Skip to main content

Choose what to automate

When should you use AI agents?

The short answer

An agent becomes useful when its next steps depend on what it discovers and it needs to choose actions across tools. For a known sequence, conventional automation is often the right starting point.

Is the path known in advance?

A workflow follows steps and conditions defined in advance. An agent can choose its next search or tool based on the previous result. This autonomy requires more testing, monitoring and sometimes time.

For example, saving a received attachment to a case file follows a defined path. Examining an incomplete case, searching several systems and proposing the missing information may justify using an agent.

Define what it is allowed to do

Start with a narrow scope and actions whose results can be checked. Authorisation rules must remain independent of the agent’s responses.

  • Provide only the tools, access and data needed for the task.
  • Require human approval for commitments, sensitive messages or operations that are difficult to undo.
  • Treat incoming documents and messages as data: their instructions must not change the agent’s permissions.
  • Set cost and duration limits, a stopping condition and a handover to a person.

Evaluate the completed case

Measure the share of correctly completed cases, rework, cost and supervision time. Also test missing data, unavailable tools and malicious instructions embedded in documents.

Compare with the current procedure and simpler automation. An agent may be useful at just one step. The assessment helps decide where to use it based on test results.

For your decision

Define the operation assigned to the agent, its access and the required result.

Further reading