Product × Finance × AI: Financial Products in the Age of AIPart Two Changing one parameter中文
Chapter 13

A new parameter: the cost of cognitive labour collapsesWhich part of the cost falls

A group of scribes has arrived on the island. They can read every fleet's report, work out every household's accounts and draft every letter; they ask almost nothing in return and never tire. The price of having someone read, calculate and write on your behalf has suddenly fallen to a few hundredths of what it was, or less. But the cost of a delegation is not made up only of this kind of labour. Someone still has to check whether the scribe got it right, and someone has to bear responsibility when things go wrong, while the scribes themselves have neither property to pledge nor a name built up over years. Their abilities are also uneven: some tasks that look hard they do very well, while some that look easy they get wrong. Once the cost of cognitive labour collapses, which part of the cost has actually fallen?

1. Scribes arrive on the island

Up to this point, every institution on the island has depended on one kind of labour. Fleet managers gather news and compare ships; custodians check accounts; those selling ship shares find out what buyers can bear; those designing defaults estimate what most people want. All of this is the labour of reading, calculating, judging and writing. Until now only people could do it, and it was expensive.

Now a group of scribes has arrived on the island, and they do this work for almost nothing. There is one new constraint: the price of cognitive labour collapses, and everything else stays the same.

Today's counterpart. Stanford's AI Index Report 2025 estimates that the price of calling a large model at a given level of performance fell from about US$20 per million tokens in November 2022 to about US$0.07 in October 2024, a drop of more than 280-fold.

The same capability, more than 280 times cheaper

The price per million tokens of calling a large model at a given level of performance. The vertical axis is logarithmic. Drag the budget to see how many reports the same money could read at the two dates.

Data: Stanford University, AI Index Report 2025, benchmarked at an MMLU score of about 64.8%: about US$20 in November 2022 and about US$0.07 in October 2024 (per million tokens). "One report" is taken as about 20,000 tokens, for illustration.

2. What the cost of a delegation is made of

When a principal hires a fleet manager to look after ship shares, what the principal pays for is not just the manager's time spent studying ships. Broken down, the cost has roughly five parts: research and comparison; drafting and judgement; carrying out instructions; checking what the manager did; and bearing responsibility when something goes wrong. The first three are mainly cognitive labour. Part of the checking can be handed to scribes, but whoever signs off at the end still has to look for themselves. Bearing responsibility requires someone with property that can pay compensation and a reputation that can be lost.

The scribes make the first three parts almost free, but the last two do not fall with them. The cheaper cognitive labour becomes, the larger the share of a delegation's cost that goes to checking and bearing responsibility.

Today's counterpart. In investment advisory services, research reports and portfolio recommendations can be generated by software, but suitability judgements made for clients, reconciliation and liability for compensation still rest with licensed firms and staff.

Which parts of a delegation's cost fall

The cost of a delegation broken into five parts. Drag the price of cognitive labour: research, drafting and execution fall almost in step with it, checking falls only in part, and responsibility does not fall at all.

Illustrative: the original cost is 100, with the five parts accounting for 30, 20, 15, 20 and 15; research, drafting and execution are entirely cognitive labour, checking is half cognitive labour, and bearing responsibility is none.

3. Evidence from experiments

How much can the scribes actually save? Several recent experiments give preliminary answers. Noy and Zhang (2023) had 453 college-educated professionals complete writing tasks from their own occupations, randomly allowing half of them to use a conversational large model: average time taken fell by 40% and quality rose by 18%. Brynjolfsson, Li and Raymond (2025) studied 5,172 customer-support agents who gained access to an AI assistant in stages: issues resolved per hour rose by 15% on average, with large differences between individuals. The earlier working paper reported that novices and lower-skilled workers improved by about 34%, while the most experienced saw almost no improvement.

Today's counterpart. These experiments mostly involve short, self-contained tasks, unlike delegation that lasts for years. They show the cost of single pieces of cognitive labour, not the cost of a complete service.

Three experiments: how much was saved

Select a study to see its subjects, method and main results.

