Choose an agent task your team can check
Start with clear records, a result someone can review and a defined place for the agent to stop.
Choose an agent’s first task by looking at work your team repeats and can check. A useful starting point has clear source records, a defined result and someone who can decide whether that result is right.
Start with a task that prepares work for review. Keep the person’s decision visible while you learn where the agent needs help.
Describe the finished work
“Help with operations” leaves too much open. “Compare completed job records with the draft invoice and prepare missing lines for review” gives the team something it can inspect.
Write down the result before choosing the model. Name the records it needs and the person who will review its work. Decide what should happen when a record is missing or two sources disagree.
Check the records it will use
For an invoice check, the agent might need the approved scope, recorded hours, material use and the current invoice draft. Having access to those files does not mean they agree.
Try a few past jobs with the person who normally checks them. Include a straightforward job, one with approved extra work and one with incomplete notes. Ask that person to show what they would include, leave out or ask about.
Those examples give you something to compare the agent’s work against. Keep them separate from any examples used to configure it.
Decide where it must stop
Preparing an invoice line and sending an invoice are different actions. Agree which actions the agent may take and which need approval.
- What may it read?
- What may it prepare or change?
- Who approves the next step?
- What happens when the answer is uncertain?
When information is missing, asking for a check can be the right result. An agent that fills a gap with a guess creates more work for the reviewer.
Measure the whole task
Track how long the team spends reviewing and correcting the result, as well as how often it is right. Compare that with doing the work manually.
A fast draft that takes longer to check has not saved the team time. Use the trial to narrow the task, improve the records or decide that this part of the work should stay with a person.