You’ve probably sat through this meeting. Someone on your team built something with AI, it’s genuinely clever, and everyone agrees it saves “a ton of time.” Then you ask how much time, and nobody can tell you.

That’s the problem with most agency AI projects. They might be working. You’ll just never know, because nobody wrote down how long the job took before anyone touched it.

Key Takeaways

• If you don’t time a job before you change it, you can’t prove AI saved anything.
• The math is simple: minutes per task, times tasks per week, times the people doing it.
• Take an honest haircut off the result. Part of most jobs is judgment, and judgment stays with a person.
• Pilot on a few accounts and leave a few like them alone, so you’re comparing apples to apples.
• The number that matters most to an agency is how many accounts each person can carry.

Why the “before” number decides everything

AI changes how a job feels long before anyone checks whether it changed how long the job takes. The new version feels faster because it’s new and somebody’s excited about it.

Feelings don’t survive a budget review. When it’s time to decide whether to keep paying for something, or whether to roll it out to the rest of the team, you want a number that was written down before the change. Without it, every AI project turns into an argument about vibes.

So before anything gets built, clock the job the old way.

The math, with a worked example

Here’s the whole formula:

Minutes per task × tasks per week × people doing it = hours the job eats every week.

Let’s run it on a made-up but very normal agency job: a monthly check on each client account, where somebody pulls the numbers, looks for anything off and writes a short note.

Say you’ve got 15 account managers, each with 30 accounts, and each check takes about 30 minutes. That’s 450 checks a month. At 30 minutes apiece, it’s 13,500 minutes, or 225 hours a month. Call it about 52 hours a week of your team’s time going into one recurring job.

Most owners have never seen that number for a job like this. It’s usually bigger than they guessed.

Take the honest haircut

Now the part people skip. Not all 52 hours are up for grabs.

Break the job into its pieces. In a typical account check, roughly half the time goes into pulling and lining up numbers, and the other half goes into reading them and deciding what matters. AI is very good at the first half. The second half is judgment, and that stays with a person.

So the real opportunity in our example is closer to 26 hours a week. Then take a little more off for the time someone spends reviewing what the AI pulled, especially in the first few months. Call it 20 hours a week back.

Twenty is a smaller number than fifty-two. It’s also one you can defend in front of a skeptical partner, and it’s still more than half a person’s week.

Pilot it against a control group

Once the job is clocked, don’t roll the new version out to everyone. Pick a few accounts to run on the new process and a few similar accounts to leave alone.

Then measure both groups the same way: how long the work takes, and whether the results hold up. If the AI-assisted accounts take less time and nothing slips, you’ve got proof. If something does slip, you found out on a handful of accounts instead of all of them.

And keep a person approving every change while the pilot runs. Trust gets earned in the pilot, not assumed on day one.

The number agencies should actually watch

Hours saved is nice. The number that changes an agency is accounts per person.

If each account manager in our example gets 20 hours back across the team, that time can go into better client work, faster responses or more accounts per person without anyone burning out. That’s capacity. It shows up as margin, as room to grow without hiring at the same pace, and as a team that isn’t drowning every month-end.

It’s the number I’d watch in yours.

Frequently Asked Questions

How long should I clock a job before changing it?

Two to four weeks is usually enough to see a normal week and a busy one. If the job is monthly, clock one full cycle. Don’t let the clocking drag on for a quarter; you’re looking for a realistic baseline, not a perfect one.

My team will hate timing themselves. Any way around that?

Keep it light and tell them why. You’re not grading anyone; you’re finding out where the work goes so you can take some of it off their plate. Rough estimates from the people doing the work, checked against a couple of timed samples, are good enough to start.

What if the numbers say a job isn’t worth automating?

Then you just saved yourself a build. That happens more than people expect, and it’s one of the most useful things the math can tell you. Move to the next job on the list.

Want the clocking done for you?

Clocking every recurring job across an agency is exactly what the AI Workforce Audit does: two weeks of interviews and timing, then a written plan with a source on every number. You can see the loop I run after that on How It Works.

Or start smaller. Book a 20-minute fit call and tell me which job you’d clock first.