Every Monday a client hands you the same report to rebuild by hand. Same spreadsheet, same summary, same formatting, an hour gone before your real work starts. At some point you wonder whether an AI agent could do the rebuild for you.
It might. But the useful question is not whether an agent can do it once. It is whether you can trust it to do the task the same way next Monday, and the Monday after. This page is a two-week test plan that answers that with your own evidence, using one task and no risk to a client.
Who this is for
- You are a virtual assistant, freelancer or small-business operator with at least one task you repeat every week.
- You can already do that task by hand and can tell a good result from a bad one.
- You can spend a few minutes a day for two weeks checking output.
It is not for you if the task is one you cannot yet explain step by step, or if it involves confidential client data and you do not have the client's permission to use a new tool. An agent cannot follow steps you cannot write down.
Step 1: Pick one task and write the steps
Choose something small and low risk: a weekly summary of a spreadsheet, a first-draft reply to a question you answer often, a list pulled from a shared folder. Write the steps as a numbered list, the way you would hand them to a new colleague. Our guides on how to write an SOP and the delegation template show a format that works.
If you cannot finish the list, stop here. The task is not ready for an agent, and the written steps are already a useful result: you now have a process someone else could follow.
Step 2: Run it with approval turned on
Whatever tool you test, set it so that nothing is sent, deleted or changed until you have looked at it. Many agent tools offer this as an approval step for actions that write or send. Give the agent read-only access wherever you can, and feed it sample or public data, not live client records.
A prompt you can paste into any agent tool to start: Here are my steps for [task], numbered. Follow them exactly on the sample data below. Show me the result and list any step you could not follow. Do not send or change anything.
Step 3: Compare for ten working days
For two weeks, do the task yourself as usual and let the agent produce its own version beside it. Keep a simple log with four columns: date, minutes you spent checking, what was wrong, what was right.
At the end, ask three questions:
- Did the agent's version match yours, or did you fix something every time?
- Did checking it take less time than doing it?
- Would you be comfortable if the client saw the unedited output?
Log template to copy: Date | Minutes I spent checking | What was wrong | What was right | Would I send it as is (yes or no).
Worked example (illustrative, not our result). The task is the Monday client report. Steps written: pull last week's numbers, add them to the summary sheet, write three sentences on what changed, flag anything that moved a lot. After ten days your log might show: checking got faster each week, the agent twice misread a column header, once it invented a total, and most days you could send the summary unedited but never the numbers. Decision: keep it for the first draft, never for the final numbers. That is a specific answer you can defend to the client.
Keep the agent only if the answers are yes. If not, you have lost a few minutes a day and learned something real about the task.
Step 4: Decide, then widen slowly
If the test works, move to the next task, not to ten tasks. Add one at a time, and keep approval on for anything that sends messages or changes records. For a wider look at the options, see AI tools for virtual assistants and how to roll out AI tools without breaking your workflow. If you prefer a structured guide, the AI Tools Mastery guide is in the Academy store.
A tool you can use for the test
You need something to run the test on. One option is Creao, an AI agent platform. Its public pages say you describe what you want in plain language, agents can run on a schedule or when a Slack message arrives, they connect to your tools with OAuth or an API key, and each agent can be set to full auto or to approval gates. Its pricing page lists a free plan with no credit card, 30 one-time credits and 5 daily bonus credits. Plans change, so check the pages yourself. We have not tested Creao ourselves.
If you want to try it for this test, here is the Creao link. We may earn a commission if you sign up.
Opened this page from Instagram or another app? Creao's sign-up page says Google and Apple sign-in are not available inside an in-app browser. Tap the three dots in the corner and choose Open in browser first, or sign up with an email address.
Link blocked or not opening? Some ad blockers block partner tracking links. You do not need the link: search for Creao AI (creao.ai) in your browser and sign up from there.
Before you paste anything in
Never paste confidential client data into a tool you have not reviewed. Check the client's rules, read the tool's privacy and data terms, and use sample data for the test. If you are unsure, ask the client first. A good habit: write down what may go into the tool, and what may not, before the first run.
We may earn a commission if you sign up. It does not change what you pay.