The report a machine wrote

The machine can write the report. The responsibility for what the report says has to stay with a person.

At the end of the first essay in this series I wrote that one of the essays to come would be about the report a machine wrote. When I wrote that, I was thinking of the reports the system produces by itself: the numbers of the day, the close of the night. But today there is also a second machine, one that does not produce the numbers but summarises them and writes them up in professional language. This essay is about both.

What I have and what I do not

What I have: in last mile, the daily KPI reports for Damage, Incomplete, Delay and Loading Time, the performance reports by DSP and by driver, and the reporting on operational problems, delays and incomplete orders. In the hotel, during Night Audit, the hotel status and occupancy reports, arrivals, departures and no-shows, the breakfast report and forecast, the end-of-day reports, and the information for the following shift. And the use of ChatGPT to organise operational material, for SOP instructions, to structure information and to draft professional texts.

What I do not have: a case in which I proved that an automatic report was wrong and corrected it. A report that was produced every day and that nobody read. An official report written by AI and sent without a person having looked at it.

So this essay is not about a report that got it wrong. It is about the question I ask every time a report looks finished to me: what does it really tell me, and what does it let me believe?

Two reports, two numbers

In Night Audit I had a case in which the number of people in the breakfast forecast differed from the number of people in the hotel status report. Two reports, from the same system, for the same night, with two different numbers.

The first reaction is to ask which one is wrong. But I cannot say that the system had made a mistake. Two reports can count with different criteria, and show different figures without either of them being inaccurate. What the case asked for was not a correction. It was a check: comparing the two sources, and understanding what each one counts.

This is the part of automatic reports that is most easily forgotten. A generated report looks final because it is orderly: tables, columns, totals. But its orderliness says nothing about the question you have. It says only that the system answered the question it was set, with the criteria it was given.

Two reports with two different numbers do not mean that one of them is wrong. They mean you have to know what each one counts.

Closing the day is not closing the issues

During Night Audit, the end-of-day reports are generated at a certain moment of the night. Before I trusted them, I checked a few things. Whether the information matched the operational state of the shift. Whether arrivals, departures and no-shows had been handled according to the process. Whether there were open issues that needed follow-up. Whether the reports that were linked to each other had discrepancies that needed clearing up. And whether the information the next shift needed had been documented.

The report does none of this by itself. The report closes the day in the system. It does not know whether the guest with the complaint about the shower got an answer, nor whether the morning shift knows it has to follow it up.

Generating a report does not automatically mean that every operational issue has been completed.

In What a Night Auditor learns I wrote about the night as the place where the day is closed and the numbers have to match. This is the other side of the same thing: the numbers can match, and the day can still have something open.

The second machine

I have used ChatGPT to organise operational material, to prepare SOP instructions, to structure information and to draft professional texts. It does well what it is asked to do: it takes a messy list and turns it into something that reads.

That is exactly where the risk lies. A well-written text creates the impression that someone has thought it through well. A report with orderly paragraphs, with words like “verified” and “resolved”, sounds like a checked report, even when nobody has checked anything. And an AI tool can add something that was not in your notes, in the same confident tone as everything else.

Before I use a text prepared with AI, the standard I stand by is this:

1. I verify the facts and figures against the original source. Not against my memory of the source.
2. I check that the AI has not added information that was not there. A name, a time, a “resolved” that I had not written.
3. I separate the resolved issues from the ones still open. The orderly text makes them look the same.
4. I check the language, the recipient and the purpose. The same summary does not go the same way to the next shift and to the manager.
5. I do not put personal data, payment data or confidential information into unauthorised tools. No summary is worth that risk.

AI can help me write a report more clearly. It must not decide by itself what is true.

The report in three parts

What I would do differently today is the structure of the report itself. I would divide it more clearly into three parts.

The facts: what the confirmed data show. The interpretation: what they could mean, and what is still to be verified. The action: what has to be done, by whom, and by when.

And I would put a visible distinction between information that the system or a tool produced automatically, and information that a person has verified. Not because one counts and the other does not. Because the reader has to know which one they are reading.

The reason is simple. I would not want a well-written report to create the impression that the problem has been understood or solved, when in reality it has not yet been checked. The case of the two breakfast numbers, written as a fact, would be “the system made a mistake”. Written in three parts, it would be: two reports give two numbers; they may count with different criteria; someone has to clear it up before the breakfast service.

Who reads it

I do not have a case in which a report was produced and nobody read it. In last mile, the KPIs were part of the operational discussions and of the daily stand-up. In the hotel, the reports were used for closing the day and for the continuity of the work.

But that does not prove that every report was read and used in the same way. And the easier it becomes to produce a report, by the system or by AI, the more reports there will be, and the less time for each one. The question is no longer who writes the report. It is who reads it, and what they do after reading it.

A report should help a person understand reality, not replace the checking of reality.

A report in three parts

The card below is for every report you write or pass on, especially when part of it was generated by the system or by an AI tool: the confirmed facts, what was generated and what was verified, the interpretation as a question, what has been solved and what is still open, and the action with a name and a deadline. Shift Handover splits the shift handover into open and closed issues, so that the night's report leaves the next shift what it needs, not only what the system produced.

The machine can write the report. The responsibility for what the report says has to stay with a person.