Business AI guides

By Joshua Hale · September 23, 2026

How to choose your first AI workflow for a small business

Start with a task you already do, can describe clearly and know how to check. A meeting follow-up draft is a better first experiment than asking AI to run your customer relationships.

The useful question is specific: what part of this job could become easier without creating more work somewhere else?

You can answer that before buying another subscription.

Find a small task that keeps coming back

Write down three jobs you repeated recently. Preparing a weekly update. Finding the same business information. Turning a set of notes into a draft.

For each one, ask:

• Does this happen often enough to be worth improving?
• Can I describe the input and the result I need?
• Can I recognize a wrong or incomplete answer?
• Can I try it using information I’m allowed to put into the tool?
• Can I review the result before it affects a customer or colleague?

Choose the task with the clearest answers. If you can’t explain what a good result looks like, clarify the job before adding AI.

For a first experiment, keep sending, publishing and changing records in human hands. You want room to notice a mistake before it becomes someone else’s problem.

Write the job down before you write a prompt

Use this short worksheet:

When this happens: What starts the task?

I provide: Which notes, documents or other inputs are needed?

I need back: What should the output contain, and in what format?

A good result means: How will I judge whether it is useful?

A person checks: Which facts, decisions or omissions require review?

It must not: What is outside the task?

That last line matters. A tool asked to organize notes should not quietly decide who owns a task, invent a due date or send a message on your behalf.

Check the step that happens after the draft

A useful first build can be a page, a booking step or a signup email. Define the finish line before you begin. A page preview lets you check the wording and layout. A booking link needs a separate check. A signup email needs to arrive in the test inbox.

Write down which checks passed and what remains unfinished. If an initial email arrived, that confirms the initial email in that test. It doesn’t confirm every later message in the sequence or tell you whether more people became customers.

Leave time to learn how to change the work afterward. A page you cannot update may create another job for someone else. Ask where the copy lives, how to make a small edit and which changes need help.

Practice defining a task: meeting notes to follow-up

Here is an invented example. These aren’t client notes, and the table below is an illustration of the desired output, not a recorded AI test.

Input:

Tuesday team check-in. Maya will draft the updated service FAQ. We need current prices before publication. Someone should follow up with the designer about the missing image. Next check-in is Friday.

Task instruction:

Using only these notes, make a table of tasks, explicitly named owners and explicitly stated due dates. Mark missing information as unknown. Put unanswered questions below the table. Do not infer that the next meeting is a deadline. Do not send or assign anything.

Illustrative result:

Questions to resolve: Who provides the current prices? Who contacts the designer? When are these tasks due?

Notice what didn’t happen. Friday wasn’t turned into a deadline. Maya wasn’t assigned every task. The missing information stayed visible.

That is a useful result because you can check it against a short source. If an actual tool fills in those gaps without evidence, correct it and record the problem.

Keep the information boundary clear

Start with a made-up or appropriately redacted example when you can. Don’t use customer records or internal documents just because they are convenient.

Check whether the specific tool, account settings and your business’s permissions allow the information you plan to use. Different products and plans can handle information differently. If you aren’t sure, keep the first test free of that information until you have an answer.

AI-generated text can sound confident while being wrong. NIST identifies this as a risk of generative AI. A polished answer still needs to be compared with its source. NIST’s Generative AI Profile.

Try a small pilot

Use five representative examples as a manageable starting exercise. Five isn’t proof of reliability; it is a way to discover obvious problems before expanding the work.

Include an incomplete example, not just the cleanest one. Keep the original input, the draft and your corrections together. Look for repeated mistakes: invented facts, missed tasks, changed meaning or instructions the tool didn’t follow.

Track the time for the whole job:

Prepare the input + run the tool + review + correct + finish the work.

Compare that with doing the same kind of task your usual way. Keep setup time and subscription cost visible, too. A fast draft that takes longer to repair has not saved you time.

Then make a decision. Keep the workflow if it helps and you can check it. Narrow it if part of the job works well. Stop if review takes too much effort or important mistakes are hard to catch.

Write down the version that works

Keep the input requirements, instructions, review checklist and known limits somewhere you can find them. Note which tool you used and when you checked the process.

A useful business system includes the person who runs it and the decisions they still need to make. Someone should be able to return to your notes next week and understand the job.

If you’re a Bellingham business owner who would rather work through that first task with someone, I offer practical AI training and system building. Bring a recurring job and an example you’re allowed to share. We’ll discuss what help would fit before agreeing on the work.