AI automation: which reports disappear, and which stay

A recurring report usually has three parts. Someone puts it together, someone receives it, and someone then decides something differently than they would have without it. The first part runs reliably in most companies. The third one gets thin in our experience, and that rarely stands out, precisely because the first part runs so reliably.

That is why "which report do we automate first?" is rarely the first question. The cheaper question comes before it. Does this report still trigger a decision? If not, the best AI automation is the one you do not build. You cut the report and save yourself both the work of today and the system of tomorrow.

When we go through the recurring work with a management team, reports are almost always at the top of the list, and there is a reason for that. They are visible, they repeat on a fixed rhythm, and everyone in the building can say in one sentence how long they take. That makes them the obvious first candidate for AI automation. It does not automatically make them the right one.

We take this list early in a mandate, because it costs nothing and the list gets shorter before anyone starts building. The sorting runs on three questions and ends in four groups. Only two of them lead to work on a system, and only one to a rebuild.

The three questions that decide a recurring report

Take a specific report. The weekly sales report, the monthly summary for the management team, the reporting to a client, the list of open items.

Which decision depends on it? Not "who reads it", but: who has done something differently after reading it in the past six months? A report that only confirms that everything is as expected is not a decision report. It is reassurance. Reassurance is not worthless, but it does not need a weekly rhythm and half a working week per year.

Where do the numbers come from? From one system, from two or three systems, or from people's heads. That is the cost question. A report whose numbers sit in one system is usually an export that nobody ever set up. Two or three systems plus a spreadsheet in between is the normal case and the actual time sink. If part of it comes from someone's head, as in "I know that this item has not been invoiced yet", then that part is not a data problem. It is a judgement, and judgements belong in group three.

What happens if it does not arrive for two weeks? The most uncomfortable of the three questions, and the only one with just two possible answers. Someone asks, or nobody asks. Whoever asks will tell you in the same email what they needed it for, and that is the most precise requirement you will ever get for this report.

If a report survives all three questions, the rebuild is worth it. For that we work with a format that has a date: an AI automation in four weeks, one workflow, measured against a number that was fixed beforehand.

Four groups, and what happens to each

Group 1: gets cut

No decision, nobody asks. The report comes from a meeting that no longer exists, or from a period in which one number was critical and has not been since. It keeps running because someone set it up and nobody feels entitled to switch it off.

This group does not get automated. It gets cut. It is the only part of the exercise that gives time back immediately without anyone building something. In our experience it is larger than the management team expects, and the most common find is a summary that was built once for a meeting that has not existed since a reorganisation.

Group 2: becomes a query

The decision exists, but it does not come at the rhythm of the report. Once a quarter someone needs the open items older than 30 days, and that is why they have been on the table every Monday for four years.

What disappears here is not the content, it is the rhythm. A weekly report becomes a question that someone asks when they need it. That saves work without anyone building something, because nothing is prepared for the worst case any more, only for the case that actually occurs. What it takes is access to the data and a clean question. The difference to group three matters, because no workflow is created here that someone has to maintain afterwards.

Group 3: stays, but gets prepared

A decision depends on it, and it needs a judgement. That is the normal case for everything that goes outside: reporting to clients, to owners, to the board of directors. Nobody wants to receive a report that no person has looked at, and that is a reasonable expectation, not laziness.

What falls away is the searching, reconciling and formatting. What has to stay is the sentence only someone who knows the client can write. Whoever produces such a report today can tell you fairly precisely how the time splits between searching and judging. If the larger part sits in the searching, that justifies the rebuild, and more clearly than in any other group, because the same work comes back every month. If most of the time goes into the judging, you have a well organised report and a different problem.

Group 4: stays as it is

Two kinds. The first are reports required by a contract, the auditors or an authority. Here the question about the decision is the wrong question, because the obligation is the purpose. Form and deadline are fixed, and simplifying how it is put together is often still worth it.

