Skip to content
Articles
ArticlesAI translation· 26 October 2026

The Blueprint of Disruption: Why Software Shows What's Coming for Knowledge Work

by 10 min read

Technological revolutions don't transform an economy everywhere at once. They start where the technology works, where deployment pays off and where organisations can adapt the way they work. From there, they spread.

Generative AI follows this pattern. Right now the shift is most visible in software development: the people building this technology are also using it to reorganise their own work.

But the significance reaches far beyond code. Lawyers, auditors, consultants and marketers solve different problems, yet many of their tasks share the same structure: understand the brief, gather information, produce a result, check it, revise it.

My thesis: what we're seeing in software development will carry over to other knowledge domains in comparable form. Not identically and not simultaneously, but along a similar pattern: from assistance to delegation to redesigning the organisation. This blog has already measured these strands one by one: centaurisation, the GenZ dilemma, the ROI gap. This article connects them.

The IMF estimates that around 60 percent of jobs in advanced economies are exposed to AI. Exposed doesn't mean replaceable: some tasks become more productive, others come under pressure. Still, the order of magnitude makes one thing clear: this is not a story about the tech industry alone.

The waves don't follow a ranking of professions. Centaurisation already supplied the criterion: what matters is how well a task can be described, fed with context and checked. Where feedback is fast and errors are manageable, delegation becomes economical first. That's why the waves don't travel from industry to industry. They travel through the work processes of every industry.

Wave 1: Software development, the first big demonstration

“Coding Is AI's Killer App” put the economics of software centre stage. This piece takes up the follow-up question: what carries over to other knowledge work?

Software offers favourable conditions: the process is digital, and compilers and automated tests provide feedback. That feedback loop isn't perfect. Software can pass its tests and still be insecure or miss what the customer actually needs. But it enables a powerful cycle: generate, run, check, improve.

As early as October 2024, Google CEO Sundar Pichai said AI was generating more than a quarter of new code at Google, reviewed and accepted by engineers (Google). That's not labour savings on the same scale, but AI had reached a core production process.

Anthropic's study of its own workforce, published in October 2026, goes further: it links the introduction of coding agents to rising technical output and examines quality indicators such as later code deletions. The observed correlation alone doesn't prove a causal effect.

The point isn't that developers type faster. The point is that the unit of delegation is changing. Autocomplete suggests the next line; an agent takes on a bounded problem, works through several steps and hands back a proposed solution for review. Human work shifts accordingly: specifying, selecting, coordinating, evaluating. Expertise doesn't become redundant. Its economic centre of gravity moves.

That's the blueprint: support in execution increasingly turns into delegation of execution. Why delegation can still pay off economically was described in “AI vs Junior” as Correction Effort: what counts isn't a flawless first draft but how quickly and smoothly the next version comes together. AI doesn't have to be error-free to become attractive. It has to offer the better cost-benefit balance in the concrete process, review and rework included. It's this cost logic, not faith in technology, that explains the pull of the waves.

Wave 2: Document-heavy work, from drafting to verifying

The next wave hits work built on documents, rules and precedents: legal advice, accounting, financial analysis, compliance.

A contract review isn't a single indivisible act. Someone has to understand the business context, identify the relevant clauses, compare them against requirements, flag deviations and assess the consequences. AI doesn't have to replace the whole process to change its economics. If first drafts and comparisons get cheaper, the effort shifts to exceptions, negotiation and sign-off.

This is where the software analogy helps, and also where it hits its limits. A developer runs a test in seconds. Whether a legal interpretation holds up may only become clear in negotiation or in court. An accounting decision may depend on information that isn't even in the document being analysed.

So the decisive question isn't only whether AI can produce a plausible answer. It's whether we can check that answer economically and assign responsibility unambiguously.

Anthropic's study also describes rising throughput in legal and administrative processes, a sign that the shift won't stay confined to software (Anthropic).

For professional services firms, this means: when preparatory work gets cheaper, staffing structures, training, pricing and review processes all have to be rethought.

The training question is particularly uncomfortable: if AI takes over entry-level work, where do young professionals acquire the review skills they're supposed to use to supervise these systems?

This blog already raised that question in the GenZ dilemma and sketched first answers: audits of AI-generated drafts, Socratic dialogues, simulated “grunt work”. These are approaches to new learning paths, not yet proof that professional experience can be fully front-loaded. What matters is how learners build real subject-matter depth and develop their review skills in a verifiable way. Otherwise it's not just the career ladder that erodes, but the supervisory authority without which no centaur works.

Wave 3: Creative and strategic work and the uneven frontier

One reassuring narrative refuses to die: routine gets automated, creativity and strategy stay protected.

In an experiment with 758 Boston Consulting Group consultants, participants using AI on suitable tasks completed 12.2 percent more tasks, 25.1 percent faster, with markedly higher-rated quality.

But on a task outside the model's capabilities, AI users were less likely to reach the correct solution. The researchers speak of a “jagged technological frontier”: performance doesn't run evenly along familiar categories.

AI delivers persuasive market analyses, campaign ideas and strategic recommendations. But linguistic persuasiveness doesn't prove that the assumptions hold, that constraints were taken into account or that a recommendation can actually be implemented.

