How to choose your first task to automate

Start with work your team is tired of repeating. Pick a task with clear steps and a result someone can check. Then test whether the new process saves time after review and fixes.

You do not need to hand over a whole department. A useful first test might cover one step in a job your team already knows well.

Find the repeat work

Ask the people who do the work where they keep copying, checking, or typing the same things. Use real examples from a normal week, not just the worst day last month.

For each task, write down:

• How often the team does it.
• How much time it takes from start to finish.
• Where the needed facts come from.
• What a correct result looks like.
• What happens when someone makes a mistake.

A task that feels dull is worth a look. That alone does not make it a good choice. A rare task with many special cases may cost more to set up and check than it saves.

Choose one step with clear rules

“Handle customer service” is too broad for a first test. “Draft replies to questions about store hours using our current hours page” gives you a smaller job to test.

Write the steps as if you were teaching a new team member. Name the starting point, the facts to use, and where the work should stop.

For the draft reply example, the steps could be:

1. Read the customer's question.
2. Check that it asks only about store hours.
3. Find the hours on the approved page.
4. Draft a reply with the right day and location.
5. Leave the draft for a person to check and send.

If the message also asks for a refund, send it to a person. Do not let the system guess a new rule.

These are examples for planning a test, not reports of client results.

Make the result easy to check

Before you build, choose a few past cases where your team knows the right answer. Include a normal case, missing facts, and an unusual request. Use copies you have permission to test with, and remove private details when they are not needed.

Set a clear check for each result. For a draft reply, check the day, hours, location, and wording against the approved source. For a copied record, compare each required field with the original.

Keep the source beside the result so the reviewer can spot errors. If review takes as much work as doing the task by hand, try a smaller step or a different task.

Decide whether you need AI

Fixed automation follows rules you set in advance. It can move a form entry into a work list or send an alert when a field meets a rule. Microsoft documents this pattern as triggers, actions, and conditions in Power Automate.

AI may help when the task involves varied text, such as drafting a reply. A single AI step can sit inside a fixed process. That does not mean you need an agent.

An AI agent can choose its next steps and tools as it works toward a goal. This is the distinction used in Anthropic's guide to building effective agents. The guide recommends starting with the simplest solution that works. Its terms are useful here, but not every provider uses them in the same way.

For your first task, use fixed rules where they fit. Add AI only where you need it. An AI answer is also not the same as an action: drafting a message does not mean the system should send it.

Count the whole cost of the work

Measure the current task before the test. Then measure the same kind of work during the test.

Count all staff time, including:

• Getting the facts ready.
• Starting or watching the process.
• Reviewing its output.
• Fixing mistakes and handling odd cases.
• Keeping the rules and source facts current.

Track setup time and tool fees separately. Record delays too, since less staff work does not always mean a faster result for the customer.

Compare the total staff time before and after for the same amount of work. Check quality as well as speed. Do not count time as saved if the team just moved it into review or cleanup. There is no promised saving until you have measured your own task.

Set approval and stop rules

Start with drafts or test copies when you can. Give the system access only to what the task needs.

Name the person who owns the test and can stop it. Decide what needs approval before it affects a customer, spends money, deletes data, or changes access. Build those limits into the tools and permissions, rather than relying on a written request alone.

Pause the process when required facts are missing or checks fail. Keep a log of what happened and a way to return to the manual process.

Anthropic recommends tests in a separate test environment, controls, and stopping conditions for agents. NIST's AI Risk Management Framework offers broader, voluntary guidance for managing AI risks. Neither source makes a test risk-free.

Your first-task checklist

Use this list with the person who does the work:

• Your team repeats this task often enough to justify a test.
• You can explain the steps and the cases that need a person.
• You have the right to use the data and connect the tools.
• You know what a correct result looks like.
• You can check the result before it causes harm.
• You will count review, fixes, setup, and ongoing upkeep.
• Someone owns approval, errors, and the stop decision.

If you cannot explain the steps yet, map them with your team first. If you cannot check the result safely, choose a lower-risk task.

Pick one task, write down its rules, and run a small test. Expand only when the results show that the process helps your team without weakening its checks.