The Line-Shaft Effect in the Age of AI (a four-part series)

science

Part 1 | New motor, same line shaft

On 4 September 1882, Edison’s Pearl Street station in New York began delivering electricity. You would expect factories to switch to electric motors quickly and productivity to take off. What actually happened was far slower: around 1900, electric motors supplied less than 5% of the mechanical power in American factories, and the real jump in manufacturing productivity did not arrive until the 1920s — nearly forty years later.

The economic historian Paul David spent years on these “missing forty years”. The answer he found was not in the motor but in the long steel shaft that ran through the factory. I call the phenomenon he described the line-shaft effect: the gains from a new technology are locked up by a way of organising built for the old one, and are released only when that organisation is torn down.

A factory run by one shaft

A nineteenth-century factory had a single source of power — a steam engine, or earlier a water wheel. Power travelled out along a steel shaft running the length of the building; the shaft was hung with pulleys, and belts carried power from them to each machine.

That shaft decided almost everything about the factory. The longer the transmission, the greater the losses, so buildings went up in storeys, stacked around the power source. Where a machine sat depended on how close it was to the shaft, not on the order of operations, and work in progress was often carried up and down stairs. Every belt turned at the same speed; even to run a single machine, the whole shaft had to turn, and if the shaft failed, the whole plant stopped. The layout served the distribution of power; how the product was made came second.

The motor arrived; the factory didn’t change

When the electric motor first appeared, owners did the obvious thing: take out the steam engine, put in one big motor, and keep driving the same line shaft. Fuel costs fell a little, control got a little easier, and otherwise the factory ran as before. David described this as laying a new technical system on top of the old structure.

Then came an intermediate arrangement, group drive: cut the line shaft into sections, give each section a motor, and let it drive a group of machines. It was somewhat more flexible, but the motor was still something added to the old equipment, and shafts and belts still ruled the shop floor.

The real turning point was unit drive: every machine gets its own small motor. At first glance that is just more, smaller motors; in fact it shook the organising principle of the factory. Once power no longer had to be distributed from a centre, buildings could be single-storey and long, and walls no longer needed reinforcing to carry overhead transmission gear. Machines could be laid out in process order, with materials flowing in one direction. A broken machine or a workshop being refitted no longer affected the whole plant. Workers could decide for themselves when to switch on. On the eve of the First World War, Ford brought the fixed assembly-line layout to maturity at Highland Park, using group drive and unit drive side by side. Delivering power by wire rather than by shaft was exactly what this flow-centred way of production needed.

Forty years of waiting, ten years of surge

When the change came, it came all at once. The share of US manufacturing driven by electricity was about 5% in 1899, just over half in 1919, and close to 80% by 1929. By the estimates of David and Gavin Wright, total factor productivity in manufacturing grew only about 0.2% a year from 1909 to 1919, then jumped to about 5.3% a year from 1919 to 1929. In the thirty years before, manufacturing output relative to capital and labour inputs had grown less than 1% a year. In the same decade the capital–output ratio in manufacturing turned down, reversing decades of increase. The reason is plain: with the heavy shafts, belts and reinforced buildings gone, the same output needed far less fixed capital.

The first reason it took so long was sunk cost. The old buildings still worked, and tearing down a functioning factory for unit drive did not pay. A 1923 Japanese study found that unit-drive equipment in textile mills cost 10% to 27% more than group drive — before counting the benefits available only in a new building. So the earliest adopters of unit drive were mostly new factories.

The second was supporting infrastructure. For a long time factories had to generate their own power. Only after 1914–1917 did the real price of electricity from regional utilities fall sharply and central-station capacity overtake factories’ own plants, making cheap purchased power widespread. Behind that lay changes in utility regulation and investment, not only technical progress.

The third is the easiest to overlook. David and Wright stress that the 1920s surge was not produced by the motor alone. Mass immigration essentially stopped around the First World War, and by the 1920s the real wages of industrial workers were 50–70% higher than a decade earlier. Dearer labour forced firms to redesign jobs, stabilise their workforces and raise hiring standards. The new factories needed workers who could read, do arithmetic and pick up new procedures quickly — which happened to coincide with the spread of American high-school education. Motors, factory layout, management methods, labour and institutions came together in the same decade, and only then did the surge happen.