Data: Noy and Zhang (2023, Science); Brynjolfsson, Li and Raymond (2025, Quarterly Journal of Economics; the improvement for novices comes from the 2023 working paper); Dell'Acqua et al. (2023, Harvard Business School working paper). Bar lengths follow each study's own percentages; the measures in different studies are not directly comparable.

4. An uneven edge

The scribes' abilities do not vary neatly with how hard a task is. Dell'Acqua et al. (2023) ran an experiment at a consulting firm. On tasks within the large model's capabilities, consultants using it completed 12.2% more tasks, worked 25.1% faster and produced work more than 40% higher in quality. On a task that looked similarly difficult but lay outside its capabilities, consultants using it were about 19 percentage points less likely to give the correct answer. The authors call this irregular boundary of capability the "jagged technological frontier".

An uneven edge means that whether to check cannot be decided by how hard a task looks. The most dangerous tasks are precisely those that look easy but which the scribe gets wrong.

Today's counterpart. A large model may perform well on complex reasoning yet make mistakes on simple counting or on citing sources, and it states its errors with the same confidence as its correct answers.

An uneven edge: what to check

Twenty-four tasks are arranged from left to right by how hard they look. Green ones the scribe gets right; orange ones it gets wrong. Choose a division of labour to see the time spent and the errors that slip through.

Looks easyLooks hard
Scribe gets it rightScribe gets it wrongChecked by a personError slips through

Illustrative: a person takes 1 hour to do a task, the scribe takes almost no time, a person takes 0.3 hours to check one task, and redoing a task after finding an error takes 1 hour. Which tasks go wrong is fixed in advance and only weakly related to how hard they look (what Dell'Acqua et al., 2023, call the "jagged technological frontier").

5. The weakest stave: the slowest step sets the limit

Even if the scribes make one part of the work a hundred times faster, the whole job does not become a hundred times faster. Suppose a delegation used to take ten days: nine for reading, calculating and drafting, one for checking and signing. The scribes cut the nine days to a hundredth of their former length, but checking and signing still take a day, so the whole job goes from ten days to about 1.1 days, a little over nine times faster. Even if the nine days were cut to zero, the whole job could be at most ten times faster. The one day that cannot be sped up sets the ceiling for the whole.

For delegation, the parts that cannot be sped up are exactly checking, bearing responsibility, and waiting itself: grain is still harvested once a year, and ships still sail for months. With the cost of cognitive labour collapsing, the overall amplification factor depends on whether these steps can change as well.

Today's counterpart. An investment recommendation can be generated in seconds, but account opening, suitability assessment, the arrival of funds and trade settlement each still have their own procedures and timelines.

The slowest step sets the limit

Part of a job can be handed to the scribes and sped up; the rest (checking, bearing responsibility, waiting) keeps its pace. Drag the two sliders to see how many times faster the whole job becomes.

Overall speed-up = 1 ÷ [(1 − p) + p ÷ s], where p is the share that can be sped up and s is the speed-up of that share (Amdahl, 1967). The horizontal axis is s, on a logarithmic scale; the dashed line is the ceiling 1 ÷ (1 − p) as s goes to infinity.

6. The reachable share

What proportion of all work can the scribes do? Eloundou, Manning, Mishkin and Rock assessed the tasks of each US occupation one by one. Their 2023 preprint estimated that about 80% of workers have at least 10% of their tasks affected by large models, and about 19% of workers have at least half of their tasks affected. The version formally published in 2024 uses a different measure that distinguishes two cases. With a large model used only through a simple interface, occupations with more than half their tasks affected make up about 1.8%; counting the software and tools built on top of large models, the figure rises to more than 46%. The reachable share therefore depends mainly on what is built around the large model, not only on the model itself. The authors also note that being affected is not the same as being replaced, and that being technically feasible is not the same as being adopted.

Today's counterpart. In finance, research, customer service, compliance review and report writing are all jobs whose tasks are highly exposed to large models.

The reachable share depends on what is built around the model

A hundred cells stand for a hundred occupations. Switch between the two measures to see how many occupations have more than half their tasks affected by large models.

