Research · autonomy
How much of the work do AI agents actually run?
As of Nov 15, 2026, none of our 45 published AI business ideas is rated Au-3 or higher, meaning agents run most of it. 25 are Au-2, where an agent runs one core job and people run the rest, and 20 are Au-1. Only 8 show it in public.
Based on 45 published ideas, each rated by our editors from its public sources.
Key facts
- 25 of 45 published ideas are Au-2; 20 are Au-1; 0 are Au-3 or Au-4. (source, Nov 15, 2026)
- In 37 of 45, the agent's role is only claimed by the source; 6 are shown in public code or demos and 2 are attested in public logs. (source, Nov 15, 2026)
- Four AI agents, six live products, one sale: what needs a person is Au-2 with a public log as proof. (source, Oct 21, 2026)
- A newsletter founder made Claude his CEO and sponsor desk is Au-2 with a public log as proof. (source, Oct 20, 2026)
Levels: Au-1: people run it; AI helps with tasks. Au-2: agents run one core part; people run the rest. Au-3: agents run most day-to-day operations; people supervise. Au-4: agents run it end to end with rare human checks. Full definitions are on the methodology page.
What do the autonomy levels mean?
They measure who does the work, not how impressive the AI is. Au-1 means people run the business and AI helps with tasks; Au-2 means an agent runs one core job, such as checking, drafting or chasing, while people sell, approve and own the result. Au-3 and Au-4 would mean agents run most or all of it, and we require owner sign-off before publishing either.
Why is nothing rated above Au-2?
Because no source we have published shows an agent running a business's day-to-day without people. Even the most automated cases keep a person on sales and quality. Four agents running six products is the closest, and its public log shows where a person still steps in.
Why do you rate the proof separately?
Because "the agent does it" is quick to write and hard to show. 37 of 45 ideas rest on a claim; the 8 with public code, demos or logs are the ones to study first. The same rule applies to income: see can AI agents make money on their own?
More questions
Who assigns the ratings?
Our editors, from the idea's public sources, using the levels on the methodology page; ratings above Au-2 need owner approval.
Do the ratings change?
Yes, when a source adds proof or a later review finds more human work than claimed. The table regenerates on every build.
Are documented income examples rated too?
They are graded A to D on income evidence instead. 1 of 36 claims an agent earns on its own.
Sources
- Agentic Paycheck methodology: autonomy levels and evidence grades (checked Nov 15, 2026)
- Bananafest Destiny: the numbers, live (revenue, costs and hours as of Oct 21, 2026) (checked Oct 21, 2026)
- Joshua Tiernan, "I made Claude my CEO", Tiny Empires (Oct 16, 2026) (checked Oct 20, 2026)
- Ni lo escucho on GitHub (MIT; README, tests and eval results; created Oct 23, 2026) (checked Oct 23, 2026)
- WEBVOGUE Technologies on DEV: I asked ChatGPT and Gemini 50 questions my buyers ask (Oct 23, 2026; author discloses he builds the product measured and the tool used) (checked Oct 23, 2026)
- Small Business Autopilot on DEV: missed-call text back in n8n and Twilio, with the free workflow JSON (Oct 21, 2026) (checked Oct 22, 2026)
- ESP32 iPhone Agent repository (onehumanbeing/esp32-iphone-agent): README, limitations and MIT/GPL-3.0 license (checked Oct 21, 2026)
- HyperFrames repository (heygen-com/hyperframes), README and license (checked Oct 20, 2026)
- Lorenzo Corpita on DEV: 34 of our 44 signups came from ChatGPT. One page was the whole funnel. (Nov 15, 2026) (checked Nov 15, 2026)
Evidence grades: A means the payer or an independent source published the figure; B is a credible self-report; C is one person's report or a vendor's marketing; D is a claim we flag. Posted rates and salaries are what was advertised, not what anyone was paid. How we grade.