AI agents tried real freelance jobs. 4 in 5 still fail.

AI agents were given real, paid freelance projects and judged by people.

The numbers

1. October 2025, at launch2.5%

   of real freelance projects done to an acceptable standard by the best agent.

   Judged by humans against the work a professional delivered.

   Source: Remote Labor Index, arXiv 2510.26787

2. Remote Labor Index, 230 projects, 4 Nov 2026

   A year later: 8x better

   • Claude Opus 5.521.25%
   • GPT 6 Astra20.83%
   • Fable 5.117.92%
   • Best agent, Oct 20252.5%

   Drawn on a 0 to 100% scale. The rest of the bar is the work that failed.

   Source: Scale Labs RLI leaderboard, read 4 Nov 2026

3. Graded by AI vs graded by people2.3-2.9x

   how much an AI judge overstated the agents’ results.

   So only human-graded scores count. Be careful with ‘AI judged it’ benchmarks.

   Source: CAIS, 1 Jul 2026

4. OSWorld, everyday computer tasks

   Close to people on small tasks

   AI agents, best~66%up from 12%

   Human baseline72.35%task success

   Source: Stanford AI Index 2026

5. Why the gap

   Small tasks are close. Whole projects aren’t.

   A whole project is many steps, a client brief, and a finished file someone accepts. Agents still break somewhere along that chain.

   The trend is clear, though: from 2.5% to over 20% in about a year.

Use the agent as your junior

1. Let it do the steps, you own the delivery. The 80% that fails is the part clients pay a person for.
2. Sell the accepted result. Clients buy finished work, not attempts.
3. Watch this leaderboard. When the number moves, your offer should too.

Sources

1. Remote Labor Index, arXiv 2510.26787 (Scale AI and CAIS), 30 Oct 2025
2. Scale Labs, Remote Labor Index leaderboard, read 4 Nov 2026
3. CAIS, A significant increase in digital labor automation, 1 Jul 2026
4. Stanford HAI, The 2026 AI Index Report, Apr 2026

Would you hand an agent a client project? Tell me on LinkedIn or Instagram.