Verification is a habit, not a mood

The situation

You asked for a one-page comparison of four supplier quotes. What comes back is clean — a table, a recommendation, a note on lead times. You read it twice, both times for whether it sounds right, and it does. You forward it.

On Thursday someone asks which quotes include installation. Your page says one. The answer is two: the fourth quote carries installation on a separate line under the subtotal, and the summary compressed that away. Nothing was invented. Something was dropped, and the draft read exactly the same either way.

That is the part worth noticing. Wrong output does not announce itself by sounding wrong.

The principle

Verification is a fixed routine that runs on every output, and the depth of the check is set by what it costs to be wrong rather than by how the writing reads.

You already do this with human work. You do not re-derive a colleague's spreadsheet before a meeting — you check the two or three cells the decision hangs on, and you check them in the same order every time. The routine is what makes it fast enough to actually happen. Applied to AI output it is four passes, in this order:

• Trace the numbers. Every figure, date and total that came from something you supplied: find it in the source. Anything derived, redo the arithmetic once.
• Check what points outward. Names, quotes, citations, links, standards, regulations, version numbers. This is where invention shows up, because the model is filling from memory rather than from your files.
• Find the load-bearing sentence. One sentence is the one someone will act on. It gets read slowly, on its own.
• Look for what is missing. Wrong content is visible. Absent content is not. Count the items in the source, then count them in the output.

Then size the check. Internal and reversible, where you fix it in a minute: first pass only. Anything that lands on someone else's desk: all four. Anything that leaves the building, moves money, touches safety, or goes in front of a regulator or a board: all four, traced line by line, plus a second human reader. Pick the tier before you open the draft, not after it has charmed you.

Worked example

Shown here in Microsoft 365 Copilot.

Checking is faster when the draft tells you where to look. Send this in the same conversation, after the output:

Mark up your own draft before I use it.

Tag every number, date, and proper name in the output:
[S] copied from the files I gave you — quote the exact line
[D] derived — show the arithmetic
[M] from your own knowledge, not in my files

Then list anything in the source files that a reader of this draft
would expect to see and that you left out.

Do not rewrite the draft.

What comes back:

Supplier B: 42,000 installed [S: "Total incl. installation 42,000",
quote-b.pdf p2]
Lowest of the four [D: 42,000 < 47,500 < 51,200 < 58,900]
Lead time for this class of unit is typically 8 weeks [M — not in
your files]

Left out: quote D lists installation on a separate line under the
subtotal, so its headline price is not comparable to the others.

Three things just happened. A 600-word draft became five items to check. The one line pulled from the model's own memory is now labelled, so you can decide whether an assumed lead time belongs in a document going to procurement. And the omission is stated — the failure you could not have caught by reading, because there was nothing on the page to catch.

The self-marking is triage, not proof. A model can mislabel its own sources. An [S] tag tells you which line to open; it does not tell you that the line says what it claims. You still open it. What you stop doing is reading 600 words hoping something looks off.

In your tool

What you want is the shortest path from a claim back to the thing it came from.

Product surfaces checked 2026-08-04.

Try it

Open the last AI-assisted thing you sent to another person. Five minutes.

Highlight every number, date, name and outward-pointing claim. Mark each one: from your source, derived, or from the model. Then count how many you can confirm in two minutes.

You are done when you have two numbers — how many claims the document makes, and how many you could stand behind if challenged — plus the tier you would now assign it. Most people find the first number larger than expected and the tier higher than the check they actually gave it.

Common failures

• Reading for style. Fluent prose reads as checked prose. The attention that catches a clumsy sentence is not the attention that traces a figure, and only one of them is running while you skim.
• One depth for everything. Uniform effort means either ten minutes spent on a note nobody acts on, or a board number given the same glance as an internal draft. Consequence sets depth; nothing else does.
• Missing the omission. People check what is on the page because that is what the eye is offered. Nine items summarised from a source of eleven produces a page with no defect visible anywhere on it.