The method (what the companies where AI shows up in the result did differently)

Eight hundred quotes

Seven forty on a Tuesday in August. An electrical supplies distributor in Rosario: one hundred and forty people, three warehouses, forty years in the trade. The head of administration turns her screen so the owner can see a finished quote.

"This used to take us two hours. Now it takes twenty minutes."

It is true and it checks out. For a year now the sales team has been putting quotes together with artificial intelligence: it loads the line items, pulls the prices from the list, drafts the terms note. It works. Nobody is exaggerating and nobody is lying.

"And how many quotes went out last month?" he asks.

"Eight hundred and something. Same as always."

"And a year ago, how many went out?"

A short silence.

"Eight hundred and something too, I suppose."

They are both right. The tool works, the task takes a sixth of the time it used to, and a year later there is not one line of that company's P&L anyone can attribute to it. There is no way to argue about it either: nobody wrote down how things were before.

(The scene is a composite reconstruction, and so are the figures inside it. From here on, everything carrying a number is measured and has a source.)
The sources were read in the version available in August 2026.

The myth in this last part is the most expensive of the four the series has covered, because it is the only one that arrives with an invoice: adopting artificial intelligence means buying the tool and giving people access to it. The licences, the training, the vendor who configures it: that is the expensive part and it is the part that gets approved, because it is the part with a price and a date.

And it is the one part of all this that, on its own, changes nothing in the result.

Part 3 of this series answered one half of the question: the time saved at one step is spent, two desks further down, by whoever has to verify what came out. This is the other half, and it is cheaper to fix. Even when the work is not transferred to anyone, the time saved shows up nowhere if the flow around the task stayed the same: the quote still waits for the same signature, still goes out in the same Friday batch, and the bottleneck —which was never the writing— is still exactly where it was. Speeding up a step that was not the one holding things back speeds up nothing.

Which leaves the question: what did the companies where artificial intelligence does show up in the result do differently? There is an answer, and it comes with a label that has to be read first.

What they did differently

In March 2025, McKinsey published the round of its global AI survey that put twenty-five different things a company can do or fail to do when it brings AI inside to the test: who governs it, how much is invested, how people are trained, which data gets tidied up first. Of those twenty-five, the one most associated with AI showing up in operating profit is redesigning workflows.

There is one expression worth unpacking: the report talks about EBIT attributable to generative AI, which means how much of its operating profit the company attributes to the tool. The word doing all the work is attributes: it is declared by the same person answering the survey, and nobody audits it.

And redesigning workflows is also the thing fewest companies do. Among those reporting gen AI use by their organisations, 21% say they have fundamentally redesigned at least some workflows. One in five.

The next round of that survey, the one from November 2025, explains that Rosario morning better than any diagnosis. Almost nine in ten say they use artificial intelligence regularly in at least one function. Fewer than four in ten report any effect on operating profit, and most of those put it below 5% of their total operating profit. The tool reached everywhere. The result, almost nowhere.

Of everything a company can do, which one came first and how many did it:

What “redesigning the flow” actually means

Put like that it means nothing: it is one of those phrases you have heard a hundred times without it ever telling you what to do on Monday morning. It does, however, have a short definition that can be verified in an afternoon.

A redesigned flow is one where who does what, in what order, what stopped being done and who signs at the end have all changed. All four changes, not one.

That is where the test comes from, and it is the only thing worth remembering from this section:

If nobody stopped doing something, there was no redesign: there was one more tool.

That is the test for two reasons. The first is that it cannot be faked: a company can say it redesigned the work and show a new diagram, but it cannot point to the step that disappeared if the step is still there. The second is that a step that disappears is the only thing that genuinely frees up capacity. As long as the quote still waits for the same signature before it goes out, it makes no difference how long it took to write: the flow delivers exactly what it delivered before, with less busy people inside it.

And now the warning that avoids the disaster, because removing a control step without having tested the task is worse than doing nothing. The order of what follows is the order in which it is done: first you find out which side the task falls on —which side of the limit of what the tool does well—, then you measure how things were going, and only then do you take the step out. The redesign arrives at that point, and not before.

