AI Without the Hype · Part 1 of 10
You do not need an AI strategy for every corner of your business before you try something useful. You do need a better starting point than buying subscriptions and hoping your staff figures it out.
My recommendation: start with one recurring task that takes more effort than it should, has a clear definition of good work, and can be checked by someone who knows the job.
That could mean turning rough job notes into a draft customer update. It could mean sorting incoming inquiries for an employee to review. It might mean preparing a first draft of instructions from an existing procedure.
The task matters more than how impressive the demonstration looks.
Follow the work that keeps coming back
Ask the people doing the work where they repeat themselves. Where are they rewriting the same explanation, pulling details out of a long message, or turning scattered notes into something another person can use?
Listen for specific tasks. “Email takes forever” is too broad. “Every afternoon I turn the technicians’ notes into customer updates” is a place to investigate.
For each candidate, write down how often it happens, how long it takes, what information goes in, and what someone needs at the end. Include what happens when the task goes wrong. An awkward internal summary and an incorrect customer quote have different consequences.
Keep the list short. Three real tasks are enough to start a useful conversation.
Choose work you already know how to judge
Imagine a small repair business. Its office manager takes technicians’ notes and writes a short explanation of completed work. A good update says what was found, what was done, and what remains unresolved. It avoids jargon and does not invent a diagnosis or promise a completion date.
That is a sensible task to test because the office manager can compare the draft with the source notes and recognize a problem.
Now imagine asking AI to choose a repair method the team does not understand. Who is going to catch a convincing but unsuitable recommendation? If you lack the expertise to evaluate an answer, the review problem gets harder.
Start where your team’s existing knowledge gives you a way to check the work. You can explore unfamiliar territory later with appropriate expertise involved.
Check the inputs before blaming the tool
If the notes say only “fixed issue,” the draft cannot reliably explain what happened. Someone has to supply that information.
In our repair example, a simple note template might ask for the reported problem, findings, work completed, unresolved items, and next action. That improvement may help even before AI enters the process.
Use fictional examples or information approved for the tool during the first test. There is no reason to upload a folder of customer records just to find out whether a draft format is useful.
Keep the first attempt inside a small boundary
Give the tool one job: produce a draft update from the supplied notes. Have the office manager approve every update before it leaves the business. Leave scheduling, pricing, and changes to customer records outside the test.
This boundary makes the result easier to assess. If you connect five systems and change three procedures at once, you will have a much harder time figuring out what helped and what caused trouble.
Run several examples, including incomplete notes and a case with an unresolved issue. Look for extra checking work as carefully as you look for time saved.
What would make this worth keeping?
Suppose an update usually takes eight minutes. The AI produces a draft in seconds, but the office manager spends seven minutes correcting it. That might still improve consistency, but the time saving is small. Measure the whole task through approval.
Also ask whether the result is easier for the customer to understand. Speed is useful only if the work remains acceptable.
Write a short decision before expanding: keep it, change the approach, or stop. Include what you learned about the inputs, the review effort, and the exceptions.
Anthropic’s engineering guidance recommends starting with simple approaches and adding complexity when it is justified. For a small business, I would apply the same discipline to the work itself: get one useful task working before building a larger system around it.
Your first step this week
Sit with the person who handles a repetitive task and watch one example from beginning to end. Ask them what they check, what they correct, and what experience tells them that the written procedure leaves out.
Choose a task from that conversation. You will have a much better test than anything that begins with “we should probably be using AI.”
Source
Anthropic: Building effective agents. The repair-business example and selection method above are illustrative recommendations, not reported client results.