Autonomous Business
Autonomy begins with accountability
Before a company gives an AI agent access to its operations, it should be able to answer a surprisingly ordinary management question: what, exactly, has been delegated?
By FABCODE2 min read

Give the task a boundary
An agent that can prepare an invoice has a different mandate from one that can send it. An agent that can identify a late payment has a different mandate from one that can renegotiate a contract. Treating these as variations of the same automation hides the decision that matters: which consequences the business is prepared to accept without another person intervening.
Anthropic's engineering guidance distinguishes predefined workflows from agents that choose their own sequence of actions. It also recommends starting with the simplest approach that meets the task. That distinction remains useful even as the underlying tools change. A fixed approval process does not become more valuable merely because an agent could improvise around it.
Consider an illustrative payment-reminder process. The first useful deployment might retrieve the invoice, check the agreed payment terms and draft a message. Sending the message could remain a separate permission. Changing the amount or the terms would fall outside the mandate entirely. That is a small system, but its responsibility is understandable.
Design the exception before the happy path
The interesting cases are often the awkward ones: a disputed invoice, a customer with a side agreement, an incomplete record. A production system needs a response to uncertainty that is more precise than asking the model to be careful. It needs an owner to contact, the evidence to include and a rule about what happens while a decision is pending.
This changes how progress should be measured. Completion rate alone rewards systems that keep moving. It says little about whether they should have moved. A team should also inspect incorrect actions, unnecessary escalations, recovery time and the cost of review. The right balance will differ between a draft newsletter and a change to a customer's billing record.
Make the mandate expandable
A useful first mandate should leave a record: what the agent saw, what it did, which checks passed and where a person intervened. That record is the basis for expanding authority. Without it, a wider deployment is largely a judgement about confidence rather than evidence.
For founders, this suggests a practical order of work. Choose one recurring process. Define the permitted action, the forbidden action and the escalation route. Run it against ordinary cases and deliberately difficult ones. Increase the mandate only when the remaining errors are understood and manageable.
This approach has a cost. Review, permissions and recovery mechanisms require engineering effort, and some processes will never justify it. The value of autonomy is the amount of useful responsibility a business can delegate while still being able to explain its decisions. That is a more demanding ambition than simply increasing the number of agents at work.