Three common tasks at a mid-sized company, with and without redesign:

In all three rows the proof is the same: somebody stopped doing a step and there is a new signature. If you cannot point to both, you bought a tool. 32sur schematic: none of the three rows is a measured case —they are three flows common at a mid-sized company, written to show what changes and what does not— and the sampling percentages are illustrative: each company sets its own from the result of its own test (see Part 1 of this series).

Before redesigning, you need to know which task holds up

The first part of this series showed that the limit of what the tool does well cannot be deduced: two tasks that look equally demanding to a professional can fall on opposite sides, and the one that fails does not come marked. The authors gave that shape a name: the jagged frontier. The consequence for this method is a single one: the unit of work is the task, not the department and not the company. Anyone who redesigns the flow of an entire department because the tool worked well on one task is betting that the tasks next to it are the same, and that is the error Part 1 measured.

Locating a task costs an afternoon and nothing more. You test it on already-solved cases, with the right answer written down before anything is opened, and you count two things: how often it got it right and how often it got it wrong while sounding sure. The second one decides, because it is the one your people will not catch.

In Competing in turbulence we argued that a bet on new ground is tested small and with the exit conditions written before starting. It is the same here, with one difference that matters: the criterion is not written to find out whether the bet paid off, it is written so you can see the error that sounds sure of itself.

The number your company does not have

Suppose the task held up under the test. What is still missing is the number that lets you say, a year later, whether it was worth it: how things were going, written down before anything was touched. It is called a baseline.

And it is not written in hours saved, but in the output of the flow: how many quotes went out per month, in how many days from the request, how many had to be corrected afterwards, how many were won. Hours are no use, and Part 2 of this series showed why: they are precisely what people estimate badly, even about their own work.

If you rolled the tool out eight months ago and never wrote any of that down, the good news is concrete: the baseline still exists, dated, inside your own systems. The previous twelve months are in the management system, in the mailbox, in the quotes folder. Rebuilding it costs an afternoon.

There is a second thing that also costs no money. The best field study there is on this —the one on the support agents at a software company, the same one from Part 1— was able to measure what it measured for an operational reason: the tool was opened site by site and one agent at a time, and that difference in calendar made it possible to compare each person with themselves before and after. There is a method being given away there: one slice of the company that starts six weeks later. That is a comparison group, and no consultancy sells you those six weeks.

What is still missing is the underlying reason to do it in-house. There is a meta-analysis that puts the field in order: it pools the results of another twenty-three studies and finds a real, modest effect on productivity, 0.33 on the standardised scale used to compare studies that measured different things —it is not a percentage: there, 0.2 counts as a small effect and 0.5 as a medium one—. Two caveats before using it: the twenty-three studies are about programming and not about quotes, and the meta-analysis has not been peer reviewed yet. But the number that matters is a different one: the range where that average may sit, from 0.09 to 0.58 —from almost nothing to quite a lot—. The meta-analysis itself explains the width: the more the measurement resembles a real company, the smaller the number.

The average of other companies is not a forecast for yours. It is barely proof that your number exists and that somebody has to go and find it.

Where that effect may sit, and where yours sits:

Adoption is not requested: it is designed

This is still the same station of the method: the rule is the other half of the redesign, and without it the redesign cannot even be written down. Everything above assumes something no company has guaranteed: that its people use the tool and say when they used it.

Declaring that you used artificial intelligence is not free for whoever declares it, and Part 3 of this series shows the experiment where you can see it: the same piece of work got a lower competence rating when its author said they had used the tool. Whoever paid that cost once does not say so the next time. Few people connect the consequence: use that is not declared leaves no trace, and a flow nobody knows the workings of cannot be redesigned.

And asking for it in a meeting is not enough. A Danish study —a draft, still without peer review, that crossed two survey rounds covering some 25,000 workers with administrative data on earnings and hours— found something that puts any adoption plan in order: encouraging use, rolling out the corporate version of the tool and running training are associated with better outcomes when they go together; done separately, the corporate version and the training are associated with lower reported benefits. Half an initiative is associated with less reported benefit, not more.

The study does not say why; it offers more than one explanation and does not choose between them, and this article is not going to invent the missing one. What does fit with everything above is that training and rolling out without touching the flow adds steps —learning the tool, reviewing what comes out— and removes none.

