Knowledge Work in a Post-Knowledge Era

For a brief and comfortable period, knowing how to produce a convincing document was a reasonable basis for a career. You gathered information, arranged it into an argument and applied the formatting that indicated somebody qualified had been involved. The finished object travelled upwards through the organisation until it encountered someone authorised to ignore it.

Now a machine can produce the document before you have decided what to call the folder. It can supply the argument, the executive summary and three alternative tones, including the peculiar corporate register in which nobody ever causes a problem but challenges nevertheless arise.

This does not mean knowledge has become obsolete. It means some of the things we accepted as evidence of knowledge have become extraordinarily cheap.

The document was never the work

There are real productivity gains here. In a 2023 experiment published in Science, Shakked Noy and Whitney Zhang studied professional writing tasks and found that access to ChatGPT reduced average completion time by 40%, while assessed output quality rose by 18%.

Those were defined writing tasks, not a measurement of entire occupations. But the result deserves to be taken seriously. Producing a useful first draft faster is useful. There is no moral requirement to spend Thursday struggling with a paragraph that could have been finished on Wednesday.

The difficulty is what happens when organisations mistake easier production for completed thought. A strategy document is an artefact of deciding what to do. Its existence does not establish that the decision is sound.

Imagine a business generating ten versions of a market-entry plan in an afternoon. The plans may be excellent. They may also share an assumption nobody has tested because everyone is occupied comparing the introductions. We have accelerated the supply of answers without necessarily improving the question.

Expertise can travel further

The more interesting gains are not confined to making polished people faster. AI can also give less experienced workers access to approaches that would otherwise take longer to acquire.

In Generative AI at Work, published in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that AI assistance increased customer-support productivity by 15% on average, measured through issues resolved per hour. Benefits were larger for less experienced and lower-skilled workers.

The setting matters: customer support within a particular organisation, with a tool assisting people in their work. It does not prove that expertise is now downloadable in every profession. It does show that some practical knowledge can be distributed more effectively than through the traditional arrangement of hoping the new person sits near somebody helpful.

That should be welcomed. Expertise has sometimes enjoyed an unnecessary scarcity premium. Knowing where the answer is filed is useful, but it is not quite the same achievement as understanding why the answer is right.

The frontier has awkward edges

Unfortunately, competence does not arrive evenly. A system can perform impressively on one task and struggle with another that looks similar to the person assigning it.

Research involving 758 consultants, described by Harvard Business School, found substantial improvements on tasks within the model’s capabilities. The researchers described a “jagged technological frontier”: usefulness varied across tasks rather than advancing along a tidy line.

This makes delegation more demanding than the demonstration suggests. Watching a model perform something difficult does not tell you whether it can reliably perform the next thing. The impressive answer is evidence about that answer, not a certificate covering the rest of the afternoon.

Human colleagues also have uneven abilities, but we develop ways of recognising them. We know who understands the customer, who can investigate a failure and who should never be left alone with a spreadsheet containing dates. With AI, we need comparable discrimination without assuming fluency supplies it.

Feeling faster is not the same as being faster

The evidence also resists a simple progress narrative. In an early-2025 experiment, METR studied sixteen experienced open-source developers working on 246 tasks in familiar repositories. Allowing AI increased completion time by 19%.

That finding concerned particular tools, people and tasks. It was never a verdict on all software development. In its February 2026 follow-up, METR said newer tools likely provided greater benefits, but selection effects and measurement difficulties made its updated estimates unreliable.

The lesson is not that AI is secretly useless. It is that measuring changed work is difficult, and confidence about the experience is insufficient. A tool can feel productive because something is continuously happening. So can a meeting.

For someone responsible for a real system, the relevant result includes checking, integration and whatever must be repaired afterwards. Generated output is a stage in that process, even when it arrives with the brisk assurance of a finished job.

The apprenticeship problem

This creates a question we have barely begun to answer: how do people acquire judgement if the work through which they previously learned it is increasingly delegated?

Junior assignments are not always educational treasures. Some are repetitive chores protected by senior people who confuse having endured something with its continuing necessity. We should not preserve pointless work as a heritage attraction.

But struggling with a draft, tracing a bug or assembling an analysis can reveal why an apparently obvious answer fails. Remove those experiences without replacing their educational function and we may become very efficient at producing work that fewer people can assess.

That is a risk, not an established forecast of universal deskilling. AI could also improve learning through explanation, feedback and practice. The outcome depends partly on whether organisations use it to develop capability or merely to stop paying for its development.

A business that removes every junior role and later complains about a shortage of experienced people has created a rather demanding recruitment specification.

Somebody still owns the decision

The reassuring response is that humans will concentrate on judgement. Perhaps they will. But judgement is not a mysterious substance that appears automatically when typing becomes unnecessary.

It requires relevant knowledge, an understanding of consequences and the opportunity to challenge an attractive answer. The person reviewing the work needs enough authority to reject it, not simply the responsibility of having clicked approve.

Consider an analysis recommending that a service be withdrawn. The spreadsheet may be correct. The summary may be immaculate. Neither tells you whether an affected customer has been promised something different, whether the apparent saving creates another cost, or whether the organisation is willing to defend the consequence.

Some of that context can be recorded and supplied to a model. Some must first be discovered through conversations that the request did not mention. Knowing that something important may be missing is itself part of knowing the job.

What becomes scarce

A post-knowledge era is therefore a misleading name for a real change. Knowledge remains essential. What becomes less scarce is the ability to produce an answer that resembles the work of somebody who possesses it.

That changes how we should assess contribution. The useful questions become harder to satisfy with presentation alone. What did we establish? What remains uncertain? What changed because somebody understood the situation well enough to act?

There will still be room for writing, analysis and technical craft. There may be considerably more room to attempt them. But the visible output will become a weaker guarantee of the thinking behind it.

Somewhere, an organisation is using AI to generate a report that another employee will use AI to summarise. The summary will enter a meeting where nobody has spoken to the person experiencing the problem. The machinery will work beautifully. The knowledge work will still be waiting to begin.