Autonomous agent

Autonomy is where the time saving comes from, and also where the risk comes from. A wrong turn at step two is carried into every later step, and nobody sees it until the end. A mistake with a reversible action, such as a bad draft, costs minutes. A mistake with an irreversible action, such as a sent email or a deleted record, cannot be taken back. The skill is deciding how much autonomy each task has earned.

• Ask one question first: if this goes wrong while nobody is watching, can it be undone cheaply?
• Give the agent read access and drafting rights first. Add the right to send, pay or delete only after it has proved reliable.
• Set limits on steps, time and spend, after which the agent must check in.
• Log every action, so a wrong turn can be traced and undone.
• Keep a person on anything final through human-in-the-loop approval.

A chatbot answers one request and waits. An autonomous agent takes a goal, such as "research these ten companies", and works through it. It plans a step, takes an action with a tool such as search or email, reads the result, and decides what to do next. This loop repeats until the goal is met, a limit is reached or the agent asks for help.

Autonomy is a scale and not a switch. At one end a person approves every action. At the other the agent runs unattended for hours. Most useful set-ups sit in between.

• Treating autonomy as a sign of quality. An agent can be fully autonomous and wrong.
• Giving write access to email, payments or records on day one.