Time the task before you automate it, or you will never know what AI saved

KPMG released the first edition of its Global AI Pulse last week, a survey of 2,110 senior leaders in 20 countries at companies with at least $100 million in revenue. According to IT Brief’s coverage, 74 percent said AI would stay a top investment priority even if a recession hit in the next year. The average planned AI spend over the coming year was $186 million.

The same survey found that 64 percent of organizations report meaningful business value from AI. Reported is the word to notice. Steve Chase, KPMG International’s global head of AI and digital innovation, said it directly: “spending more on AI is not the same as creating value.”

By the numbers

74%

of senior leaders would keep AI a top investment priority even in a recession

Source: itbrief.co.uk

Budgets are rising faster than proof. And the simplest way to get proof is something most companies skip: timing the work before the AI arrives.

Why “it saves time” rarely survives a close look

The UK’s Department for Work and Pensions published a careful evaluation of its Microsoft 365 Copilot trial in January, and it is a useful example of how hard measurement gets when you start late. The trial ran from October 2024 to March 2025 with 3,549 staff. About 90 percent of the people surveyed said Copilot saved them time, and 30 percent of those said it saved more than an hour a day.

The evaluators didn’t take those answers at face value. They compared license holders with a matched group of non-users and estimated a saving of 19 minutes a day on routine tasks. They also listed the limits of their own work. Every outcome was self-reported. Licenses weren’t randomly assigned, which the authors said may overstate the benefit. And there was no baseline data, so comparisons depended on people remembering how long things used to take.

That last point is the one to learn from. A government department with an evaluation team still couldn’t measure before and after, because nobody timed the work before the tool showed up.

Checking the output also eats into the savings. In March, document software company Foxit published research, reported by TechRadar, finding that executives spent about 4 hours and 20 minutes a week validating AI output. After that checking, net weekly savings came to about 16 minutes for executives and 14 for workers.

Before and after

Measured by asking

People estimate the time saved
Nobody recorded the old time
Review time is left out
The number is hard to defend

Measured by timing

The task is timed before launch
The same task is timed after
Review and fixes are counted
The number holds up

The second method takes one extra week at the start and saves months of argument later.

How to time a task in five days

You don’t need software or a consultant. You need one workflow, one manager and a spreadsheet.

Pick the unit of work. One invoice matched. One support email answered. One lead researched and logged. The unit has to be the same before and after, or the comparison means nothing.

Count the volume. How many units came in last month? Pull it from the system if you can. Volume times minutes is where the hours come from.

Time a sample. For five working days, the people doing the task log start and stop times on a sample of items, say every fifth one. If that is too fiddly, have them log total hours spent on the task each day and divide by the count. Both work. Guessing doesn’t.

Note the errors and the waiting. Record how many items had to be redone and how long items sat before someone picked them up. Sometimes the biggest win from an agent is the wait time, which a stopwatch on the work itself won’t catch.

Write it down and get the manager to sign it. This is your baseline. It should fit on one line. For example: 1,000 invoices a month, 7 minutes each, 4 percent redone, 2-day average wait.

FlowMeasuring one workflow, start to finish

01Count last month's volume
02Time a sample for five days
03Launch the agent with review
04Time the same task again
05Compare minutes and errors

The before and after measurements use the same unit, the same people and the same method.

Count the review time, or your number is wrong

After launch, repeat the same timing, and include everything a person still does. If an agent drafts the reply and a rep reads and edits it, the new time per unit is the rep’s reading and editing. If 1 in 10 drafts gets thrown out and rewritten, those minutes count too. This is where the Foxit finding bites: a draft that arrives in seconds but takes ten minutes to verify has saved nothing.

Here is a worked example. Say your support team answers 1,200 emails a month at 6 minutes each. That is 120 hours. After launch, an agent drafts every reply. A rep spends 2 minutes reviewing each one, and 1 in 10 needs a full rewrite at the old 6 minutes. That is 40 hours of review plus 12 hours of rewrites, 52 hours in total. You saved 68 hours a month, and you can show your working.

If you had asked the team instead, they might have said the agent saves them “loads of time.” That answer doesn’t survive a budget meeting. The 68 hours does.

Sometimes the number comes back small. That is useful too. If the review time is eating most of the savings, the fix is usually one of three things: the agent is missing data it needs, its instructions are vague, or you picked a task where every answer needs a close read. Each of those shows up clearly in your timing log, because you can see which items took the reviewer longest. Fix the cause, time it again for a week, and compare. Without a baseline you would only know that people felt the tool was “fine,” and you would keep paying for it anyway.

Track errors alongside minutes. An agent that halves the time per invoice but doubles the redo rate may not have saved you anything once the redos are counted.

Decide where the hours go before you get them

One more step makes the number real. Saved time that isn’t assigned disappears into email and meetings, and a month later nobody can point to it. Before launch, the manager should decide what the freed hours are for: clearing a backlog, calling at-risk customers, taking on a function you would otherwise hire for. Then you can measure that too.

This is also how you protect the people involved. When the plan says the 68 hours go to proactive customer calls, the team knows what their new week looks like. That is a better conversation than a vague promise that AI will make everyone more productive.

What to do this week

Pick the workflow you plan to hand to AI first and name the unit of work.
Pull last month’s volume for that unit from your systems.
Ask the team to log time on a sample of items for the next five working days, along with redo counts and wait times.
Write the baseline in one line and have the manager sign it.
Decide, in writing, what the saved hours will be used for if the project works.

Do not turn anything on until the baseline is written down. Once the agent is live, the old timing is gone for good.