One more detail is worth pausing on. Britain and Japan started electrifying their factories later than the United States, but once they started, they moved faster. By 1930 Japanese factories had electrified more thoroughly than American ones, despite relying almost entirely on imported technology. Latecomers did not have to repeat the detours; they could build straight to a design already proven to work, and many small Japanese factories jumped directly from hand workshops to unit drive.

The same story has played once already

When David wrote his well-known 1990 paper “The Dynamo and the Computer”, the question he was really answering was a different one: computers are everywhere, so why don’t they show up in the productivity statistics? He and Wright later pointed out that the spread of the personal computer in the 1980s looked a lot like group drive. Many PCs were used to emulate mainframe terminals, using a fraction of their capability; documents looked nicer and reports were easier to produce, but the organisation of work barely moved. Only when networks and systems integration forced firms to redesign their processes did the gains slowly appear.

In hindsight, the most important thing about the motor was that it let power be decentralised. Once power could be decentralised, the factory could be organised around the process rather than around the power. So the line shaft was never just a steel shaft. It was an organising principle: everything connected to one centre, turning to one beat.

Most companies today are bringing in AI much as factory owners did in 1900 — treating it as a better power source, hooking it up to the existing organisation and processes, and then wondering why the results disappoint. What the line shaft of the AI age is, is the question for the next part.

Part 2 | What is the line shaft of the AI age?

Most organisations today are still built on the premise that human cognitive labour is scarce. Hierarchy, division of labour, approvals, weekly meetings — nearly every arrangement we take for granted comes from there. AI is loosening that premise, yet the arrangements stay exactly as they were. I think they are the line shaft of the AI age.

Individuals got faster; companies didn’t get more profitable

Start with two sets of evidence that sit awkwardly side by side.

At the level of the individual, AI’s effect is fairly solid. Researchers at Stanford and MIT followed more than 5,000 customer-support agents at a Fortune 500 software company. When a generative-AI assistant was rolled out in stages, issues resolved per hour rose 14% on average; for novice and lower-skilled workers the gain was 34%, while the most experienced saw almost no change. Studies of coding, copywriting and data analysis have found similar results.

At the level of the firm the picture is different. The preliminary 2025 report from MIT’s Project NANDA estimated that enterprises had spent $30–40 billion on generative AI and that about 95% of pilots produced no measurable profit. The report was not peer-reviewed and both its sample and its definitions have been questioned, so the exact figure should not be taken too literally. Larger surveys point the same way, though. In PwC’s January 2026 survey of 4,454 CEOs across 95 countries and territories, 56% said they had so far seen no significant financial benefit from AI. In McKinsey’s global survey published in August 2026, 80% of respondents said AI had made their own work more efficient, but only 37% of organisations reported a positive contribution from AI to EBIT — essentially unchanged from the year before.

These are not the same kind of measurement. One is task efficiency measured after a staged rollout inside a real company; the other is self-reported financial results, with a time lag in between. Erik Brynjolfsson and colleagues have proposed the “productivity J-curve”: a general-purpose technology like AI demands large intangible complementary investments early on, which national accounts struggle to measure, so early productivity growth is underestimated — and later, once those investments start paying off, overestimated. Part of the gap here may simply be that it is too early.

Still, the NANDA report’s diagnosis of the cause matches the line-shaft story. It found the problem lay mainly in AI failing to connect with the organisation’s day-to-day workings: the tools could not learn from feedback and could not be embedded in how people actually work. The quality of the models themselves was secondary.

Weightier evidence comes from the previous wave of technology. In 2002, Timothy Bresnahan, Brynjolfsson and Lorin Hitt used detailed data on US firms to test the relationship between information technology and organisation. They found IT was complementary to a new form of work organisation: broader responsibilities for front-line staff, more decentralised decisions, more self-managing teams. Only where technology and organisational change moved together was the productivity effect significant. That is almost exactly the computer-age version of unit drive.

