Most AI pitches lead with a percentage: "cut admin time by half." I don't use those numbers, because nobody selling the tool has timed your work.

What you can know is how much time a task takes today. I call that the hours at stake. It's the most AI could possibly give back, before setup, checking and mistakes. It's a ceiling, not a promise. If the ceiling is low, no tool will make the task worth automating.

The formula

Hours per year = times per week × minutes each time ÷ 60 × people doing it × 48 working weeks

I use 48 weeks to allow for holidays and time off. Change it if your year looks different.

Here's an example with made-up numbers. It's an illustration, not a benchmark:

  • Task: entering vendor bills into QuickBooks
  • 60 bills a week, about 4 minutes each, 1 person
  • 60 × 4 ÷ 60 × 1 × 48 = 192 hours a year

If you know roughly what that person's time costs, multiply. At $30 an hour (also made up), that's $5,760 a year at stake.

Now compare a second task: a monthly report that takes the owner 90 minutes. That's 1.5 hours × 12 = 18 hours a year. It might feel more painful, but there's a lot less at stake.

Time it for two weeks first

Estimates are almost always wrong, usually low. Before you decide anything, time the task for two weeks.

  • Tally sheet. A sticky note or spreadsheet with a mark for each time the task happens.
  • Time a sample. Time 10 instances start to finish with a phone timer. You don't need to time all of them.
  • Count the fixing. If a bill gets entered wrong and someone fixes it later, that's part of the task.
  • Count the retyping. If the result gets typed into a second system, like from an email into QuickBooks or from a spreadsheet into your CRM, time that too. Retyping is often where the hours hide.

At the end of two weeks you have a before number. Without it, you can't tell later whether the tool helped.

Four questions beyond the hours

  1. Is it the same each time? Repeatable work, like the same vendor or the same fields, fits well. If it's different every time, write the steps down first. AI won't fix a process nobody can describe.
  2. Does it start digital? Emails, spreadsheets and data already in your software are easy to work with. PDFs and scans are workable. Paper and phone calls need to get into a digital form first.
  3. Does someone retype the result? If so, the answer may be connecting the two systems or producing an import file, not an AI tool.
  4. What does a mistake cost? If an error means money, legal trouble or a lost customer, keep a person checking every item. That checking time counts against the hours at stake.

A task with high hours that's repeatable, digital and low-risk is a strong candidate. A task with low hours that's different every time and high-risk is one to leave alone for now.

Then compare it to the cost

Once you have hours at stake, put the tool's price next to it. Include setup time and the time someone spends checking output.

The rule of thumb I use: if the hours at stake are small and the tool needs real setup and checking, skip it. Put your effort where the hours are.

Test on 10 real examples

If a task passes, don't roll a tool out to everyone. Try it on 10 real examples, with a person checking each one, and time it the same way you timed the before number. Include the checking.

  • Week 1: time the task (if you haven't already)
  • Week 2: try the tool on 10 real examples, with a person checking each
  • Week 3: compare before and after minutes, then decide: keep, adjust or drop
  • Week 4: if you keep it, write down the new steps so it doesn't live in one person's head

Do this without a spreadsheet

The Where should AI start? prompt does this math for up to 10 tasks. Paste it into your business AI account. It interviews you one question at a time, then ranks the work and names what to leave alone.

Example figures in this post are made up for illustration. This is general information, not tax, accounting or legal advice.