The Report Reviewer: Five Rules for AI-Written Reports

The new problem

If your team uses Copilot or ChatGPT to write reports, analysis and email, the work did not get worse. It got too polished, and you cannot always tell anymore whether the person who sent it understands what is in it. These are the five things I would put in place this week if I were still running a team. The last one turns AI back on the problem it caused.

1. One page. Whatever the source.
2. The two-minute test. Explain it to me, nothing in your hands.
3. Show your work. One line: your assumption, your source.
4. You signed it. Your name on it means you can defend it.
5. Make it grade itself. A saved reviewer in front of every report.

The five rules

Rule 2 is the place to start. Rule 5 is the build.

Rule 1: One page

Whatever the source, the version that reaches you is one page. Not because AI can't write more, but because more room is exactly where the wrong claim hides. A report that grew from one page to twelve almost never has more facts in it. It has more places to bury the one line you needed to catch.

Tell your team once, plainly: what reaches you is one page, however it was produced.

Rule 2: The two-minute test

A polished document used to be evidence that someone had done the work. Today it is evidence that a tool did. So the next time AI-assisted work lands on your desk, don't open it first. Ask the person to explain it to you in two minutes, with nothing in their hands.

It costs nothing and takes less time than reading the deck would have. It tells you immediately who understands the work they are sending you. If you only do one of the five, do this one.

Rule 3: Show your work

Asking "did you use AI on this?" is the wrong question. Everyone says no, and now you have taught your team to hide it. Instead, add one line to how anything AI-assisted gets submitted to you:

The one line

What were you assuming, and where did the number come from?

One line, every time. Not a policy binder. The sentence to watch for is the confident one with nobody's name behind it: "a critical dependency has been completed," with no source and no owner.

Rule 4: You signed it

Whoever signs a number owns that number, whether they built the model or an analyst did. Nobody gets to say "the spreadsheet did that," and the rule does not change because the tool got smarter.

Say it once to your team, calmly, not as a threat: your name on it means you can defend it, not that you typed every word of it.

Rule 5: Make it grade itself

Slowing the team down or restricting tools does not fix this. It teaches people to do the same thing somewhere you cannot see it. You do not fix a bad scale by asking everyone to eyeball the weight again. You build a better scale and put it in front of everyone.

This is the one from the video: a saved Report Reviewer agent that does the first pass for you. It summarizes any report in five lines and quotes every sentence that states a fact, a number or a conclusion without a named source or owner.

Build it in Microsoft 365 Copilot. Open Agent Builder, start a new agent, and fill in three fields.

1. Name

Report Reviewer

2. Description

First-pass reviewer for draft reports. Summarizes in five lines and flags every claim with no named source or owner.

3. Instructions

You review draft reports before they reach a manager. For any report I paste, do two things. First, compress it to exactly five lines, each a complete sentence, covering what happened, the numbers that matter, what changed, what is at risk, and what is being asked for. Second, list every sentence in the original that states a fact, a number or a conclusion without naming a source or an owner, and quote each one. If there are none, say so plainly. Do not soften or rewrite the flagged sentences. Do not add anything that was not in the report.

Then select Create. That is the whole build. Saved once, it is there for every report from here on.

Try it on this sample. It is invented, and it has three unsourced claims planted in it so you can watch the flagging work. Paste it into your new agent:

Sample report (invented)

Q3 Field Operations Update, Northwind Logistics Regional throughput finished the quarter at 41,200 units, up from 38,600 in Q2, per the warehouse management system export dated Sep 30. On-time delivery held at 94.1% against a 95% target, tracked in the carrier scorecard. The dip in on-time performance is mostly down to the new routing software. Driver retention has clearly improved since the schedule change. Overtime spend was $142,000 for the quarter, confirmed by the payroll report from Finance. Two of the four depots are now running the revised pick process; the remaining two are scheduled for October, owned by the depot operations lead. We expect the routing issues to resolve themselves next quarter. Requesting approval for two additional dock staff at the east depot, at an annualized cost of $96,000.

You should get five lines back, then a list quoting the three sentences with no source or owner: the routing software conclusion, the driver retention claim, and the forecast that the routing issues will resolve themselves. The sentences that name a system, a document or a person are left alone.

Two honest limits. The reviewer flags claims that name no source or owner. It does not check that a named source is real, so it is a first pass, not a verdict. And Agent Builder is part of Microsoft 365 Copilot on a work account. The free personal Copilot does not have it.

No Agent Builder? The idea is the reusable review step, not the specific feature. Paste the instructions above into a chat in Copilot Chat, ChatGPT or Claude, add the report below them, and you get the same five lines and the same flags. In ChatGPT or Claude you can save the instructions in a Project so you only paste the report next time.