Not even the companies that push manage all three: among those that encourage use, only 29 out of 100 rolled out the corporate version and ran training. The other end is the finding that convinces: where all three coexist, 93 out of 100 workers say they have used the tool, and more than one in four use it every day. No amount of exhortation produces that level of use.

What replaces exhortation is a one-page usage rule. It is not an information security policy and it is not a code of ethics: it is four lines, none longer than a single line.

One more rule is missing, and it comes from Part 1: asking the tool to review its own work scales the persuasion, not the truth. Verification happens against the criterion written beforehand, and never inside the conversation with the tool. A review that starts after you have read the answer is not a verification: it is an argument, and on the other side there is something better equipped to win it.

The full loop

Put in order, the five things this article talks about —locating the task, testing it small, writing the baseline, redesigning and setting the rule, scaling or stopping— are a loop: you come in through a task and come out with a written decision, which is how the next one comes in. A task is not adopted until it has gone round the full loop, and almost every company does the first quarter and calls that adoption.

The loop ends in scaling or in stopping, and both are legitimate outcomes. Stopping a task after testing it is the cheap way of finding out it was on the wrong side of the limit: it costs an afternoon and not an entire rollout. What almost no company has is the record of the tasks it tested and discarded, and on what criterion it discarded them; that is the asset that makes the second loop cost half.

And then the scale, which is where you decide whether this is executable at all. A full loop over one task is weeks and not quarters: an afternoon for the criterion and the test, an afternoon for the baseline, six weeks of rollout in batches, one page of rule. The expensive part was never the time. The expensive part is having four pilots open at once and being unable to decide on any of them.

The five stations, with the question each one answers and the paper it produces:

The four questions of the system

The four parts of this series were not four subjects. They were four questions put to the same task, in order, and failing just one is enough for the result never to appear.

And they are put to a task. None of the four questions of the series can be answered about “artificial intelligence in the company”: all four are answered about “the task of putting quotes together”, which is the unit this series has worked with since its first line. That, incidentally, is why an AI committee never answers them: a committee deliberates at company level, and all four only have an answer at task level.

The order is not decorative either. If the task falls on the wrong side of the limit, measuring it well is of no use. If the number that excites everyone was measured by somebody else, at another company, on other work, your own baseline never gets written. If nobody declares when they use the tool, there is nothing to review. And if the flow stayed the same, the three previous answers stay on the desk of whoever answered them.

All four, with what it takes to answer each one:

The method is the barrier, and the method can be learned

The two numbers that opened Parts 1 and 2 come from the same survey and have never been read together. The first: 41.6% of the Argentine SMEs surveyed by CEPE —the public policy centre at Universidad Torcuato Di Tella— and Fundar already use at least one artificial intelligence technology. The second: the barrier they report most, above cost, is the lack of experience and knowledge about the subject, 46.2%.

Read them one after the other and the conclusion assembles itself. What those companies say they are missing is on no price list. It is the method, and the method is the only thing in everything this series covered that can be learned.

The other half is a matter of calendar. A full loop over a single task is weeks, and nobody finishes it without having started it. Whoever sits down this month to reconstruct how many went out per month before the rollout is in time. Whoever leaves it for next year will have the same Rosario conversation, one year more expensive.

The tool reached everywhere; the design of the work, almost nowhere:

The three figures come from three different kinds of source and describe the same thing from three places: the tool came in, and the work stayed where it was.

Questions for Monday

The four questions on the card are put to a task; these seven you put to your company. The first three can be answered right now, without opening anything and without asking anyone.

Go back to the office in Rosario, seven forty. Nothing that company did was wrong: it picked a reasonable task, bought a tool that works, and its people use it every day without anyone having to insist. What is missing fits in two lines. Nobody wrote down, a year ago, how many quotes went out per month and in how many days. And nobody took a step out: the quote gets written in twenty minutes and still waits for the same signature until Thursday.

With those two lines written down, this morning's conversation would have lasted forty seconds and would have ended in a decision. Without them it can last another year.

Four parts, four questions, one task. The series was not about artificial intelligence: it was about how decisions get made when the number that reaches the table was measured by somebody else, at another company, on other work. The tool will keep changing every six months, and the four questions will keep being the same.