Scarce cognitive labour is an invisible shaft

The nineteenth-century line shaft came from a physical constraint: power was scarce and could only be generated centrally. Modern organisations face a different constraint: expert judgment, professional analysis and precise writing all depend on expensive, limited people. So organisations are designed around how to allocate a limited supply of human judgment.

Because good judgment is scarce, decision rights concentrate in a few hands; information goes up level by level and instructions come down level by level. Because no one can learn everything, a piece of work is split across several roles in a relay, and every hand-off produces waiting, misunderstanding and lost information. Because attention is limited, decisions are batched — hence weekly meetings, monthly reports, quarterly planning and annual budgets, the whole organisation turning to one beat, much like belts running at the same speed off the line shaft. A department’s budget and influence are usually tied to headcount. Knowledge is scattered across documents, email, meeting notes and people’s heads — readable by people, hard for machines to use.

In a world where cognitive labour is scarce, all of this is perfectly reasonable. Once AI makes a great deal of cognitive work cheap and on tap, it starts to hold things back.

Three stages

By analogy with electrification, the ways companies use AI fall roughly into three stages.

The first is the equivalent of swapping the power source: give employees chat assistants and coding copilots, and leave processes and organisation alone. Individual tasks do get faster, but the time saved is often eaten by hand-offs, waiting and approvals. It is much like using a PC to emulate a terminal — a powerful tool, used the old way. My guess is that most companies today are stuck here.

The second is the equivalent of group drive: automate a whole department or a segment of a process — customer support, code review, first-pass contract review, report generation. Local metrics improve noticeably, but upstream and downstream don’t change. There is an old rule in operations management: speed up a non-bottleneck step and the system’s output does not increase; you only pile up more work in progress in front of the bottleneck. If support replies faster but a refund still needs three levels of approval, the customer’s experience does not fundamentally change.

The third is unit drive: redesign work and organisational boundaries with AI as the basic unit. The question becomes “if cognitive labour were nearly free, how should this be done?” What that might look like: a small team with a set of agents owns something end to end, replacing a cross-departmental relay; decision rights move to the people closest to the problem, because analysis and information-gathering are no longer scarce; problems are dealt with when they arise, not at the weekly meeting; teams are measured by outcomes, not headcount.

Survey data points the same way. In McKinsey’s 2026 survey, only about 6% of organisations could attribute more than 5% of EBIT to AI. Nearly three-quarters of these “high performers” had fundamentally redesigned their workflows, against only about a quarter of the rest. The NANDA report also found that the few firms seeing returns had handed implementation to business owners and front-line managers while keeping them accountable. Both are cross-sectional surveys and cannot prove causation, but the direction is consistent with the line-shaft logic.

This shaft isn’t easy to remove either

In the electrification era, removing the line shaft was hard because of sunk costs, infrastructure and labour. All three are still here, in new forms.

Sunk cost has moved from buildings to organisational capital. Role structures, approval processes, IT systems and performance-review schemes are the product of years of investment. They still work; the cost of starting over is obvious and the benefit is hard to state. Infrastructure has moved from the grid to data and systems: whether data is readable by AI, whether systems talk to each other, whether permissions and compliance rules are clear, whether models can keep learning in a specific business. Many companies’ data is scattered across dozens of systems and countless inboxes; however capable the AI, it has nothing to get hold of.

The labour problem is sharper. The new factories of the 1920s needed workers who could learn new procedures quickly. Today’s new organisations need people who can define problems, judge AI output and take responsibility for outcomes, and managers who can lead mixed human–machine teams. Incentives are harder still: if a department’s power and budget are tied to headcount, a genuinely effective AI solution may look first of all like a threat to the department head.

And there is a problem electrification never had: accountability. If a motor ran backwards, it was clear who was responsible. When AI makes a wrong call, who is? Until there is an institutional answer to that question, many firms will instinctively put a human approval step behind the AI — which puts the line shaft right back in.

A diagnostic table

The table below can be used to place your own team.

