Measuring an admin task before you change it

“Saves five hours a week” is the most common claim in AI marketing, and the hardest to check. Not because suppliers are dishonest, but because almost nobody knows how long the task took before.

Without a baseline, every claim about time saved is an impression. Impressions are not evidence, and they are not what a board, a commissioner or an inspector will find convincing.

The method below takes about two weeks of light effort and gives you something you can actually stand behind.

Three things that go wrong

We have run this exercise enough times to know where it fails. All three failures are avoidable.

People do not know how long tasks take now

Ask a care coordinator how long a referral summary takes and you will get an answer. Ask five, and you will get five answers spanning a factor of three. Estimates are confidently given and unreliable, which is the worst combination.

This is why the measuring comes first. A baseline reconstructed after the fact is a guess wearing a number.

New tools are slow at first

In the first weeks with any new tool, staff take longer, not less. They are learning the interface and checking the output carefully, which is exactly what you want them doing.

Measure in week one and a genuinely useful tool will look like a failure. Give it time to settle before you take the “after” measurement, and say in your write-up when you took it.

Checking takes time too

This is the one that invalidates most claims.

Time saved on drafting only counts if the drafting-plus-checking time is lower than the original writing time. A tool that produces a draft in ten seconds has saved nothing if a care worker then spends four minutes reading it against what actually happened.

So measure the whole task, end to end, including the check. Anything else is measuring the tool rather than the work.

The method

Step 1: define the task precisely

One sentence, with a clear start and end. For example: “From opening the referral email to saving the summary in the system.”

Vague boundaries are the most common reason two measurements cannot be compared. If one person starts the clock when they open the email and another when they start typing, you have two different tasks.

Step 2: record each instance

Staff note the start and end time each time they do the task. Rough times are fine — to the nearest minute is plenty.

Aim for at least 20 records, over at least two weeks so you catch a normal spread of days. Our admin time tracker has the columns already set up.

Keep it genuinely light. If recording the task takes real effort, people will stop, or worse, fill it in from memory at the end of the week.

Step 3: use the median, not the average

One chaotic afternoon can drag an average a long way. The median tells you about a typical instance, which is what you want.

In any spreadsheet: =MEDIAN(D2:D40).

Record the spread as well — the fastest and slowest — because a task that varies from four minutes to forty is telling you something an average would hide.

Step 4: record quality, not just time

For each instance, note whether the finished work needed correcting, and roughly how much.

A faster process that produces more errors is not an improvement, and in care records it is a risk. This column is the one that stops a time-saving from being reported as a success when it was not.

Step 5: repeat it afterwards, the same way

Same task definition, same people if you can, same recording method. Take the “after” measurement once the tool has settled — a few weeks in, not on day three.

Then you can say something specific: the median time went from 14 minutes to 9, and corrections went from 1 in 10 to 1 in 4. That is a finding. It may even be an argument against continuing, which is what makes it worth having.

What to report

For each task, three numbers and one sentence:

• Time before — median, with the range
• Time after, including checking — median, with the range, and when it was taken
• Corrections — how often the output needed fixing, before and after
• What you could not measure — always something

That last line matters. Quality of the record, staff confidence, whether the thing people did with the saved time was useful — these are real effects that a stopwatch does not capture. Saying so is more credible than pretending the numbers cover everything.

A word on honesty

If you measure properly, some results will be disappointing. A tool that saves ninety seconds on a task done twice a day is saving about six hours a year, which is not nothing but is not the pitch either.

Publishing that internally is the point. A service that can say “we measured it, it saved less than we hoped, here is what we did next” is in a stronger position than one that can only say it felt faster.

We have not published our own figures anywhere on this site, because we do not have any yet that came from real cohorts. When we do, they will appear with the same caveats we are asking you to apply.