Your support team spends 4,000 hours a year typing things a computer already knows.

Before anyone tries to sell you an AI transformation, do this exercise. It costs three days of one person's attention and it will tell you more than any vendor demo.

Take your last 500 resolved support tickets. Ignore the subject lines entirely — subject lines are how customers describe problems and they are useless for this. Instead, for each ticket, write down the action your agent finally took.

The list is shorter than you think

Every time we have run this, the same thing happens. Somewhere between twelve and twenty distinct actions cover more than 90% of the volume. A support organisation that feels infinitely varied from the inside turns out, from the outside, to be about fifteen things done over and over.

A representative result:

Those six rows are just over 3,000 hours a year. Add the next four and you are past 4,000. That is roughly two full-time people, and not one of those hours involves a judgement call, a difficult conversation, or anything a person would describe as their job.

Two columns, and the second one is the important one

Now add a second column to every row: what is the worst thing that happens if this is done wrong, and can it be undone?

• Pasting a tracking number wrong: customer is briefly confused. Trivially reversible.
• Issuing a $40 credit wrongly: you are out $40. Reversible.
• Changing a shipping address wrongly: freight goes to the wrong place. Expensive, mostly recoverable.
• Cancelling a contract wrongly: you may have lost a customer. Not reversible.

Sort by hours descending. Filter to reversible. The top of that list is your automation roadmap, in order, and you built it yourself in three days without talking to a single vendor.

Why do the counting first

Three reasons, and they are all about not wasting money.

It tells you whether you have a problem worth solving. If your top ten actions total 400 hours a year, an AI programme is not your best investment and anyone who tells you otherwise is selling. We have told companies this and walked away from the work. The assessment is cheap precisely so that the answer can be no.

It changes what you ask for. A company that has done this exercise does not ask for "an AI chatbot". It asks for "something that handles tracking, POD and address corrections, with approval on the address changes". That is a scoped, testable, thirty-day project. The other thing is a nine-month disappointment.

It gives you a baseline. The most common reason AI projects cannot prove value is that nobody measured the before. If you do not know that a short-shipment credit took 9 minutes in March, you cannot claim anything in September.

The objection you will hear from your own team

"Our tickets are more complicated than that." They will say it sincerely and they will be partly right: the hard 10% is genuinely hard, and it is also what they remember, because the boring 90% does not stick in the memory. That is exactly why you count instead of asking.

It is also the honest reassurance to give them. Nobody's job is 4,000 hours of pasting tracking numbers. Taking that away does not remove a person; it removes the part of the day everyone hated, and it is worth saying so out loud before the project starts rather than letting people work it out for themselves.

We do this as a fixed-fee assessment and hand you the spreadsheet whether or not you hire us for anything else. But you do not need us to do it. You need somebody to spend three days counting.