The line doesn't run cleanly between routine and creativity, junior and senior. It runs straight through work processes.

For consulting and marketing, this raises the business model question: what is the client paying for when a credible first draft takes less effort? The answer probably lies in understanding context, robust diagnosis, implementation and accountability.

Not every presentation is a strategy. Not every piece of copy creates demand. Not every analysis changes a decision. Anyone who has been selling slide production will have to explain why it actually makes the decision better.

Creativity and strategy aren't disappearing. But drawing the line between them and mere output generation becomes more important.

Wave 4: Rebuilding the company, when local productivity no longer cuts it

The McKinsey comparison on this blog put it in a nutshell: agentic AI isn't a software roll-out, it's organisational design. The fourth wave is exactly that. Not the next industry on the list, but the organisational consequence of the first three.

Suppose engineering builds a prototype faster, legal reviews contracts sooner, marketing produces materials with less effort. The company as a whole still doesn't have to get any faster: budget approval takes weeks, teams disagree about the target audience, nobody has the authority to decide on the next step. That's the difference between task productivity and organisational performance.

Alongside higher output, Anthropic's study describes growing effort spent instructing, monitoring and reviewing agents; meeting time for technical staff without management responsibility went up (Anthropic). When ten agents deliver at once, the old production bottleneck quickly turns into a new review backlog.

So the fourth wave doesn't simply mean: “AI replaces middle management.”

Some coordination tasks are easier to automate, others become more important. The central organisational question is: which decisions need human judgement, which activities can be delegated, and how do we connect the two?

An AI-supported organisation may need fewer handovers, but clearer accountabilities. A Copilot subscription is not yet a new operating model.

The economic test: more business per employee?

The common economic denominator of these waves is leverage: more business without a proportional increase in headcount. Revenue per employee is an interesting indicator for this, but it doesn't explain its own cause.

An AI-native company can achieve high revenue per head because its people use AI productively. Or because it benefits from enormous product demand, scales via compute infrastructure or sources services from outside its own workforce.

The metric describes an outcome, not an isolated productivity effect. Cost per correctly completed process instance is an operational efficiency metric; revenue per employee is a company outcome. Real ROI additionally requires investment, total cost and attributable benefit over a period of time, and that benefit can come from lower costs, higher quality, reduced risk or additional business. Revenue per head isolates neither the one nor the contribution of internal AI use. For consulting firms in particular, this distinction is crucial.

A report produced faster initially frees up capacity. It doesn't automatically create additional revenue. The firm has to win more business, offer a more valuable service or invest the time gained in better client outcomes.

If every competitor can produce the same output more cheaply, the productivity gain may end up with the client: as lower prices rather than higher margins. And cheaper generation doesn't mean cheaper overall delivery. If context building, agent loops and oversight grow along with it, the supposed gain vanishes somewhere else. More efficiency is no guarantee of more revenue per head, and more code is only valuable if it solves a relevant problem.

The competitive advantage doesn't lie in maximum output volume. It lies in translating additional production capacity into value someone is willing to pay for.

Different speeds are not permanent immunity

The four waves aren't a fixed timetable. They overlap, and even within a single profession, tasks can be affected very differently.

Nor can geographical differences be reduced to “America acts, Europe regulates”: data access, integration, liability and customer requirements determine how a specific organisation can use AI.

The danger lies elsewhere: slow visible change can be mistaken for lasting safety. McKinsey's State of AI survey (cited here via the blog post) shows a gap between expectation and reported reality: in 2025, 32 percent expected an AI-driven reduction in headcount; for 2026, 14 percent reported one. Whether tasks and career paths are changing regardless has to be examined separately, by looking at junior hiring, task mix and learning opportunities.

This is where the Social Innovator's Dilemma kicks in: institutions built around established career paths and staffing models have to change precisely the structures they owe their success to.

The task isn't to preserve every job unchanged. It's to safeguard learning capacity, professional development and economic participation under changed production conditions, and to find new answers to how experience is built, how performance is assessed and how responsibility is distributed.

Conclusion: Software is a preview, not a timetable

The lesson from software development isn't that every knowledge worker needs to learn to code. It's that the relationship between expertise and execution is changing. As more execution becomes delegable, the scarce contribution shifts: from producing the first answer to defining the problem, evaluating the result and taking responsibility for its consequences.

That's neither automatically successful nor purely positive. A badly designed rollout can generate more output without added value, accelerate mistakes and weaken the very learning processes that produce tomorrow's experts.

Which brings us full circle: centaurisation describes the working model, the GenZ dilemma the endangered learning paths, the ROI gap the work of translating it all into business results. If entry-level tasks are cut without creating equivalent learning opportunities, the same delegation undermines the review skills its responsible use depends on. That's exactly why the fourth wave isn't a management fad. It's the moment the Social Innovator's Dilemma becomes an organisational task.

2026: The Tipping Point of the Social Innovator's DilemmaDocIsInDaHouse

Software is the first big demonstration. For every other knowledge organisation, the strategic question is: which parts of this pattern apply to us, and what do we have to change to make them work?

Live long and prosper 😉✋

▶Listen