Agentic AI · Guide
Designing enterprise agents with human approval built in
Agents that act on business systems need clear boundaries. A practical pattern for deciding what an agent may do alone, and what needs a person.
Techcerv editorial team · · 1 min read
The difference between a chatbot and an agent is action. Once an agent can update a record, send a message or start a workflow, the design question shifts from 'is the answer right?' to 'what is this agent allowed to do, and who is accountable?'
Classify every action
- Read-only: retrieving and summarising information the user can already access.
- Reversible: drafts, tasks and changes that are easy to undo.
- Consequential: payments, external communications, access changes, anything hard to reverse.
Let agents perform read-only and reversible actions within a scoped identity. Route consequential actions to a named person for approval, with the agent's reasoning and evidence attached.
Make it observable
Log each action with the agent, the triggering user, the inputs and the outcome. Review a sample regularly. When an agent's approval rate stays high and errors stay low, consider widening its autonomy — deliberately, not by default.