Before You Automate It, Run the Workflow by Hand

Last updated: 2026-08-15

A task that irritates you once may be random. A task that irritates you three times is usually asking for a system.

That does not mean you should immediately build an automation.

It means you should get curious about the task.

Where does it start? What information does it need? What would a useful result actually look like? Where do you make a judgment call that a new teammate would miss?

Those questions matter even more with AI. AI can produce a lot of work very quickly. If the brief is fuzzy, it can produce a lot of work that creates a new mess for you to clean up.

Start with the annoying repeat

Think about a task that has shown up a few times lately. Maybe it is turning a client call into a follow-up email. Maybe it is sorting meeting notes. Maybe it is preparing the same report every Friday.

The task does not need to be huge. In fact, smaller tasks are often better places to start.

A founder we coach had a workflow in mind for turning meeting transcripts into organized notes. The temptation was to build the whole thing immediately. Instead, I suggested a slower first pass: run the transcript through an AI chat, inspect the result, adjust the instructions, then repeat it with another meeting.

That manual run reveals the stuff you cannot see from a blank automation canvas. One meeting may have a promised introduction. Another may have a decision with no owner. A third may contain a detail that belongs in a customer record rather than the project notes.

If you automate before seeing those differences, you end up designing for an imaginary version of the work.

Write down what “done” means

The first part of AI delegation is often the most valuable part. You are defining the finish line.

For a post-meeting follow-up, “done” could mean:

• A draft email names the decisions from the call.
• It lists every promised action with an owner.
• It includes dates only when the conversation actually included dates.
• It leaves uncertain details clearly marked for review.

Now the AI has something useful to work toward. You have also made your own expectations visible.

This is the difference between asking, “Summarize this meeting,” and asking for an email draft that helps the conversation move forward. The first request may give you tidy notes. The second one can save you from reopening the transcript two days later, trying to remember what you said you would send.

When I delegate work to AI, I think about it as 80/10/10. The opening 10 percent is the brief: context, resources, constraints, and what a good result looks like. The middle is execution. The last 10 percent is review.

That final pass is where your voice, relationships, and judgment stay in the picture.

Let the AI ask for missing context

Many people get stuck trying to write the perfect prompt. I would flip that around.

Tell the AI what you want to accomplish, then ask: “What do you need from me before you begin?”

For a recurring meeting workflow, it may ask who should receive the follow-up, how direct the tone should be, where tasks belong, or whether introductions need approval before they are drafted. Those are good questions. Answer them once. Save the answers somewhere you can reuse.

You are building a small operating manual for the task.

And you do not need to get every detail right on day one. Run the workflow. Look at what it produced. Add one instruction after a miss. Remove one instruction that did not matter. The workflow gets better because it is based on actual work, not a theory of the work.

Document the judgment before the automation

Some recurring tasks have a hidden decision inside them.

Take a client update. A good update does more than repeat what happened. It decides what belongs in the message, what can wait, what needs a clear owner, and what would create unnecessary noise. If you want AI help with that, document those choices first.

You might write a few simple rules:

1. Include decisions and commitments from the conversation.
2. Do not invent deadlines or certainty.
3. Keep the email short enough that someone will read it.
4. Flag anything sensitive for me rather than sending it onward.

Those rules give the system boundaries. They also make review faster because you know what you are checking for.

This is a useful test for any proposed automation. Can you explain how a skilled person decides whether the output is good? If the answer is no, stay manual for another few rounds. The task may still be worth doing. It just is not ready to run unattended.

Keep the human pass

Automation is great at getting a first draft onto the screen. It is less reliable at knowing when a sentence sounds off, when a detail will land badly, or when the right move is to say less.

That is why I like draft-first workflows. Let AI prepare the follow-up email, pull action items from a transcript, or organize a rough report. Then read it before it leaves your hands.

The review is not busywork. It protects trust. A generic email with your name on it can cost more than the five minutes you saved.

So this week, skip the giant automation project. Pick one task that has annoyed you three times. Run it manually with AI beside you. Write down what good looks like. See where your judgment enters the process.

You may end up with an automation. You may discover the task only needed a template. Either way, you will have a system that matches the work you actually do.