On Artificial Intelligence
ANNOTATED
The one chapter where the printed textbook is newest and therefore most confident, and most wrong. A practitioner's POV on what AI changes in a business, what it quietly doesn't, and where it actually belongs in a process.
The Tool, Not the Tradesman
The current literature presents AI as a transformative force poised to reshape every industry. It is described as a new kind of colleague, an autonomous agent, an intelligence that will absorb whole categories of work. The framing is one of replacement: tasks done by people will now be done by machines, and the organisation that adopts fastest will win.
What is true beneath the framing is more specific and more useful. AI is exceptionally good at tasks with high volume, tolerance for approximate answers, and abundant examples: drafting, summarising, classifying, translating, extracting structure from mess. It is unreliable precisely where business is most demanding: where the answer must be exactly right, where the situation is novel, and where the cost of a confident error is high.
When ATMs arrived, the obvious prediction was the end of the bank teller. The opposite happened for decades: cheaper branches meant banks opened more of them, and tellers — freed from counting cash — moved up into relationship and sales roles. The total number of tellers rose for years after the ATM. The pattern to watch: a tool that automates a task rarely deletes the job; it changes what the job is, usually pushing the human toward the judgement the machine can't do.
In the agency work, AI didn't replace a single role — it changed what the roles do. First drafts, variations, and research got fast and cheap; the humans moved up to briefs, taste, and the client conversation. Exactly the nail-gun pattern: same trade, different hands-on-tool ratio. — A.P.
AI will replace the work. AI changes what the work is made of — and moves the human up to judgement.
AI Inside the System
The common adoption pattern is to acquire the capability first and search for the use afterward — to purchase AI because competitors have, then look for something to point it at. This produces impressive demonstrations that never quite make it into daily operation, because the demonstration was never designed around a real bottleneck.
AI adopted well follows the discipline of any process improvement. Identify where value is lost — the step that is slow, error-prone, or starved of attention. Ask whether the loss is of a kind AI is suited to: high volume, tolerant of approximation, rich in examples. Deploy it there, measure whether the system output improved, and hold it to the same standard of trust as any other component.
The trust requirement is decisive. A business runs on components it can depend on, and a component that is brilliant most of the time but confidently wrong on occasion imposes a hidden tax: everything it touches must now be checked. Where that checking costs more than the work saved, the capability is a net loss regardless of how advanced it is.
Klarna reported that an AI assistant handled the workload of hundreds of support agents, resolving routine queries in minutes. The useful detail is where they pointed it: high-volume, repetitive, tolerant-of-approximation queries — the exact profile AI suits — while complex and sensitive cases still routed to humans. That's the discipline: they didn't ask "where can we use AI," they found the bottleneck of repetitive tickets and aimed the tool precisely at it. Aimed at the sensitive cases instead, the same bot would have been a liability.
My favourite proof that the bottleneck comes first: in one content workflow the constraint wasn't producing work — it was client approvals. Bolting AI onto production would have made a bigger queue in front of the same door. We fixed the approval step instead; only then did faster drafting matter. — A.P.
Deploy AI at a measured bottleneck, hold it to the same trust standard as any component, and count the cost of checking its work.
- AI is a power tool, not a tradesman. It changes what your people do with their hands, not whether you need hands.
- Sort every use by the cost of a confident error. Cheap-and-wrong → automate. Costly-and-wrong → AI drafts, human decides.
- Find the bottleneck first. Capability-first adoption is the feature trap with a shinier tool.
- Count the checking cost. If verifying its output costs more than the work saved, brilliance is still a net loss.
What Doesn't Change
Every general-purpose technology arrives wrapped in the claim that this time the fundamentals are rewritten. Electricity, the spreadsheet, the internet — each was said to make prior wisdom obsolete. Each instead lowered the cost of a category of work so dramatically that the scarce resource moved elsewhere, to the judgement about what to do with the new abundance.
The disciplines in this manual are unaffected by the arrival of AI, and arguably become more important. A process still has to be designed before it can be automated. A metric still lies in exactly the ways Chapter II describes, and does so faster when a machine is generating the numbers. The transitive trap still snaps when the ruler changes, whether the reasoning was done by a person or a model. Efficiency is still respect for the people downstream, even when some of the steps between are now machine-run.
The honest conclusion is neither the breathless one nor the dismissive one. AI is a genuine and large change in the cost of certain work. It is not a change in what makes work worth doing, in the discipline of building systems that hold under pressure, or in the fact that trust, once lost to a confident error, is expensive to win back. The margins of the old textbook still apply. They just matter more.
When VisiCalc and later Excel arrived, they automated overnight what armies of clerks did by hand — and the prediction was mass unemployment for bookkeepers. Instead the spreadsheet created the modern financial analyst, the FP&A function, entire industries of people whose judgement about what to model became the valuable thing once the calculation was free. History rhymes: when the doing got cheap, the deciding got precious. That is the exact shape of the change arriving now.
The MBA finance units drove this home for me: once the spreadsheet made calculation free, the marks — and later the money — went to knowing which model was worth building. I'm watching the same shift with AI from inside two businesses, and I'm betting my career on the judgement side of it. — A.P.
New tools lower the cost of doing. They raise the value of judgement about what is worth doing. Master the margins, and every new tool is leverage rather than threat.
- Cheap production moves value to judgement. When making things is free, knowing what to make is everything.
- The old disciplines matter more, not less. Process, metrics, and the transitive trap all still apply — faster.
- Least worried, most careful. Judgement and trust appreciate; confident nonsense scales. Chapter II is the seatbelt.
- Bet on the scarce thing. The tool got cheaper; judgement got scarcer. Build your career on judgement.