Is the autonomous company the wrong question?

Autonomy hides several questions

A system can draft, calculate, route, negotiate, or execute a workflow with little intervention. That is evidence about particular tasks. It does not yet tell us whether the organization has found a worthwhile activity, understands the people it serves, or can sustain the work economically.

The phrase “autonomous company” bundles together distinct questions: which tasks can be delegated, how decisions are coordinated, who chooses the goals, and whether the resulting activity is valuable. Treating one as proof of the others makes a technical demonstration look like an economic result.

Research on decentralized autonomous organizations offers a useful warning about the word itself. A 2026 study reviewing 80 articles and analyzing a large DAO dataset separates automation, where code executes decisions, from independence, freedom from outside influence; it finds that most DAOs achieve neither. DAOs are not AI-run conventional firms, but the distinction shows why autonomy needs to be defined as specific properties rather than treated as one achievement.

Field evidence already shows why the distinction matters. In a study of 5,179 customer-support agents, access to a generative AI assistant increased issues resolved per hour by 14% on average, with larger gains for less experienced workers and little measured effect for the most experienced. That is a consequential result for one workflow. It is not evidence that an entire firm can be run autonomously.

Generating a strategy is not validating it

AI can also widen the search over possible strategies. In an experiment using start-up accelerator applications, investors rated AI-generated plans at least as favorably as plans written by entrepreneurs. A separate comparison found that AI evaluations of 138 start-up competition plans were positively correlated with investor evaluations. These results make AI-assisted strategy work worth examining. They measure judgments about plans, however, not customer adoption, operating performance, or returns from launching the ventures.

That gap is visible in a field experiment with 640 entrepreneurs in Kenya. Giving entrepreneurs access to a GPT-4 business assistant did not produce a detectable average increase in revenue or profit over the study period. Effects varied with entrepreneurs’ starting performance. Advice can be available, fluent, and sometimes useful without reliably improving a business outcome.

Even on knowledge work, performance depends on the task. In a randomized experiment with 758 consultants, GPT-4 improved performance on tasks within its capability frontier, while performance fell on tasks beyond it. People did not always recognize where that frontier lay. The relevant unit was the task and the surrounding judgment, not the nominal autonomy of the organization.

Evidence at the firm level points to organizational change, too. A study of U.S. firms found that AI investment was associated with more educated and specialized workforces and flatter hierarchies. The analysis describes changes that occur alongside adoption; it does not establish that AI caused them or that these firms became autonomous.

The economic question comes first

An activity becomes worth building when it addresses a real need under conditions that can support its delivery. That means learning how people deal with the problem now, what it costs them, which alternatives they trust, and what would make them change. It means testing whether a proposed service can be delivered reliably, at a cost and price that make sense, with the people affected willing to use it.

AI may change those conditions. It may lower the cost of research, reasoning, or coordination enough to reopen an activity that used to be impractical. But the technology does not tell us which unmet need matters, whether the old constraint was the decisive one, or whether the new arrangement works outside a demonstration.

Make autonomy a question inside the inquiry

A stronger inquiry starts with the activity: what valuable work is not being done, what has kept it from being done, and what evidence would show that those conditions have changed? Only then can we ask which parts should be automated, where human judgment remains necessary, and what the whole system costs when it is used in the field.

This makes autonomy a design choice to test rather than a destination to announce. Some activities may benefit from extensive automation. Others may depend on relationships, unusual cases, physical work, or decisions whose consequences demand human authority. The answer should come from the work and its evidence, not from a preference for either people or machines.

An autonomous company can be a revealing experiment in coordination and execution. It is not, by itself, proof of a valuable company. The more useful question is: what becomes worth doing as the cost of discovery and coordination falls, and what would convince us that it works?