How to use AI agents at work in one month

The best way to use AI agents at work is to start with one repeatable task, give the agent a written brief and check its results before handing over anything else. A small team does not need to automate an entire department at once. It needs a clear way to test whether an agent can do useful work without creating more work for the people supervising it.

Give the experiment a month. In week one, choose a routine job and write the instructions. In week two, review completed runs and correct recurring mistakes. In week three, introduce a second job with its own permissions and budget. In week four, compare the time saved with review time, corrections and running costs. Each stage gives the team a reason to decide what happens next.

Week one: choose a task and write the brief

Start with work that arrives regularly and follows a recognisable pattern. It should take enough staff time to justify delegation, but a mistake should be easy to catch before it affects a customer, supplier or colleague.

Consider a small company with a shared support inbox. Someone reads incoming messages, identifies the subject, checks whether the customer has provided enough information and assigns each request to the right person. An AI agent could prepare this work by sorting messages and producing a short summary for the support team.

Avoid giving the agent a vague instruction such as 'handle support requests'. It needs to know what information to use, what to produce and where its responsibility ends. Write a brief that covers:

• The source material, such as the shared inbox or a specific folder.
• The task, including how to identify subjects, extract questions and assign categories.
• The expected result, such as a summary, category, urgency assessment and suggested owner.
• The limits, including a clear instruction not to reply to customers, promise refunds or change account records.
• The exceptions, such as incomplete requests or unusual cases that need a person's judgement.
• The budget, which sets the maximum amount the delegation can spend.

Add examples of ordinary and difficult requests. A question about delivery should not receive the same treatment as a report of an incorrect charge. Explain how the agent should handle missing information instead of allowing it to fill gaps with assumptions.

For the first week, let the agent prepare the work while a team member checks every result. Do not grant permission to send replies simply because its summaries look accurate. The first assignment is a test of the brief and the agent's ability to follow it.

Week two: use the log to fix recurring mistakes

After several runs, look at the original requests alongside the agent's results. Pay attention to errors that keep appearing. One incorrect category might be an isolated misunderstanding. The same incorrect category appearing repeatedly usually points to a problem with the instructions or the information supplied.

Suppose the support agent keeps classifying cancellation requests as ordinary delivery questions. It may be treating the presence of a tracking number as the main clue. Adding a specific rule for cancellation requests, even when a tracking number is present, gives it a clearer basis for the next run.

During the review, ask whether the agent received the correct source material, which decisions needed the most correction and whether its summaries helped the person taking over. Check whether it recognised cases outside its permissions or invented details when information was missing. Record how long the reviewer spent checking each result.

The activity log helps explain what happened during a run. The deliverable shows what the agent produced. The log records tool calls, approvals and policy changes, giving the team a way to investigate where a problem started. An append-only record is useful when you need to distinguish an instruction error from a source problem or a permissions change.

Keep a short record of corrections. Change one part of the brief, run another example and see if the same mistake returns. Rewriting every instruction after each error makes it difficult to tell which change improved the work. The aim is to build a brief that handles the cases the team actually receives.

Week three: add a second job with separate boundaries

If the first delegation now produces results that people can review without constant repairs, introduce a second task. Pick something useful to the same team, but give it a separate brief, budget and expected deliverable.

An operations team might add a meeting-preparation agent. Before an internal meeting, it gathers relevant notes and documents, identifies outstanding decisions and prepares a briefing for the person leading the discussion. That person checks the material before using it.

This agent has a different job from the support triager. It needs access to different information and follows different instructions. The first agent's successful runs do not give the second agent permission to use the same tools or records.

Write the second brief from scratch. State which documents the agent can access, what it should return and which decisions remain with a person. If an old meeting note contains an unresolved decision, the agent can identify it as outstanding. It should not decide the issue itself.

Keep the two delegations separate in your records. This lets you see which agent produced each result, how much each assignment cost and where corrections were needed. It also helps identify problems if the first agent's work changes after the team introduces another one.

Set the team's approval rules before either agent receives broader permissions. Sending an external message, moving money, publishing publicly or changing records can have consequences beyond the workspace. These actions need a person's decision. The same applies to accepting terms, changing access permissions or connecting an integration with wider access.

In Delagents, the approvals inbox presents an external action as a card showing the exact action, cost so far and deadline. A person can approve, change or reject it. Every delegation also has a hard budget cap, so an agent cannot spend beyond the amount set for its assignment.

Week four: measure the time saved and the cost

An agent finishing a task quickly does not automatically mean the team has saved time. If a job previously took fifteen minutes and now takes five minutes to run but twenty minutes to review, the new process is slower.

Compare the old process with the actual work required by the new one. Include the time spent preparing instructions, checking deliverables, correcting mistakes and dealing with exceptions. Use staff working time rather than the agent's background running time.

Take a task that previously took thirty minutes. The agent prepares a result in five minutes, but the reviewer spends twelve minutes checking it and another eight correcting recurring errors. The team has saved five minutes of working time, not twenty-five. The running cost must still be deducted from any financial benefit.

A weak result does not always mean the task should be abandoned. The brief might need a clearer exception rule, or the source documents might need organising before each run. Give the team a chance to correct a specific problem, then measure the next set of results. Do not keep an inefficient delegation running simply because time has already gone into setting it up.

Assess each agent separately. A meeting-preparation agent might save useful staff time even if its briefing needs careful checking. A support triager might be a poor choice if it regularly misses urgent requests. Each job deserves its own budget and review standard.

Decide what deserves another month

At the end of week four, give each delegation one of three decisions. Keep it running if its deliverables are useful, the review burden is reasonable and the cost makes sense. Revise it if a specific problem can be fixed through better instructions, cleaner inputs or narrower permissions. Stop it if supervision and corrections take more time than the task saves.

Some work should remain with people. A task that depends on personal judgement, happens rarely or carries a high cost when it goes wrong may not be suitable for an agent. There is no benefit in keeping an assignment active simply to increase the number of automated jobs.

A shared workspace should make each assignment easy to inspect. In Delagents, a delegation record connects the brief, budget and result. The team can check what was requested, what the agent was allowed to spend and what it delivered before deciding whether the work should continue.

For the next month, choose one additional task only if the first two have earned their place. Give that task a named owner, a written brief and a measurable result. Set its budget before the first run, then put a review date on the calendar. Start with one real example of the work your team already does, and write down exactly what a person should receive when the agent finishes.