The Tool Saved 20 Minutes a Draft. Checking and Fixing Took 12 of Them Back.

A 14-person trading company in Beirut sends about 40 supplier-quote drafts a week. By hand, each one took 25 minutes. With an AI drafting tool, each takes 5. That is 20 minutes saved per draft, 800 minutes a week, 13.3 hours, and the figure has already gone into a slide for the partners.

This is a composite example. The company, the people and every number are illustrative, built from patterns in my work, and none of it describes a real client.

Then a partner asks one question: who reads the drafts before they go out? The senior buyer does. Every one, against the price list and the customer's terms, with dollar lines and lira lines that must never be mixed up. That reading is not in the 800 minutes.

My claim is simple. An AI tool's "time saved" is a gross figure. The number worth quoting is what remains after you subtract the time spent checking its output and the time spent fixing what fails the check. I call that subtraction the Review Tax Count, and it has four columns: Claimed, Check, Fix and Net.

Where the 800 minutes came from

A claimed saving usually arrives from one of three places: the vendor's demonstration, the team's own estimate, or a stopwatch held on the tool alone. All three measure the same cheap thing, which is how long the draft takes to appear.

That part is easy to time because it is visible. Someone clicks, waits, and a draft is on the screen. Checking is different. It happens afterwards, often by a more senior person, in small gaps between calls and WhatsApp messages, and nobody writes it down. Drafting minutes get a number. Reading minutes get absorbed into the day.

So the Claimed column is not false. It is incomplete. It describes what the tool does and says nothing about what the company now has to do because the tool exists. A draft produced in five minutes but read for eight has not saved you the full twenty.

In the composite, Claimed reads 20 minutes per draft, and I leave it exactly as the team reported it. The count does not argue with it. It only stops it from being the last word.

Check, Fix, Net: the review tax in one table

Now add the two columns nobody had. Check is the time to read a draft against its source before it is allowed out. In the composite it is 8 minutes, on every draft, because a wrong price or a wrong currency line is expensive. Fix is the share of drafts that fail the check multiplied by the time a repair takes. Here 30% of drafts need a fix, and a fix takes 14 minutes, so the average is 4.2 minutes per draft.

Check plus Fix is the review tax: 8 + 4.2 = 12.2 minutes per draft. Net is Claimed minus the review tax: 20 - 12.2 = 7.8 minutes.

Composite example, illustrative figures. Net is about 39% of the claim.

The tool still wins. Five hours a week is real, and it is a number someone can say to a partner without flinching. But it is 39% of the claim, and the other 61% went into reading and repairing.

Now ask whose hours these are. Checking is rarely done by the cheapest person in the room. In the composite it is the senior buyer, who knows the price list and the customer terms by heart, and sometimes the owner approving a large quote over WhatsApp before it goes out. If the drafting minutes were saved at a junior desk and the checking minutes are spent at the owner's, the net minutes understate the problem. Count whose time, not only how much.

How to run the count in two weeks

You do not need a project. You need a stopwatch habit and one shared sheet. This is the version I suggest.

1. Pick one repeated task with a clear source to check against. A quote, a summary, a support reply. One tool, one task.
2. Collect a sample of 30 real items over two weeks. Thirty is enough to see a pattern in a small business without turning the exercise into a project.
3. Time the check on every item. Start when the draft is opened and stop when it is approved, and let the person who really does the checking hold the stopwatch. Round to the minute.
4. Tag each item as you check it: accept (sent as it was), light edit (a few corrections) or rewrite (started again or mostly redone).
5. Time the fix on every light edit and rewrite. Your fix rate is the share of items carrying either tag. Your average fix time is the mean of those timings.
6. Compute the Net. Claimed, minus the average check, minus the fix rate multiplied by the average fix time. Put the four numbers on one page and quote Net, with the other three in a footnote.

In the composite, 9 of the 30 items would carry a tag, which is the 30%. Your sample will not be so tidy, and it should not be.

Lowering the tax beats buying speed

Once the count exists, a different decision opens up. The obvious upgrade is a faster or smarter tool, but the Claimed column has little room left: in the composite the drafting already takes 5 minutes. A tool that shaved off another 2 minutes would add 2 minutes of net.

The review tax has more room. Suppose a one-page checklist of the five things the buyer verifies on every quote cuts the check from 8 minutes to 5. Suppose that narrowing the tool to the quote template it handles best, and keeping odd cases away from it, cuts the fix rate from 30% to 20%. Then the tax is 5 + 2.8 = 7.8 minutes, and Net rises from 7.8 to 12.2 minutes per draft, about 8.1 hours a week. These are illustrative numbers, but the shape is general: 4.4 minutes recovered from the tax beats 2 minutes bought from speed.

Limits of the count

Thirty items is a sample, not proof, and a different fortnight can move the numbers. Check time often falls as people learn where the tool goes wrong, so re-time it every quarter instead of quoting an old figure. If your by-hand baseline already included some self-checking, do not subtract it twice. And a tool can be worth keeping when the minutes are modest: a quote that reaches the customer an hour earlier can matter more than the arithmetic.

Whatever net hours remain still have to land somewhere, and I covered that in Why I Stopped Reporting Hours Saved.

What to do this week

Pick one AI tool and one task. For the next two weeks, time the check on every item and tag each as accept, light edit or rewrite. Write Claimed, Check, Fix and Net on one page. The next time someone asks whether the tool pays off, quote the last number and say that it is net of checking.

If you run the count, which of the four columns had you never written down before?

Frequently asked questions

What is the review tax on AI output?

The review tax is the time a company spends checking an AI tool's output and fixing what fails the check, per item. It is the check time plus the share of items that need a fix multiplied by the time a fix takes. Subtract it from the minutes the tool claims to save and what remains is the net saving. In the composite example it is 8 + 4.2 = 12.2 minutes against a claim of 20.

How many items do I need to time to get a usable figure?

I suggest 30 real items collected over two weeks for one tool and one task. That is enough to see how long a typical check takes and how often an item needs a fix, without turning the exercise into a project. It is a sample, not proof, so treat the result as an honest estimate and re-time it later. Even the first five items will show whether the claimed figure is close or far off.

Does checking time shrink as staff get used to the tool?

Often it does, because people learn where the tool tends to go wrong and read those parts first. It does not shrink on its own to zero, and it can grow again if the tool is pointed at new kinds of work. For that reason I re-time a sample every quarter and quote the latest figure, rather than carrying forward the first one.

What if nobody checks the output today?

Then the saving you are reporting is gross, and the risk of errors going out unchecked is not in your numbers at all. Have someone check a sample of 30 items as if the check were required, time it, and record how many needed a fix. The result tells you either what the real net is or how often unchecked mistakes were reaching customers.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.