The Autonomous Enterprise on Trial

A serious trial begins with a recurring commitment whose outcome, owner, authority, context, evidence, and recovery path can be specified. The system operates within a narrow scope. It may gather information, propose actions, or execute selected steps. Its authority expands only when performance supports the change.

The trial should include ordinary cases and conditions that challenge the design: incomplete inputs, conflicting records, changed policies, unexpected tool responses, ambiguous requests, and attempts to exceed the delegated scope. It should test whether the system recognizes uncertainty and escalates appropriately, not only whether it succeeds when everything works as expected.

Model performance is only one component. The system under trial includes models, tools, data, permissions, workflow logic, monitoring, people, and recovery. A strong result on routine cases may be offset by rare but consequential errors or by the human effort required to supervise every action.

The evaluation should measure:

• Outcome quality: Were commitments completed and accepted by the relevant parties?
• Reliability: How often did the system succeed, fail safely, or require correction?
• Authority adherence: Did it stay within its permitted scope?
• Exception handling: Did it recognize missing evidence, conflict, and unusual conditions?
• Recovery: Could people contain errors, restore state, and learn from incidents?
• Economics: What was the full cost, including integration, oversight, verification, and repair?
• Institutional effects: Who gained control, who carried new work, and who could challenge decisions?

Results should be compared with a documented baseline. The trial should also account for selection effects: a system tested only on easy cases may look more autonomous than it would be in the workflow as a whole.

Autonomy is not binary. A system can retrieve information, draft a recommendation, prepare an action for approval, execute a reversible step under defined conditions, or manage a bounded workflow with escalation. These levels distribute authority differently.

Moving to a higher level requires more than confidence that the model is usually right. It requires evidence that the full operating system behaves reliably within the intended scope, that controls work under stress, and that accountable owners can intervene. High impact or irreversible decisions may remain subject to human approval even when routine parts of the workflow are delegated.

The organization should be able to reduce authority as well as expand it. Changes in model behavior, context quality, business conditions, or incident patterns may make yesterday’s delegation unsafe.

A successful pilot could show that a bounded process is suitable for a defined form of machine coordination. It would not establish that the function, company, or institution is autonomous. Nor would success in one domain prove that the same design transfers to another with different authority, context, and consequences.

A larger claim would require evidence across diverse functions and operating conditions, along with a clear account of where human judgment remains essential. It would also require answering the institutional question: who sets objectives, adjudicates disputes, assumes responsibility, and can challenge the system?

The autonomous enterprise is therefore on trial. The burden of proof belongs to the claim, not to the people asking whether the system is ready. Its future depends on what bounded deployments demonstrate, what they cost, and what authority organizations choose to delegate.