Maybe this task doesn’t need AI at all

Maybe this task doesn’t need AI at all

Imagine you spend every Friday turning a spreadsheet into a weekly report.

It takes thirty minutes. Most of the steps are the same each time.

It sounds like a reasonable task to give AI.

But first, I’d want to know what is actually happening inside those thirty minutes.

Are you:

• moving numbers into the right places?
• calculating the same totals?
• checking that the numbers are correct?
• reading customer comments?
• deciding why sales changed and what to do next?

We call all of this “making the report.” But it is not one kind of work.

Some parts follow clear rules. Some require interpretation. Some require judgment.

That difference matters more than the tool.

Start with the parts that follow rules

If you are adding the same numbers and placing them in the same boxes, a formula or template may be enough.

You could use AI to help you build that solution. Once it works, the report may run without asking a model to interpret the numbers every Friday.

I think this is easy to overlook:

AI can help build the solution even when the task does not need AI every time.

Before adding AI to a repeated task, check whether it can be handled by:

• a clear rule
• a spreadsheet formula
• a reusable template
• an existing feature in the software
• normal automation

The simpler option may be cheaper, faster, and easier to check.

Use AI where interpretation helps

Now suppose the report includes hundreds of customer comments.

Someone needs to read them, notice recurring problems, and explain what might be happening.

That is a more useful place to test AI.

It could:

• group similar comments
• identify recurring themes
• draft a short summary
• point to comments that deserve a closer look

A person would still need to check whether the summary represents the comments fairly. And the final business decision would still need judgment.

The question is now smaller and more useful:

Which part of this task would benefit from AI?

Anthropic gives similar advice when discussing AI systems: start with the simplest workable approach and add complexity when it improves the result. More capable systems can also add cost and waiting time. Their guidance is written for people building agents, but I think the principle applies to everyday work too: what does the extra machinery actually buy you? Source: Anthropic

Decide what “better” means

Saving time is useful. It is not the only possible improvement.

AI might help you:

• do the same work faster
• produce a better result
• examine information you previously had no time to review
• do useful work that was not practical before

There is evidence that AI assistance can help in the right setting. A study published in The Quarterly Journal of Economics in 2025 examined 5,172 customer-support agents. On average, access to an AI assistant increased the number of issues resolved per hour by 15%. The gains varied across workers. Source: Generative AI at Work

That study tells us something useful about customer support. It does not tell us whether AI will improve your weekly report.

You still need to test that.

Count the whole process

First, write down what a good result must get right.

For the weekly report, that might mean:

• the totals match the source data
• customer comments are represented fairly
• facts and guesses are clearly separated
• the report is ready in time for the meeting

Then test the approach on a few older reports where you already know what happened. Include an awkward example with missing information or conflicting comments.

Count all the work:

• preparing the input
• waiting for the output
• checking it
• correcting it
• getting it ready to use

Suppose the old report took thirty minutes.

The new version takes two minutes to generate and twenty-five minutes to repair.

You saved three minutes.

If setup took four hours, you would need eighty weekly reports to recover that time through those savings alone.

That may still be a useful experiment. You might improve the report or learn something valuable. But you should be clear about what improved.

The kind of mistake matters too.

A formatting problem is easy to notice. A confident explanation based on a misread number could reach a meeting before anyone questions it.

If you cannot check the result confidently, keep AI’s role narrow while you work out how to evaluate it.

Try this on one task

Choose one repeated task from your own week and ask:

• Which parts follow clear rules?
• Where would interpretation or generation actually help?
• Which decisions still need human judgment?
• After setup, review, and corrections, is the result meaningfully better?

You may end up with a formula for the totals, AI to help interpret comments, and a person deciding what the business should do next.

You may also discover that a simpler report solves most of the problem.

I’d count that as progress.

The task takes less effort. The result is useful. And you understand why the solution works.

That gives you a better basis for deciding where AI belongs next.