Dimension	Stage 1: new power source	Stage 2: group drive	Stage 3: unit drive
Role of AI	Personal assistant	Automation tool for one step	Basic building block of the process
Process	Unchanged	Partly reorganised; upstream and downstream unchanged	Redesigned backwards from the outcome
Decisions	Centralised, layered approvals	Faster locally; approval chain unchanged	Pushed to the people closest to the problem
Rhythm	Weekly meetings, monthly reports, quarterly plans	Real-time locally; batched overall	Continuous, triggered on demand
Measures	Usage rates, hours saved	Step metrics (response time, throughput)	End-to-end outcomes (cycle time, quality, cost)
Organisational boundaries	Functional departments	Functional departments	Small teams organised around outcomes
Where the bottleneck is	Employees’ willingness to adopt	Unreformed upstream and downstream steps	Human judgment and accountability

A simple self-test: take away every AI tool you use. Does the org chart need to change? If not, you are probably still in stage one or two.

The history of electrification also reminds us that unit drive appeared first in new factories, and latecomers often moved faster. Today, a new organisation built AI-native from day one may reach the gains sooner than an old one in the middle of a retrofit. But old organisations have things new ones don’t: customers, data, domain knowledge and trust. The key is whether they are willing, in some well-bounded area, to design a process from scratch as if building a new factory — rather than bolting new motors onto the old process everywhere.

The next part turns to the industry I know best: what the line shaft of drug discovery is, and why AI’s record there is almost a replay of “new motor, same line shaft”.

Part 3 | The line shaft of drug discovery

First, where I stand: I have a start-up in this industry, so my judgments below inevitably carry a practitioner’s bias. Readers can discount accordingly.

My view is that the line shaft of drug discovery is the R&D pipeline itself — from target to launch, run as a staged relay with a gate at each step. Over the past decade AI has been installed in nearly every step of that pipeline, and locally things really have got faster. But the judgment that decides whether a drug succeeds does not sit in those steps.

A factory run as a staged relay

Think of drug discovery as a factory. The production line runs roughly like this: find a target that is linked to the disease and can be modulated by a drug, and validate it; screen a vast number of molecules for hits that act on the target; modify the molecules over and over to get potency, selectivity and drug-likeness right; then run animal studies, toxicology and pharmacokinetics, and file for the clinic. The clinic is itself split into Phase I, II and III, testing in turn safety in humans, early efficacy and confirmatory efficacy, before approval and launch.

Each stage is run by a different team, often at a different company: academia discovers the target, a biotech does early R&D, contract research organisations run the experiments, and a large pharma company buys the asset and takes it through the clinic. Between stages there are hand-offs, and one go/no-go review after another.

The pipeline has its logic. Experiments are expensive and slow, expert judgment is scarce, and risk can only be controlled in stages — so use cheap experiments first to cut most candidates, then concentrate expensive resources on the few that remain. In an era when both cognition and experiments were scarce, this was a very rational arrangement.

The cost is plain too. A new drug typically takes more than ten years, and only about one in ten projects that enter the clinic is eventually approved. The industry also talks of “Eroom’s Law” — Moore’s Law backwards: since 1950, the number of new drugs per billion dollars of R&D has halved roughly every nine years, falling eighty- to a hundredfold over sixty years. Around 2010 the decline stopped and the curve has since been roughly flat, but low R&D productivity remains the consensus.

Like the factory’s line shaft, the pipeline shapes everything it carries. Projects can only move serially: the next stage does not start until the last one ends, every exit is guarded by a gate, and overall speed is set by the slowest step and the most conservative review. Each stage optimises its own metrics — chemists watch potency and drug-likeness, toxicologists watch safety, clinical teams watch enrolment and endpoints. Each does its part well, yet no stage is directly accountable for whether the drug will ultimately help patients. When a project fails, the reasons usually stay in an internal report and rarely flow back upstream, so the next target choice learns little from it; across the industry, most negative results are never published at all. Data is generated all the time, but decisions wait for the quarterly review.

Where AI has been installed

AI has entered nearly every section of the pipeline: target prediction, protein structure prediction, molecule generation, virtual screening, ADMET prediction, clinical trial design and patient recruitment — each with an exciting story of acceleration. To judge the overall effect, you have to look at the data coming out of the end of the pipeline.

