Is the Future of Work Autonomous?

This article brings together a three-part series from our Human Team newsletter where we ran an experiment to understand whether AI could run an entire company autonomously...and the implications for the human workforce. Links to all three parts, including the demos, are at the bottom.

Could AI run an entire company?

In March 2025, I started a three-month experiment with my friend and former colleague, Kishore, to explore that question. We built a digital agency run by AI to understand what autonomous work could look like and which human skills would be harder to automate.

The experiment

We created eight AI agents covering functions such as research, product, and marketing. An AI General Manager handled client communication, broke work into tasks, and assigned them to specialist agents.

Clients could explain an idea by phone or through a website. The team would then prepare a project brief, research the market, develop a digital strategy, and build and deploy a landing page. We wanted the agency to keep identifying and completing work in the background with minimal human involvement.

We focused on the agency’s core services. Finance, accounting, and HR were outside the scope, so the experiment did not test whether every part of a company could run autonomously.

What the experiment showed

Within a few weeks, we had a basic version that could take an idea through to a launch with very little intervention. Agents could coordinate a sequence of work, write code, deploy websites, and produce marketing content.

But the system was brittle. A workflow could succeed once and fail the next time. Errors compounded across steps, and agents sometimes forgot client context or made poor recommendations. Getting a prototype working was much easier than making it dependable.

The team also struggled to decide what mattered next. When we asked it to develop its own business idea, it kept producing more research, requirements, and copy without making meaningful progress toward finding customers. More activity did not necessarily mean more progress.

Two problems kept coming up: reliability and context. The agents needed better ways to retain information, recognize gaps, and ask for clarification. They also needed quality checks and human review. The first version failed in real-world conditions.

By our December 2025 follow-up, we were working on a second version. Some tasks needed less supervision and recommendations were becoming more grounded, but reliable autonomous operation remained unproven.

From player to coach

The biggest lesson was the change in our own role. As AI took on more execution, we spent more time setting direction, providing context, guiding the work, and reviewing the result.

Those responsibilities require judgment. You need enough knowledge of the work to distinguish an answer that is right from one that only sounds right.

That creates a challenge for people entering the workforce. Junior workers traditionally build judgment by doing the work and learning through feedback. If AI takes on more of those tasks, we need other ways for people to develop that understanding. AI could help teach those fundamentals while supporting the work.

What this means for human work

More autonomous systems could give individuals and small teams much more capacity. Deciding what deserves that capacity becomes more consequential: which problems to pursue, what outcomes matter, and where human attention is needed.

The effect on employment remains uncertain. Lower costs could create more demand and new work, or companies could need fewer people to deliver the same output. More people may also build businesses or work independently as AI expands what they can accomplish.

Our experiment could not settle those questions. It made the support people will need much clearer: durable skills, resources for solo and independent workers, and human connection as work changes.

The full three-part experiment, including the demos, what broke, and our conclusions, is linked below.