The Information Theory of Business - Part 2

In the last article I argued that AI’s value shows up when workflows disappear rather than when they speed up. What I didn’t explain is why so many workflows turn out to be removable in the first place. That comes down to what a business fundamentally is.

Most management theory treats a business as a collection of processes. We optimise them, automate them, document them, measure their efficiency. I think the abstraction is wrong. A business runs on information.

Every economically meaningful event starts as information. A customer places an order. A supplier changes a price. A project slips. A sensor detects a fault. A customer threatens to leave. On its own, none of that is worth anything. Value appears only when the information triggers the right decision and someone acts on it.

Strip away the industry and every business runs the same loop to get from one to the other. Gather the data. Turn it into information. Turn that into intelligence, the pattern and the implication. Move it to whoever can act. Apply judgement. Execute, and feed what happens back into the next round.

Six operations, and this article is about one of them. Movement adds nothing to information. It does not change what the information is. What it decides is whether value created upstream reaches a decision while the decision still matters.

The same shape twice

Take the expense claim I pulled apart last time. Photograph the receipt, fill the form, pick a category, write a justification nobody reads, get approved, survive the sample check, wait six weeks for the money. Seven steps, and every one of them exists because the company found out about the spending long after it happened. Put the policy on the card and the transaction arrives already coded, already checked, already evidenced. The process did not get faster. It stopped existing.

Those seven steps existed only because the money moved before the company knew about it. Twenty years of software made each one faster and removed none of them.

Ford in 1990 is the same move with a database instead of a card. Five hundred people in accounts payable, employed to reconcile three documents describing the same lorry, because the three documents arrived in three places at three different times. Ford stopped having invoices and checked the delivery against a shared record at the dock. Headcount fell by seventy-five per cent. The work went away because the delay went away.

Both cases have the same shape, and both are failures of movement. Information about a single event reached the people who needed it later than the decision it should have informed. Nobody could act at the moment of the event, so the business built machinery to recover the information afterwards. A claim. An approval. A sample check. A reconciliation. Then it hired people to operate the machinery.

Which gives me a first principle:

Information is the raw material of enterprise value. Hands-on time is what you pay to convert it. Elapsed time sets the cost. The longer information waits for the decision it should inform, the more hands-on time the conversion needs.

What delay costs

That hands-on time comes out of a fixed pool. Every employee has roughly forty hours in a week, and once a role is saturated, additional demand converts into queue rather than output. The request just waits behind everything already in front of it. Asking is free. Capacity is not.

Delay is therefore charged twice. Once as the wait itself, and again as the wages of everyone employed to compensate for the wait. At Ford those wages covered five hundred people.

So the question for any business that has run out of hours (and doesn’t want to buy more) is this. How do we get the same outcome, or a better one, from the same information while spending less hands-on time on it?

The usual answer is to make people faster, and that is exactly how AI gets sold. That answer assumes the work deserves to survive.

Eliminate, delegate, or keep

Take any piece of work that consumes a person’s hours and ask two questions in order. Does this need to exist? If it doesn’t, it goes. If it does, can something other than a person do it? If it can, it gets delegated. If it can’t, it stays with a human.

Eliminate, delegate, or keep. There is no fourth branch, because there is nowhere else for work to go.

There are two answers people reach for before they ask either question, and both fail for the same reason. They add throughput to a problem that is made of delay.

Hire more people and you have bought the ability to do more things at once. Brooks made the point in 1975: the bearing of a child takes nine months no matter how many women are assigned to it. Extra hands parallelise work that can be divided, and do nothing whatever to a sequence that has to happen in order. Nobody can approve the claim before it has been submitted. Ford’s clerks could not reconcile three documents until the third one arrived. Hiring clears a queue. It cannot shorten a chain, and the chain is what produced the delay in the first place.

Better tools fail the same way, one level down. A faster form is still a form. You have taken minutes out of a step that shouldn’t be running and left the six weeks standing.

Elimination happens when the reason for a step goes away. While the company only found out about spending six weeks after the spend, the claim form had to exist. Put the policy on the card and the same step becomes unnecessary. Nothing about the step changed, only the delay it was built to cover. Lateness is the target, not labour.

This is why I prefer workflow compression to workflow automation. Automation swaps a person for software and leaves the process standing. Compression asks why the human touchpoint is there at all, and the answer is usually historical rather than economic. Someone built that step years ago for reasons nobody has revisited since.

Neither of my two examples used AI. Ford used a shared database and a receiving clerk. The credit card runs a rules engine at the point of sale. Both removed a delay with ordinary deterministic software, and deterministic software is what you should reach for whenever the rules can be written down. It is cheaper, it is testable, it can be audited, and it behaves the same way on Tuesday as it did on Monday. AI earns its place where the variety is too high to enumerate in rules, which is a real category and a large one. It is not the same category as everything currently done by hand.

What stays with people

Everything you eliminate or delegate is in service of what stays.

Go back to those wages. The hours Ford’s five hundred clerks spent reconciling paperwork did not come from a separate budget. They came out of the same forty hours a week that strategy, customer negotiation, creative problem solving and leadership are drawn from.

Which makes the wages the smaller half of what delay costs. The larger half is that it spends the hours in which people would otherwise have been exercising judgement, and judgement is the only thing we actually hire them for. Nobody is employed to move data between systems. They are employed to make calls where the answer isn’t obvious, and then handed a queue of work that requires none.

There is also a floor here that no language model release will lower. Judgement is what somebody is held accountable for. A system cannot be held accountable, so judgement cannot be handed to one.

Some steps also stay with a person even when a machine could do them, and I laboured this point last time because it is the one easily lost in the enthusiasm. A payment release needs a second authoriser, and not because one person is incapable of pressing the button. One person pressing the button is precisely the risk being managed. A pharmacist checks the prescription because the cost of being wrong lands on the patient. On a process map those steps look identical to the compensating kind, and the test is what the step is for. If the answer describes a limitation, it can probably go. If it describes a consequence somebody deliberately chose to guard against, it stays, whatever a machine is capable of.

So the design principle is short. Eliminate the work that exists only because information arrived late. Delegate what remains to the simplest thing that can do it reliably. Keep the judgement, and keep the controls somebody chose on purpose. The order matters, because the AI roadmaps I get shown start by asking what a machine could do, and that question quietly grants every existing step the right to survive. Ask whether the work needs to exist first, and much of what you were about to delegate disappears before you get there.

That is a bigger shift than another wave of automation, because it changes the economics of the firm. Find the places where information arrives later than the decision it should have informed, and count the people employed in the gap. Every one of them is paid to move information that should have arrived on its own. That headcount is what the redesign is worth, and the number of AI tools you have bought tells you nothing about it.