In 2024, researchers from Boston Consulting Group published the first systematic analysis in Drug Discovery Today, counting the clinical pipelines of AI-native biotechs. Since 2015, these companies and their pharma partners had taken 75 AI-discovered molecules into the clinic. Of the 24 molecules that had completed Phase I by the end of 2023, 21 succeeded — a success rate of 80–90%, against a historical industry level of roughly 40–65%. Of the 10 that had completed Phase II, 4 succeeded, about 40%, comparable to the historical 30–40%.

These numbers need handling with care, and the authors themselves repeatedly stress that this is a first analysis. Phase II has only ten molecules, and four out of ten carries enormous statistical uncertainty — the 95% confidence interval runs from about 12% to 74%. More importantly, when the authors looked case by case, only 2 of the 6 molecules that did not advance past Phase II were stopped because of negative trial data; the other 4 were stopped for reasons such as shifting commercial priorities or difficulties in running the trial. In other words, the Phase II data says almost nothing about how effective AI-discovered molecules really are. The high Phase I success rate may partly reflect selection: early AI companies tended to pick better-validated targets. The authors think this cannot explain the whole difference, but cannot rule it out either. They also ran a thought experiment: if Phase I and II success rates held at these levels, a molecule’s probability of getting through the whole clinic would rise from 5–10% to 9–18%, nearly doubling R&D productivity. It should be said that the authors work for a firm that advises pharmaceutical companies, and the study was funded by that firm.

Even taken only as an early signal, the structure of the data is worth thinking about. Phase I tests mainly safety and pharmacokinetics — properties of the molecule itself: is its toxicity manageable, can it reach the right concentration in the body? That is exactly what AI is best at: computable, predictable, with plenty of historical data to learn from — and it is where the advantage in the data is clearest. Phase II tests efficacy: does acting on this target actually improve this disease? In the industry’s historical statistics, more than half of Phase II failures are for lack of efficacy, and efficacy depends largely on the decision at the very top of the pipeline — whether the target and the biological hypothesis are right.

AI has shown it can make good molecules. But the real bottleneck in drug discovery is choosing the right problem. The first can be optimised in a closed loop within a single stage; whether the second was right is revealed only five to ten years later, when Phase II data arrives. AI is moving upstream too: by 2023, molecules involving AI-discovered targets made up more than 30% of this clinical pipeline, and we can confirm at least three molecules against entirely novel targets that have passed Phase I. But whether AI can help us past the efficacy hurdle is something the existing data cannot yet answer.

Speed up a non-bottleneck step and the system’s output does not rise. However fast and well a molecule is designed, if the target it acts on is wrong, it will still fall in Phase II — just more “cleanly”.

Why this shaft is especially hard to remove

Applying the three stages from Part 2: pharma’s first stage is giving scientists AI assistants for literature search, report writing and code. Individual productivity goes up; the process does not change at all. The second stage is deploying an AI platform in one step — generative chemistry, virtual screening, AI-driven contract services. This is today’s mainstream and the business model of most AI drug-discovery companies. One thing deserves a second thought: most partnerships between big pharma and AI companies set milestones by pipeline stage. The AI company delivers a preclinical candidate, collects a milestone payment, and the project then enters the pharma company’s existing development machine. That deal structure itself reinforces the old line shaft; the AI company becomes a new motor hooked up to the old shaft. As for the third stage — redesigning R&D around a new basic unit — it is still rare in the industry.

Beyond the reasons in Part 2, drug discovery’s line shaft is especially hard to remove for several more. Part of the shaft is statutory: clinical phases and preclinical standards belong to the regulatory system and cannot — and should not — be casually dismantled, so the room for redesign lies mainly in preclinical work and clinical design. Experiments have a physical bottleneck: computation can finish instantly, but cells grow at their own pace; AI can reduce the number of experiments needed, not make an experiment finish sooner. The business model is built on the pipeline: licensing deals, milestone payments and stage-based valuations in financing — the industry’s entire financial structure prices assets by pipeline stage, so R&D not organised by stage is hard even to value. Data is fragmented across stages and companies, in different formats, with most negative results unpublished, so what AI can learn is often just one section of the pipeline. And evaluation rewards pipeline length: many teams are judged on the number of projects and how far they have progressed, rarely on how sound the hypothesis behind them is.

