“Can’t AI just do all of this now?”
I hear some version of that question a lot lately. From founders staring at a messy pipeline. From freelancer friends wondering if their work is about to disappear.
Here’s the short answer, based on the best real-world measure I know of: today’s top AI agent can fully complete about 16% of real, paid freelance projects to a client-ready standard. That’s a big jump from 2.5% a year ago. It also means most projects still need a person.
So the useful question isn’t “AI or human?” It’s “which parts go where?”
What the Remote Labor Index measures
The Remote Labor Index (RLI), built by the Center for AI Safety and Scale AI, tests AI agents on 240 real freelance projects across 23 domains, including graphic design, video, architecture, data analysis and web apps. The human freelancers who first did this work were paid a combined $143,991.
These aren’t quiz questions. Human evaluators compare each AI deliverable with the professional’s original, and it only counts if it’s at least as good as what a paying client accepted.
That’s why I pay attention to it. It measures finished work, not clever answers.
From 2.5% to 16%: what changed, and what didn’t
When the original study came out in October 2025, the best agent completed just 2.5% of projects. In July 2026, the Center for AI Safety published new results: the top model reached 15.8%, more than quadrupling the best score in under eight months.
Part of that came from better models. Part came from a better setup. The agents got a full desktop with professional apps, up to 24 hours per project, and a separate “critic” agent that reviewed each draft so the worker could revise it.
What didn’t change: roughly five out of six projects still fell short. The researchers also found that automated AI judges overrated the newest models by two to three times compared with human reviewers.
Not panic. Not hype. Steady progress, with people still checking the work.
Treat AI like a fast junior
My candid take, and I say it with love: AI is a fast, eager junior. It drafts in minutes. It doesn’t get tired. It also, sometimes, hands you something that looks finished and isn’t.
The research team saw exactly this. One model delivered a polished bathroom render that turned out to come from an image generator, while the actual 3D model underneath was malformed and untextured.
Polish isn’t proof. Before anything goes to a client or a prospect, someone with context checks it.
The draft, sort, repeat split
Here’s the simple way I decide what goes to AI in a small business.
- Draft. Let AI produce first versions: research summaries, outreach first drafts, call notes, report outlines.
- Sort. Sort each task by risk. Low-stakes and easy to check? AI can take more of it. Client-facing, high-ticket or hard to undo? A person owns it.
- Repeat. Each week, look at what AI got right and wrong, and adjust the split.
In gentle outbound, that might mean AI drafts a first-pass summary of a prospect’s company while I write the actual message. A real question to a real person is the part I won’t hand over.
Where freelancers and WFH parents fit
If you freelance from home, maybe in the small pockets between school runs and nap time, here’s the reassuring part. The skills clients pay for are shifting toward the ones AI is weakest at.
- Briefs. Turning a founder’s fuzzy idea into clear instructions.
- Quality control. Spotting what’s off before the client does.
- Judgment. Knowing when “good enough” isn’t.
That’s not “just help.” It’s the reason the work lands.
Building the system around the tool
The tool is only half of it. The system around it is what frees up hours without adding risk.
- A short checklist for every AI-assisted deliverable: facts checked, links tested, tone matched, names spelled right.
- A named reviewer for anything client-facing.
- A stop rule. If AI output needs more than two rounds of fixes, a person takes it from there.
- A weekly 15-minute review to decide what to hand over next.
Small, repeatable steps. Fewer surprises. More breathing room.
Key takeaways
- The best AI agent can complete about 16% of real freelance projects at a client-ready level, up from 2.5% in October 2025.
- Most projects still need a person, and AI judges tend to overrate AI work.
- Give AI the drafts. Keep people on judgment and anything client-facing.
- Checklists and review steps make sure the hours you save don’t come with new risk.
FAQ
Can AI replace a freelancer yet? For most projects, no. On the Remote Labor Index, about five in six real freelance projects still weren’t completed to a client-ready standard by the best AI agent.
What does a 16% automation rate mean? It means the top AI agent produced work judged at least as good as the original professional’s on roughly 16% of the 240 test projects.
You don’t have to choose between AI and people. You get to choose the split.
If you’d like help mapping which parts of your outbound and operations can go to AI, and which should stay human, book a 30-minute call and let’s build yours.