How to use AI agents at work: start with one repeatable job

Observe how a person completes the task. Record the systems they open, decisions they make, information they trust and exceptions that require help. Measure the time spent doing and checking the work before introducing an agent.

Pick a frequent task with manageable consequences. A draft brief, document comparison or ticket triage process can be easier to evaluate than an agent allowed to change customer accounts or commit spending. The purpose is to learn where the system is useful under real conditions.

Specify the input, required output, allowed sources, prohibited actions and escalation conditions. Define who reviews the result and what evidence they should see. A task is not complete merely because the agent has produced a long response.

For an account brief, require a summary of recent activity, links to records, open issues and a separate list of unknowns. Tell the system how to handle conflicting records. Do not ask it to fill missing commercial facts with assumptions that look authoritative.

Begin with read-only access to a small set of approved systems. Check whether the agent selects the correct records and respects user permissions. Add narrow actions only when the workflow has a clear need and a reliable verification step.

A tool should have an understandable purpose and return actionable errors. If a record is missing, the agent should distinguish that from a permission failure or a service outage. Those conditions need different responses.

• Time from request to an accepted result.
• Human checking and correction effort.
• Missed facts, unsupported claims and unnecessary actions.
• Model and integration costs, including retries.
• The proportion of tasks that need escalation.

Compare these with the existing process. A fast draft can still be a poor investment if an expert must spend longer verifying it than writing the original. Ask users whether they can understand the evidence and recover from mistakes.

Keep a failure log and improve the source, instructions or tool design based on specific examples. Re-run evaluations after changes. Once one job works reliably, extend to adjacent tasks with similar controls. Avoid connecting every company system before proving that a single workflow creates value.