Unit drive brought a leap because the factory could at last be organised around the process. In drug discovery, the question becomes: when molecule design, data analysis and reasoning over the literature are no longer scarce, what should R&D be organised around? My answer is the hypothesis, which is the subject of the final part.

Part 4 | After the line shaft

This part sets out how I imagine R&D after its line shaft is removed. It is also the direction we ourselves are betting on, so it is a position as much as an observation.

Briefly: I think the next form of R&D organisation should take the hypothesis as its basic unit, close the loop tightly between computation and experiment, and bring the most upstream, most critical judgment forward so it can be tested repeatedly and cheaply.

From molecule to hypothesis

What unit drive really did was change the basic unit of the factory, from “one shaft driving a group of machines” to “one motor driving one machine”. Only when the basic unit changed did layout, process and management follow.

The basic unit of drug discovery today is the molecule. Molecules are what flow through the pipeline, what review meetings discuss and what deals buy and sell. By a hypothesis I mean a falsifiable biological statement — for example, “in a certain class of patients, modulating a certain target reverses a certain disease phenotype”. A molecule is just one of the tools for testing it. Organise R&D around the molecule and the central question is how to make that molecule good. Organise it around the hypothesis and the central question becomes how to find out, as fast and cheaply as possible, that the hypothesis is wrong.

There is evidence for this direction. A 2015 study estimated that drug mechanisms supported by human genetic evidence were about twice as likely to go from the clinic to approval as those without; an updated study in Nature in 2024 put the ratio at 2.6, with the advantage showing up mainly in Phase II and III — the stages that test efficacy. Testing the hypothesis more rigorously before the molecule even exists can directly change the final success rate.

What unit-drive R&D would look like

Mapping the changes unit drive brought to the factory, we can roughly picture an R&D organisation with its line shaft removed.

The factory went from stacked storeys to a single-storey flow line; R&D goes from a serial relay to closed-loop iteration. Computation proposes hypotheses, experiments test them quickly, and results flow straight back to correct the models, with each loop measured in days or weeks rather than stages. The system’s speed is set by the cycle time of the learning loop, no longer by the slowest review meeting.

The emphasis shifts from “making a good molecule” to “falsifying early”. Evidence bearing on efficacy should be moved as far forward as possible, using experimental systems closer to humans — patient-derived cells, organoids, functional genomics — to see whether the hypothesis holds up in human-relevant systems before large chemistry resources are committed. The question Phase II has to answer should, as far as possible, be partly answered upstream.

The nature of failure changes too. Unit drive meant one broken machine no longer brought down the whole plant. In the new R&D organisation, every falsified hypothesis and every negative experimental result should flow back into the models and the knowledge base, making the next hypothesis better. In the old pipeline, failure is a sunk cost to be forgotten as fast as possible; in the new organisation, failure is training data.

In organisational form, a small team with a set of AI systems owns the work from hypothesis generation through to key validation, with no relay across five departments, and decisions are made when the data comes in. The measures change accordingly: the old pipeline counts projects and how far they have progressed; the new organisation should instead look at what it costs to test a hypothesis, how long a learning loop takes, and what share of the hypotheses that reach the clinic turn out to be right.

This vision can be tested, and should be. If it holds, projects that are hypothesis-centred and thoroughly tested for falsification upstream should have a clearly higher Phase II efficacy success rate than the industry baseline. That will take years of data and a large enough sample, but it is the gate this argument ultimately has to pass.

New factories and old

The lesson of electrification is that unit drive appeared first in new factories. Drug discovery will probably be the same: an AI-native R&D organisation can be designed around hypotheses and closed loops from day one, without the baggage of an old pipeline. The new factory’s weaknesses are just as clear: it lacks clinical development capability, large-scale patient data, and the money and experience to go through a full R&D cycle — exactly what big pharma has.

So the more realistic path, I think, is to pick a well-bounded area — a disease area, say, or an early discovery unit — build a process on the new logic as if building a new factory, get it working, and then expand. Big pharma rebuilding its R&D system from the ground up is unrealistic; so is an AI company going the whole way alone. Compared with fitting AI into every section of the old pipeline, this has a better chance of producing system-level returns.