More than half of tasks affectedOthers

Data: Eloundou, Manning, Mishkin and Rock (2024, Science): about 1.8% and more than 46%. Cells are rounded to the nearest percentage, so 1.8% is drawn as 2 cells. Being affected is not the same as being replaced.

7. A village where cognitive labour is cheap: Civilization Simulator, version 13

The simulator adds a new parameter: "the price of cognitive labour". Each year, forty households have four money-related tasks: keeping and reconciling accounts, research and selection, monitoring managers, and adjusting to plan. Each task is partly cognitive labour and partly checking and bearing responsibility. A household does a task only when its benefit exceeds its cost. Drag the price of cognitive labour to see how many households, and how many tasks, become affordable; then hand checking to the scribes as well, and see what happens.

Civilization Simulator · Chapter 13: a village where cognitive labour is cheap

Forty households are arranged from least to most property, and each has four money-related tasks a year. A task is done only when its benefit exceeds its cost. Drag the price of cognitive labour to see which tasks become affordable; then hand checking to the scribes as well, and see what happens.

DelegationRegulation & custodyDefaultsParameter: price of cognitive labourPersonal agent · Ch. 14
Least propertyMost property
AffordableNot affordableScribe erred, nobody noticed

Illustrative model: median property is about 100 portions, the least a dozen or so, the most over a thousand. The four tasks bring benefits of 0.3%, 1%, 0.6% and 0.8% of property and require 2, 10, 4 and 3 hours of labour, of which cognitive labour accounts for 80%, 90%, 50% and 70%, the rest being checking and bearing responsibility; human labour costs 1 portion per hour. The price of cognitive labour is given as a percentage of its original level. When checking is also handed to the scribes, all labour is charged at the price of cognitive labour, but each task has an 8% chance of an error nobody notices, with a loss of twice the task's benefit; each household decides whether to do a task based on its expected gain or loss.

8. What this step leaves open

The price of cognitive labour has collapsed, but checking and bearing responsibility have not fallen with it, and the scribes' capabilities have an uneven edge. The earlier steps therefore have to be derived again:

  1. The boundary of delegation. Once research and comparison are almost free, can households without much wealth also afford a personal agent?
  2. Commitment and guarantees. The scribes have no property to pledge and no accumulated reputation. Who will guarantee their actions, and how can what they did be checked?
  3. Competition over information. When everyone has a scribe reading reports for them, information advantages will be worn away faster. Where will new asymmetries appear?
  4. Participants in the market. When scribes place orders directly, execution is cheaper, but will they all make the same decisions at the same time?
  5. Human bias. Can scribes help people keep to their plans, and will they also be used to exploit people's biases?
  6. What does not change. Time, risk and irreversible losses do not disappear because of the scribes.

What this step leaves open

Six questions to be derived again. Select a question to see the assumption behind it.

ScribesThe price of cognitive labour collapses

9. Implications for products

Financial products built on large models today all stand at the start of this change in a parameter. Three implications follow. First, to judge a product's value, look at which part of the cost it lowers. If it lowers only the cost of research and drafting while the cost of checking and bearing responsibility stays the same, the improvement users get has a ceiling. Second, because the edge of capability is uneven, checking cannot be allocated by how hard a task is; it must be allocated by the consequences of an error and by how far the result can be checked. Third, the overall amplification factor depends on the slowest step; a product must redesign account opening, authorization, settlement and checking together, rather than only making one step faster.

Chapter summary: the cost of cognitive labour collapses

Four conclusions, each tied to one mechanism in this chapter.

PriceCognitive labour is almost free

The price of reading, calculating and writing falls to a few hundredths of what it was; everything else stays the same.

CompositionChecking and bearing responsibility do not fall

The cheaper cognitive labour becomes, the larger the share of a delegation's cost taken by checking and bearing responsibility.

EdgeAbilities are uneven; checking cannot follow difficulty alone

The most dangerous tasks are those that look easy but which the scribe gets wrong.

Weakest staveThe slowest step sets the limit

Speeding up only what can be sped up leaves a ceiling on the overall factor; the reachable share depends on what is built around the model.

References