Chapter 9
Reports and Dashboards
The Utopia of the Agentic Enterprise, Part II. The Agents Who Take Your Job
From around 1100, the king’s treasury in England recorded money paid to it on tally sticks, an early kind of receipt. The sum was cut into a stick as notches and the stick split lengthwise, so each side kept a half that matched only the other, which made it hard to forge. Paper and double-entry bookkeeping did the job far better, and the sticks had been redundant for centuries by the early 1780s, when reformers asked why they were still being cut.
Parliament abolished the tallies in 1783, but the Act was to take effect only once the last of the officials who kept them had died, which took until 1826. In October 1834 the Treasury told the Clerk of Works to clear out the old sticks, and two labourers spent a day feeding them into the furnaces under the House of Lords. That evening the furnaces overheated and set the building on fire, and by the next morning most of the Houses of Parliament had burned down. Dickens told the story at a reform meeting in 1855. The sticks could have gone as firewood to the poor who lived round Westminster, he said, but “they never had been useful, and official routine required that they never should be.”
Reports and dashboards turn into tallies the same way. One is started to watch a problem, the problem gets fixed, and the report goes on, because by then it has recipients and a slot in a meeting, and stopping it would mean someone explaining why that’s safe, which is harder than sending it again.
Each recipient gets a small task every time the report goes out: read it or decide not to, and explain the odd numbers when someone asks. Graeber called the people who make up tasks for others taskmasters. A report that has outlived its problem does a taskmaster’s work, and the recipients pay for it in reading. Reports are among the first candidates for AI, because the data is structured, the format is fixed, and whoever prepares them hates it. The reformers of the 1780s wouldn’t have asked if the sticks had cut themselves.
Workflows fossilise too. Something broke years ago, and the workaround became routine, and then policy. New hires were told this was how the company worked, and the original defect disappeared into the procedure.
One kind is the manual handoff inserted because two systems disagreed on every unit of measure and nobody had the authority to fix either one. It made sense when it was created, and the person doing it today has been trained to do it very carefully. Graeber called people in that position duct tapers.
Before AI, this kind of accumulated burden eventually made itself felt. Sooner or later a budget cut made someone ask why five teams were preparing the same number in five formats. The friction worked as an alarm.
AI switches the alarm off. The report writes itself and the approval routes itself. Automating a fossil preserves the workaround at lower cost and leaves the problem it compensated for where it was. Some of this work deserved retirement long before it got a model.
Functions in a large organisation have mandates that overlap. Compliance reviews decisions and legal adds disclaimer language, and both like to be consulted early, which in practice means being copied on everything. Before AI, human bandwidth limited how much each function could produce. People got tired and comment threads died.
AI removes that limit. A proposal to simplify a workflow now reaches the next leadership meeting wrapped in a risk memo and a communications plan, each drafted overnight by someone who had the time because a model wrote the first version. Everyone agrees the proposal is very promising, and the simplification is deferred to next quarter.
So each function’s output goes up, and the organisation’s throughput goes down. A proposal to cut work has produced more of it, which is Graeber’s iron law of liberalism at machine speed.
Deletion Authority
AI can give every team more dashboards and policies than it can use. But who is allowed to delete one?
Someone has to be able to say that a report has become a tally, and to stop a dashboard from becoming a meeting, and the meeting from becoming a department with its own headcount request.
Aircraft designers have a phrase for this: weight begets weight. Add a kilogram, and the structure has to get stronger to carry it, which adds weight, which needs more wing and more fuel, which adds more. Rockets have it worse. A Saturn V weighed nearly 3,000 tonnes on the pad to send about 45 tonnes to the Moon. So when Grumman’s lunar module grew past its weight limit, the engineers went looking for grams. The gold foil the lander is remembered for is one of the things they found: blankets of aluminised film, some layers a hundredth of a millimetre thick and under 20 grams a square metre, in place of rigid heat shields. It took 50 kilograms off the lander, which was worth tonnes on the launch pad. An organisation has no weight limit, so nobody goes looking for grams unless someone has the authority to.
The authority to do that is hard to hold, because old work has constituencies. A report can be useless to the organisation and very useful to its owner, since it gets them a seat in the meeting it feeds. And a control that no longer reduces risk can still protect the function that points to it during an audit.
AI raises the political value of keeping things. When a dead artefact costs nothing to produce, arguing for its removal looks petty. But the report still asks for attention, and for an explanation whenever the numbers look odd. It keeps a small hook in the organisation, and given enough of them, the hooks become the operating model.
Deletion authority also ends where the organisation does. Someone inside can kill a report. Nobody inside a lender or a public body can kill the impact assessment the AI Act asks it to file before a high-risk system goes live, whatever they think of it.1 And the rational response to a rule you can’t remove is to satisfy it as cheaply as possible, which happens to be the work agents do best. So every artefact needs marking as a rule the organisation made or one it hosts. Its own rules it can delete. Hosted ones it can only minimise, and a compliance review that treats the two alike ends up keeping both.
The first candidates are the artefacts AI made cheap to start. The internal app someone built over a weekend with a coding assistant, which other teams now depend on and nobody maintains. Or a finding that would fit in one bar chart, written up as 50 pages for people who read the summary and trust that someone else read the rest.
Working Backward
Before improving a piece of work, find out why it exists. Start with the artefact, the weekly report or the three signatures on a $50 expense, and work backward.
Before Automating It
- What event created this, and when?
- What problem was it solving?
- Which decision depends on it?
- What risk does it control, and what would show us if that risk came back?
Sometimes the answers are humbling. The annoying control turns out to prevent an expensive failure that simply hasn’t happened lately. The spreadsheet is ugly because the ERP never represented the business correctly, and the spreadsheet is how the business runs.
And sometimes they find nothing but inertia: the process exists because nobody got round to removing it.
The aim is to tell those apart. Controls that detect and constrain a live risk deserve to be kept and, if possible, automated well. Controls that mainly produce evidence that a risk was once discussed are candidates for retirement.
Routine work needs the same care. Some of it is waste, like duplicate entry and ceremonial approvals.
A tally did real work once. It proved that a payment had been made, and paper had taken that over centuries before the sticks went into the furnace. A report starts the same way, as somebody’s way of knowing something in time to act on it. If that person is still there, an agent can do the report’s work better than the weekly send did, by telling them when the number moves and saying nothing when it doesn’t.