The second kind are reports whose value lies in being read together. If eight people sit in front of the same number once a month and disagree about it, then the meeting is the purpose and the document is only the occasion. Such reports survive any automation, and rightly so.

The report is rarely the whole workflow

One addition to the sorting, and it makes the difference in groups two and three. "The monthly report costs us two days" almost never means that writing it takes two days. The time goes into rounding up the inputs beforehand, into the email to sales, the missing response from production, the question of whether the last correction is already in there. The report is the visible part, and usually the shorter one.

That shifts the starting point. An AI automation that only takes over the last hour, meaning the formatting and the assembly, leaves the actual work untouched. One that starts at the inputs goes after the place where the time goes. Who delivered what, what is missing, who was reminded and when, visible to everyone who is waiting for it.

So it is worth adding one line to every report in groups two and three: how many people have to contribute something beforehand? At zero it is a data problem and usually solved quickly. From two onwards the coordination comes with it, and that is often the bigger item.

Where AI automation of reports gets expensive

The half rebuilt report is the most expensive case, because you pay for both. The assembly runs, but someone still recalculates the numbers every time, because nobody decided when the recalculating stops. Then you pay for the old work and the new workflow. The recalculating does not disappear through encouragement, it disappears through an agreement with a date: check three runs completely, then spot checks only, and back to complete checks at every deviation.

The second expensive case is the exception that is written down nowhere. Reports are full of them, and they are the reason a workflow works in the test and no longer works in the third month. Honestly, there is never a way to say in advance how many there are. We described this in more detail using the example of exceptions in automated workflows: the knowledge about the special case sits in people's heads, not in the work instruction, and it only comes out if you ask the person who produces the report today.

Both can be priced in beforehand if you know what the report costs today. There are five items, namely working time, handovers, external costs, errors and rework, waiting time. How to add them up for one workflow is covered in a separate article. You need that number anyway, because it is the only one you can no longer collect after the rebuild.

Pause it for two weeks instead of deleting it

For group one there is a test that costs nothing. Pause the report instead of deleting it, and do not announce the pause. Write down the date and wait two cycles.

If nobody gets in touch, you have your answer, and you have evidence for it instead of an assumption. If someone does ask, you get something better than a report that keeps running, namely a person who tells you what they need it for. That puts it into group two or three, and you already know what it has to look like afterwards.

One note from practice: a report that falls away takes a habit with it, and some people only notice weeks later that something is missing on Monday. That is why the result of the pause belongs on the agenda and not only in your notebook. Whoever missed the report should be able to say so before it disappears for good.

The sorting itself is not a consulting service, all you need for it is your own list. If two or three reports are left at the end where you are torn between a query and a rebuild, you can talk them through in a free first conversation. Nothing beyond the list is needed for that.

Three columns next to the list of reports

Write down every recurring report, weekly, monthly, quarterly, including the ones from Outlook and Excel that are in no system at all. Next to them three columns: which decision depends on it, where the numbers come from, and who would ask about it.

Then sort. Cut, turn into a query, have it prepared, leave it alone. Start with the cuts, not with the technically most interesting case. This list is written in an hour, and it decides more about the outcome of the first rebuild than the choice of tool.

Frequently asked questions

Which reports can AI automation take over?

Reports that repeat on a fixed rhythm, whose numbers sit digitally in two or three systems, and on which a decision depends. If part of the work is a judgement, that stays with people and only the assembly is prepared. Reports without a decision are not rebuilt, they are cut.

How do I spot a report that can be dropped?

By two answers: nobody has done anything differently after reading it in the past six months, and nobody asks when it fails to arrive. Do not test this in your head. Pause the report for two cycles without announcing the pause, and write down the date.

What happens to reports that fulfil an obligation?

They stay. If a contract, the auditors or an authority require the report, the obligation is the purpose and the question about the decision is misleading. Form and deadline are fixed. Simplifying how the inputs come together is often still possible.