Japan back then carried none of the first-mover’s baggage, and electrified more thoroughly for it. Whether R&D ecosystems today with less accumulated legacy pipeline and fast experimental execution can skip the “new motor” step is worth watching.

The limits of this argument

Analogies have their limits. Much of the gain from electrification was physical — no more shafts, belts and reinforced buildings. The fundamental bottleneck in drug discovery is biological knowledge itself. Our understanding of many diseases still has large gaps; AI models are good at interpolating within the range of existing data and may not be reliable in the face of genuinely unknown biology. Redesigning the organisation can help us find out faster that we are wrong, but it does not guarantee we will find the right answer.

Experiments also have physical speed limits. However tight the loop, cells grow at their own pace, and animal studies and clinical trials take time; AI can let us run fewer experiments but cannot speed up biology. And some line shafts protect patients. Clinical phases and regulatory review are not inefficient relics; what needs to be dismantled is how R&D is organised, not the safety floor.

Nor does history simply repeat. The 1920s surge depended on motors, factory layout, the labour market, electricity prices and the macro investment climate all falling into place at once, and David himself warned repeatedly against using historical resemblance to predict the specific path of the future. There is no guarantee that AI’s gains will eventually be released in a concentrated burst the way electrification’s were. Drug discovery cycles are long to begin with, and whether our current judgment is right will also take time to tell.

Where is your line shaft?

Back to the motor. It replaced the steam engine, but what really changed the factory was that distributing power from a centre was no longer necessary. AI is replacing a great deal of cognitive labour; its biggest effect may be that organising everything around scarce human judgment is no longer necessary — and in drug discovery, perhaps, that relaying molecules through stages is no longer necessary.

So for an organisation bringing in AI, a question more worth its time than “which model should we use?” is this: what is our line shaft, what constraint was it built under, and does that constraint still hold today?

References

Part 1

1. Paul A. David. The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox. American Economic Review, 80(2), 1990.
2. Paul A. David & Gavin Wright. General Purpose Technologies and Productivity Surges: Historical Reflections on the Future of the ICT Revolution. In The Economic Future in Historical Perspective, Oxford University Press, 2003.
3. Warren D. Devine Jr. From Shafts to Wires: Historical Perspective on Electrification. Journal of Economic History, 43(2), 1983.
4. IEEE. Milestones: Pearl Street Station, 1882.

Part 2

5. Erik Brynjolfsson, Danielle Li & Lindsey Raymond. Generative AI at Work. NBER Working Paper 31161, 2023.
6. MIT Project NANDA. The GenAI Divide: State of AI in Business 2025. 2025 (as reported by Fortune).
7. PwC. 29th Global CEO Survey. 2026.
8. McKinsey & Company. The state of AI in 2026: On the road to ROI. 2026.
9. Erik Brynjolfsson, Daniel Rock & Chad Syverson. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics, 13(1), 2021.
10. Timothy F. Bresnahan, Erik Brynjolfsson & Lorin M. Hitt. Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence. Quarterly Journal of Economics, 117(1), 2002.

Part 3

11. Madura K. P. Jayatunga et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 29(6), 2024.
12. Jack W. Scannell et al. Diagnosing the decline in pharmaceutical R&D efficiency. Nature Reviews Drug Discovery, 11, 2012; for the trend after 2010, see Scannell’s paper for the OECD, Eroom’s Law and the decline in the productivity of biopharmaceutical R&D, 2023.
13. John Arrowsmith. Phase II failures: 2008–2010. Nature Reviews Drug Discovery, 10, 2011 (as reported by Fierce Biotech).
14. John Arrowsmith & Philip Miller. Phase II and Phase III attrition rates 2011–2012. Nature Reviews Drug Discovery, 12, 2013.
15. Richard K. Harrison. Phase II and phase III failures: 2013–2015. Nature Reviews Drug Discovery, 15, 2016.

Part 4

16. Matthew R. Nelson et al. The support of human genetic evidence for approved drug indications. Nature Genetics, 47, 2015.
17. Eric Vallabh Minikel et al. Refining the impact of genetic evidence on clinical success. Nature, 629, 2024.