The four-hour tax: the time you spend fixing what AI gives you

Nobody counts the time they spend fixing what the AI gave them.

The figures I keep seeing land in the same place. People save around six hours a week with AI, then spend around four of them re-prompting, checking, correcting and reformatting what came back.

Net gain: two hours. I call the difference the four-hour tax.

Why the tax happens

The tax isn't an argument against AI. It's an argument against picking the wrong tasks to give it.

Some work is cheap to check. You glance at the output and know in seconds whether it's right.

Other work is expensive to check. If you have to read every line to know whether it worked, you haven't automated anything (you've moved the work and added a step).

Tasks that pay off

The tasks that pay off are the boring ones. They share three traits:

• They happen often, ideally every day or every week.
• They take the same shape each time: same input, same kind of output.
• A wrong answer is obvious, and cheap to fix.

Sorting incoming email, pulling fields out of invoices, turning meeting notes into a list of actions, formatting a weekly report. If one comes back wrong, you notice straight away and it costs you a minute.

Tasks that eat the gain

The tasks that eat the gain look impressive in a demo. Long documents, nuanced analysis, anything written in your voice for someone important.

The output reads well, which is the problem. Reading well and being right are different things, and the only way to tell them apart is to check everything.

I leave a lot of things alone for exactly this reason. Anything client-facing that carries my name, and anything where being confidently wrong is expensive.

A quick test before you automate anything

Ask three questions about the task:

1. How long does it take to check the output? If it's close to the time the task itself takes, stop here.
2. How bad is a wrong answer that slips through? A misfiled email is fine. A wrong figure in a client proposal isn't.
3. How often do I do this? Frequency multiplies the saving, and it multiplies the tax too.

Low checking cost, low damage, high frequency: automate it. High checking cost or high damage: keep it with a person, and use AI for the fetching and the first draft at most.

Two tasks, one test

Take two tasks a sales team might hand to AI.

The first is logging call notes into the CRM. Checking takes a ten-second glance, a mistake is easy to spot and fix, and it happens fifteen times a week. Automate it.

The second is writing a tailored proposal for a large client. Checking means reading every paragraph, a wrong number could cost the deal, and it happens twice a month. Keep a person on it, and let AI pull together the background research.

Both tasks save time on paper. Only one of them saves time in real life.

Choosing what not to hand over

Choosing what not to automate is most of the skill. It's also the part nobody puts in a demo.

In the audit we run with teams, this is the third bucket: the human-only work. Far from being a limitation, it's what keeps the other two buckets honest.

Before you automate something this week

Ask what it costs you to check it. If the answer is "I'd have to read the whole thing anyway", leave it where it is.

Then find something boring instead. That's usually where the real hours are.