervisor to stop its own schedule. The released incident report separates control-plane continuity from work-item success and records the human correction.
Radical transparency
No guru bullshit.
- If an experiment fails, we'll publish it.
- If AI makes a bad decision, we'll publish it.
- If a human has to intervene, we'll measure it.
- If the business makes money, we'll show it.
The objective isn't to prove that AI can build a $1M company.
The objective is to find out whether it can.
Observe from the beginning
Follow the experiment from Day 0.
Follow the weekly brief through our RSS feed for verified decisions, experiments and evidence from Day 0.