Three things an agent does
An agent in a business process essentially does three things. It assesses a signal: an incoming email, a meeting transcript, a document in a folder. It retrieves context from systems you have allowed it to read: the CRM, knowledge base, project board. And it proposes an action: a draft reply, a task with an owner and date, a CRM update with a visible difference.
What it does not do, unless you explicitly configure it: write without anyone watching.
Proposing or executing: the boundary
The question is not whether AI can do something, but what happens when it goes wrong. That is why we draw the boundary according to the consequences of an error.
- Read-only: the agent reads and reports but changes nothing. Suitable for knowledge bases, reporting and progress monitoring.
- Proposals with approval: the agent prepares a change and a person approves it. Suitable for CRM updates, client replies and tasks affecting others.
- Automatic with a log: the agent executes without intervention, and every step is logged with a way back. Only where the consequences of an error are small and reversible.
A process almost always starts in the first or second category. Only when a second implementation shows that proposals are consistently good does a step move to the third.
What you leave for a person to decide
Some decisions you do not leave to a model, however well it performs:
- Anything creating an obligation: sending a proposal, accepting a contract, releasing a payment.
- Anything with a human element the model cannot see: a client who has already called three times, a colleague under pressure.
- Anything about which the agent itself is uncertain. A good workflow has an explicit way out: I do not know, back to a person.
How to set this up concretely
Start with one process and describe, for each step, who currently decides and what happens when it goes wrong. Choose the control level for each step. Make approval visible in the system where work happens, not a separate dashboard nobody opens. And log everything, including rejections: that is the source you can later use to change the control level based on evidence.