Guides · AI at work
Does AI Save Time? Time One Job and Find Out
By Josh Croom · October 27, 2026 · 10 min read
AI can save time, but often less than it feels like, because the minutes spent checking and fixing its work count too. In a 2025 timed study, experienced coders felt 20% faster with AI while the clock said they were 19% slower, so time one job by hand and then run AI on it for 30 days.
The question nobody could answer
Friday, 4:30 PM. The owner of a small roofing company asks her office manager a simple question before he leaves for the weekend: is the AI tool worth what we pay for it? He says it saves him hours on quotes. She asks how many hours, and he thinks about it for a moment. Then he admits he doesn't know, because he never timed a quote before he started using AI.
His feeling may even be right, but a feeling can't go into a budget or answer a partner's question. It also leaves out the minutes he spends reading each draft, catching the wrong prices and fixing the parts that don't sound like the company.
What a timed trial found
METR, a research group, ran a randomized trial with 16 experienced software developers working on 246 real tasks. The tasks came from projects the developers had worked on for about five years on average, so they knew the work well.
- 24%faster, they guessedbefore the work started
- 20%faster, they feltafter the work was done
- 19%slower, on the clockthe timed result
Why AI feels faster than it is
AI makes a first draft appear in seconds, and that moment is the one people remember when they think about the tool.
The minutes after that are easy to forget. Someone reads the draft, catches the wrong price, rewrites the paragraph that sounds like nobody, and checks the facts before anything goes out. Each of those steps feels like ordinary work rather than AI work, so nobody adds it to the total. When you count only the minutes to a first draft, AI will almost always look faster than it really is.
The METR trial caught the same gap between the feeling and the clock. The developers were experienced coders working in code they knew well, and they still overestimated the gain both before the work and after it.
Most of the new tasks are setup and checking
Workers in the same study who used the chatbots reported saving 2.8% of their work hours on average. In a 40-hour week, that's a little over an hour. The authors set that against controlled trials in the same kinds of jobs, which often found gains above 15%. An ordinary office week can land far from what a careful trial found, so measure your own work before you plan around anyone else's number.
AI studies go out of date quickly
AI tools keep changing, so every study describes the tools as they were on the day it ran.
In February 2026, METR said it was changing how it runs its developer study. Between 30% and 50% of the developers it surveyed said they were choosing not to submit some tasks, because they didn't want to do them without AI. METR said that meant its study was missing the tasks where developers expected AI to help the most, and it began redesigning the experiment.
So run your own test. A 30-day test on one of your own jobs, like quotes or follow-up emails, measures the tool you use today on work you know.
Going by feel, or going by the clock
Going by feel
- Counts the minutes to a first draft.
- Skips the time spent checking and fixing.
- Tries AI on five jobs at once.
- Calls it a win or a fail with no starting number.
Going by the clock
- Times the old way for two weeks first.
- Counts drafting, checking, fixing and cleanup.
- Tests one job with one named owner.
- Decides on day 30 against a line set on day one.
The 30-day AI test, step by step
The whole test takes about six weeks: two weeks of timing the old way, then 30 days with AI. It works for a two-person office as well as a large staff, because it only ever measures one job at a time.
Before you start
Pick one job and one owner
List the jobs that fill your week, and take the one that eats the most hours. Give that one job to one named person, so somebody owns the count. The job should come up at least 20 times during the test, because a handful of tries only measures luck.
Two weeks by hand
Time it by hand first
Don't build anything yet. Write down how many minutes the job takes the old way, every single time, because that number becomes the starting line you'll compare everything else against.
Before day 1
Set the keep line and a mistake limit
Decide now what result on day 30 earns a yes, and how many errors that reach a customer you'll accept, like one in fifty. Ask each person to guess how much faster AI will make them, and write the guesses down so you can compare them with the clock later.
All month
Keep three rules
Nothing goes out unread, nothing private goes into the tool, and a person checks anything that would cost a lot if it were wrong. If you reach your mistake limit, stop using AI on that job, however much time it seems to save.
Every run
Write down the minutes and the misses
Track the minutes to a first draft, the minutes spent checking and the minutes spent fixing. Count the errors you caught, along with the ones that got out and the person who had to clean them up.
Weeks 1 and 2 with AI
Make the prompt better
In the first week, run AI on real work at least three times and edit everything it gives you. In the second week, add what those first drafts were missing to your instructions, and save the better prompt so it becomes the default.
