The Reports Nobody Reads Are About to Multiply

Somewhere in your monthly reporting pack is a page that nobody reads. Probably several. It was requested in a specific meeting, by a specific person, for a specific reason that made sense at the time. That person may have changed roles or left the company. The reason may have expired years ago. The page is still produced every month, reviewed by no one, and filed.

Nobody has removed it, because removing it would require someone to establish that it is not needed, and nobody is quite sure. So it continues. And what makes this worth thinking about now, rather than as a permanent minor annoyance, is that the one force that historically limited how many such pages accumulated has just been removed.

Reports do not die, for a rational reason

The persistence of unread reports is not evidence of a careless organization. It is the predictable result of how the costs and risks are distributed.

Consider the position of anyone who might retire a report. If they stop producing it and someone did need it, that failure is immediate, visible, and personally attributable. There will be a moment in a meeting where a number is missing and a name is attached to its absence. If instead they continue producing something nobody reads, the cost is a few hours a month, absorbed into a general sense that reporting takes a long time, and attributable to no one.

Faced with a visible personal risk on one side and an invisible distributed cost on the other, continuing to produce the report is the correct individual choice, every time, for everyone in the chain. Nobody has to be lazy or thoughtless for the inventory to grow permanently in one direction. The incentive structure guarantees it, and no organization I am aware of has a standing process that asks, on any schedule, which recurring outputs should stop.

The result is accumulation. Reports are added when needed and never subtracted when not, so the total only rises, and the pack that took a day to produce in 2019 takes a week now.

What the accumulation actually costs

The obvious cost is production time, and in a finance or operations function that is not trivial. But it is the smaller cost, and treating it as the main one leads organizations to the wrong solution, which is to make production faster.

The larger cost is attention, and Herbert Simon identified the mechanism in 1971, long before anyone was worried about dashboards. In an essay on designing organizations for an information-rich world, he pointed out that information consumes something, and what it consumes is the attention of its recipients. Hence, in his formulation, a wealth of information creates a poverty of attention, and a need to allocate that attention among far more sources than can possibly receive it.

Applied to a reporting pack, this is precise rather than poetic. The attention available to read the pack is fixed. It is a small number of senior people with a limited number of hours. As the pack grows, the share of that attention available for any individual page falls, which means the number that genuinely matters this month is competing with forty pages that do not. Adding unread reports does not have a neutral effect on the read ones. It degrades them, by diluting the finite resource they depend on.

There is a second cost that is subtler and worse. A report on a topic creates a strong impression that the topic is being managed. "We report on that monthly" is often offered, sincerely, as an answer to whether something is under control. It is not the same claim at all, and the gap between producing a number and acting on one can persist for years precisely because the existence of the report suppresses the question. This is a close cousin of the problem of measuring things that do not connect to any decision, covered in the metrics that lie about AI success.

Why this is about to get much worse

Until recently, one thing constrained the growth of report inventories: production cost. Producing a new recurring report meant someone doing real work, repeatedly, forever. That cost was visible enough that requests were at least lightly rationed, and occasionally someone asked whether a report was worth the effort it consumed.

That brake is now gone. When generating a report takes a prompt rather than two days of analyst time, nobody has to weigh whether it is worth producing. The marginal cost approaches zero, the requests stop being rationed, and the inventory grows at whatever rate people can think of things to ask for.

What has not changed, and will not change, is the amount of attention available to read the output. Human attention remains exactly as scarce as it was, which means the ratio of produced information to available attention is deteriorating quickly, and the deterioration is being described internally as an improvement in reporting capability.

Simon anticipated even this. In the same essay he set out a design principle for information systems that is worth applying directly to AI reporting: a new information-processing unit reduces the net demand on an organization's attention only if it absorbs more information than it produces. It has to condense. In his phrase, it must listen and think more than it speaks. He noted that it is conventional to design such a system by considering the information it will supply, and that in an information-rich world this is doing things backwards.

Most AI reporting deployments are doing exactly the thing he warned against. They are designed around what they can generate. By Simon's criterion, a tool that produces more reports than it absorbs is not an attention conserver, whatever it does for productivity. It is a net consumer of the organization's scarcest resource, and the scarcity is what determines whether any of it gets read.

This is a different problem from the collapse of effort as a signal, which is about what you can infer from a document, examined in when effort stops being evidence. The issue here is volume against a fixed ceiling of attention. Both stem from production becoming cheap, and they compound: you receive more documents and can infer less about which ones deserve your time.

The honest part: this is not a case against reporting

It would be a mistake to read this as an argument for producing less information, and the aggressive version of report-culling causes real damage.

Some reports have readers you do not know about. An auditor who uses one schedule annually. A lender whose covenant package depends on a specific breakdown. One owner who reads their variance page with genuine attention and would notice its absence immediately. Regulatory and compliance reporting is not optional and its readership is often external and invisible from the inside.