Every Friday
Meet for fifteen minutes
Sit down with the owner for fifteen minutes each Friday, and write down that week's number and what changed. On day 30 you'll have a record to decide from instead of a memory.
Day 30
Decide out loud
Start with the time the job took by hand, then take away the new time, the checking, the fixing and the cleanup of errors that got out. Put what's left next to your keep line, and decide whether to keep AI on the job, fix how you use it or stop. A stop should mean the numbers showed no gain. When nobody kept the numbers, run the test again with an owner who will, because an unmeasured month can't tell you anything.
The math on day 30
The real gain is the time the job took by hand, minus everything the AI version costs you. Here's a sample for a team that writes repair quotes.
The first draft took 10 minutes instead of 40, which felt like a 30-minute win. The real gain was 10 minutes a quote, a third of what it felt like, and that's still worth keeping. At 25 quotes a week, 10 minutes each adds up to more than 4 hours a week, and over 50 weeks that's more than 200 hours a year.
| Each quote | Minutes |
|---|---|
| By hand, before the test | 40 |
| First draft with AI | 10 |
| Checking the draft | 12 |
| Fixing what was wrong | 6 |
| Cleaning up errors that got out, averaged | 2 |
| Real gain per quote | 10 |
The Friday check-in can be as short as a text, as long as the number counts every minute the job took.
What's this week's number on quotes?
30 minutes each, all in. Down from 40 by hand. We caught 2 wrong prices before they went out.
Good. Keep the new prompt, and send me next Friday's number.
Is your AI test a real test?
Answer five quick questions about the way your team measures AI right now.
Did you time the job by hand before you started using AI?
Does one named person own the test?
Do you count the minutes spent checking and fixing?
Did you set a keep line and a mistake limit before day one?
Will the job come up at least 20 times during the test?
You're running a real test. Keep the Friday numbers going, and compare day 30 with the guesses people made on day one.
You're close to a real test. Fix the first question you answered No before the next run, because each gap makes the result look better than the truth.
Start over with one job and one owner, and time the old way for two weeks. Without a starting number, the test can't tell you anything the feeling hasn't already told you.
Fill in the test card before day one
Tick each one off as you fill it in. Your ticks stay on this device.
0 of 6 done
Say what AI is worth in one sentence
When someone asks what AI is worth to your team, give them a number in one sentence instead of a feeling.
Start by turning the hours into dollars. Multiply the hours saved each week by what an hour of team time costs you, and then multiply by 50 weeks to get your yearly number. Divide that by what one typical sale, job or gift brings in, and you can say what AI is worth in your own terms.
Then try it this way: AI gives our team back (hours) a week. That's worth about (dollars) a year, and we're putting those hours into (what you'll do with them). Every part of that sentence comes from your own records, so a board member or a doubtful partner can check it.
Then pick next month's job, because one job a month adds up to twelve tested jobs a year. If the first drafts keep missing your voice or your prices, teach AI about your business before you start the clock, so the test measures the tool at its best.
Talk it through with us
We run Google Ads for small businesses, nonprofits and online stores, and every call and form our ads bring in shows up in your portal with what it cost. If you'd like to talk about where your leads and your team's hours go, book a call with Miguel.
Questions people ask
Does AI really save time at work?
It can, but often less than people feel. In a 2025 timed trial, experienced coders took 19% longer with AI while believing they were 20% faster. To know for your own team, time one job by hand, then with AI for 30 days, and count the checking and fixing.
How do you measure AI time savings?
Time the job by hand for two weeks to get a starting number. Then track the minutes for drafting, checking and fixing on every AI run for 30 days. The real gain is the starting number minus all of those minutes.
Why does AI feel faster than it is?
The first draft shows up in seconds, and that's the moment people remember. The minutes spent reading, checking and fixing feel like ordinary work, so people rarely count them.
How long should an AI time test run?
Plan on two weeks of timing the old way, then 30 days with AI. The job should come up at least 20 times during the test, so pick one that repeats often.
See how we run Google Ads, with every call and form counted
Read next
- Call Log Template: What to Write Down From Every CallA ready-to-copy call log, the one question to ask every caller, and how to sort wrong numbers.
- ChatGPT Custom Instructions Examples: Make Each Fix Only OnceEight custom instructions examples to copy, and a four-line card that carries your fixes into every chat.
- ChatGPT Prompt Templates: Write the Weekly Ask Once and Save ItCopy-ready ChatGPT prompt templates for the jobs you run every week, and how to save each one.