There is also real risk in removal, and it is asymmetric in the other direction. You frequently do not discover who depended on something until after it has gone, sometimes at an inconvenient moment such as a close or an audit. The people who defend unread reports are not always being irrational.

And cheap production genuinely enables analysis that was not viable before. There are questions worth answering that nobody asked when answering them cost a week of analyst time, and getting those answers is a real gain. The problem is not that AI makes reporting cheap. It is that cheapness removes a rationing mechanism without supplying a replacement, and the replacement has to be deliberate.

The discipline that replaces the cost brake

If production cost is no longer limiting the inventory, something else has to, and it should be a small number of habits rather than a governance program.

Every recurring report needs a named reader. Not a distribution list, a person who would notice and complain if it stopped. If no such person can be named, the report is a candidate for retirement, and the exercise of trying to name one is usually more informative than any review.

New reports get a sunset date. Set at creation, typically six or twelve months, after which the report expires unless someone renews it. This inverts the default from persist to expire, which addresses the underlying incentive problem without requiring anyone to make the uncomfortable case for killing something.

Use the stop-and-see test. Rather than trying to establish in advance whether a report is needed, stop producing it, keep the ability to reproduce it on request, and see whether anyone asks. This converts an unanswerable question into a cheap experiment with a safety net, and it is far more reliable than asking people whether they use something, which reliably overstates.

Distinguish reports that drive decisions from reports that create records. Both are legitimate, and they have different requirements. A decision report needs to be read, so it competes for attention and should be short. A record exists to be available if needed, so it does not need to be in the pack at all, and moving it out is not a deletion.

And apply Simon's test to anything new. Before deploying AI to generate reporting, ask whether it will condense more than it produces. A tool that summarizes six existing reports into one is an attention conserver. A tool that generates thirty new dashboards nobody asked for is not, however impressive its output, because it is spending the resource that determines whether reporting works at all.

The organizations that handle this well over the next few years will not be the ones that generate the most analysis. They will be the ones whose leadership can still find the number that matters, because they resisted the temptation to produce everything they were suddenly capable of producing.

FAQs

Q1. Why do unread reports persist for so long?
Because the incentives make continuing rational for everyone involved. Retiring a report someone actually needed produces an immediate, visible, personally attributable failure. Continuing to produce an unread one costs a few hours absorbed into general overhead and is attributable to no one. Given that asymmetry, the individually correct choice is always to keep producing, so inventories only grow.

Q2. Isn't the main cost just the time spent producing them?
That is the smaller cost, and focusing on it leads to the wrong fix, which is faster production. The larger cost is attention. The capacity to read the pack is fixed, so every added page reduces the share available for the pages that matter. Unread reports do not sit neutrally alongside read ones, they compete with them for a finite resource.

Q3. What did Herbert Simon actually say about this?
In a 1971 essay on designing organizations for an information-rich world, he observed that information consumes the attention of its recipients, so a wealth of information creates a poverty of attention. He also set out a design principle: a new information-processing unit only reduces net attention demand if it absorbs more information than it produces, meaning it must condense rather than add.

Q4. Why does AI make this worse rather than better?
Because production cost was the only real constraint on report proliferation, and AI removes it. Requests stop being rationed when generating a report takes a prompt instead of days of analyst work. Meanwhile the attention available to read output is unchanged, so the ratio of produced information to available attention deteriorates while being described internally as improved capability.

Q5. How do I know whether a report has readers?
Try to name a specific person who would notice and complain if it stopped, rather than pointing to a distribution list. If no name comes to mind, that is informative. The more reliable method is the stop-and-see test: cease production while retaining the ability to reproduce on request, and observe whether anyone asks. Surveys about report usage consistently overstate it.

Q6. What is the risk of retiring reports too aggressively?
Real, and asymmetric in the opposite direction. Some reports have invisible external readers such as auditors, lenders, or a single attentive owner, and compliance obligations are not negotiable. You often discover who depended on something only after it is gone, sometimes during a close or an audit. This is why the stop-and-see approach, which preserves reproducibility, is safer than deletion.

Q7. Should we just stop using AI for reporting?
No. Cheap production genuinely enables analysis that was not viable when answers cost a week of work, and that is a real gain. The point is that removing a cost constraint without installing a replacement produces uncontrolled growth. Apply Simon's test: a tool that condenses six reports into one conserves attention, while one generating thirty new dashboards consumes it.

Q8. What is the single most effective habit to adopt?
Sunset dates on new recurring reports, set at creation. This flips the default from persisting indefinitely to expiring unless renewed, which addresses the incentive problem directly. Nobody has to make an uncomfortable case for killing anything, because the burden shifts to justifying continuation, which is the position that should have required justification all along.