Start from the outcome.
Efficiency, growth, quality, risk or experience: name the one the business case must move, the driver behind it and the KPI that proves it. Then choose the AI.
AskWhich number must move, and by how much?
an essay on AI transformation
Electricity took forty years to pay off. Not because of the motor, but because of the factory around it. It is the most useful lesson we have for AI.
01 1883 — 1910s
1883. A year after Edison begins selling electricity as a commodity, the first factory machines run on electric power.
Most factories do the obvious thing: they replace the steam engine with one large electric motor. Same shafts, same belts, same multi-storey building.
Then came group drive: smaller motors, each turning a section of shaft. Not a mistake: it was how factories learned electricity without stopping production.
Exhibit A How American factories were powered
02 1899 — 1919
In 1899, electric motors supplied less than 5% of the power driving American factory machines.
By 1919 they supplied more than half. Electrification was the defining investment of a generation of managers.
Yet productivity barely moved: about 0.7% a year in the decades to 1909. Dynamos were “everywhere but in the productivity statistics”.
Exhibit B Electrification of US factories, and productivity
Electric motors, share of factory drive power (US)
| Year | Share |
|---|---|
| 1899 | under 5% |
| 1909 | about 25% |
| 1919 | over 53% |
| 1929 | about 78% |
03 The 1920s
Then the logic flipped. Every machine got its own motor: unit drive, not widespread until the 1920s.
With no shafts to align, the factory became single-storey, lit from above and laid out along the flow of materials.
Productivity growth rose to 5.3% a year in 1919–1929. Group drive had built the skills. Redesigning the factory turned them into growth.
Exhibit A, continued Unit drive and the single-storey factory
04 Today
Adoption puts AI in everyone’s hands: copilots, assistants, chat. It is today’s group drive, and it is necessary: it builds literacy, trust and quick wins.
Integration wires AI into existing workflows: a step is drafted, checked or routed by a model. Same process, faster steps.
Transformation redesigns the work around what AI makes cheap. The process changes, and so does the P&L.
Each stage is measured differently: usage, then cycle time, then outcomes. Mixing them up is how adoption gets reported as transformation.
Exhibit C Three stages, three ways to measure them
Stage 1
Stage 2
Stage 3
05 2022 — 2025
The cost of running a model at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024.
Open models closed the gap with closed ones, from 8% to 1.7% in a year. The two best models are 0.7% apart.
When everyone can rent the same model, the advantage moves to what surrounds it. First, data and technology: proprietary data, compute, integration into your systems.
Then people and organisation: domain experts, skills, leaders who steer the change, redesigned workflows, trust and governance. This is where most of the work is.
And finally the market: access to customers, positioning, brand, partners, regulation. The model is the motor. The moat is everything around it.
Exhibit D The motor gets cheaper; the advantage moves around it
Cost of running a model at GPT-3.5 level (index, Nov 2022 = 100)
280× cheaper in two years
Hardware costs fell about 30% a year; energy efficiency improved about 40% a year.
Performance gap between models, % of score
a year earlier in the 2025 report
AI model
1Data & technology
2People & organisation
3Market & positioning
06 2024 — 2026
53% of people used generative AI within three years: faster than the personal computer or the internet.
78% of organisations say they use AI somewhere in the business.
Official statistics are far lower: about 18% of US firms used AI in a business function at the end of 2025.
And only about 5% of custom enterprise AI tools reach production. Between using AI and running on AI is where the value leaks.
Where the work itself is redesigned, the effect is concrete: physicians report up to 83% less time writing notes.
Exhibit E Four measures of AI adoption, four answers
0%
of people used generative AI within three years of launch
People · Stanford AI Index 2026
0%
of organisations say they use AI somewhere
Organisations, self-reported · Stanford AI Index 2025
0%
of US firms use AI in a business function
Firms, official statistics · US Census Bureau, end of 2025
0%
of custom enterprise AI tools reach production
Tools in production · MIT NANDA 2025, indicative
up to 83%
less time spent writing clinical notes, where hospitals redesigned the work around ambient AI scribes
Work redesigned · Stanford AI Index 2026
Different populations and definitions: four measures, not one funnel.
07 Redesign the work
As it runs today: seven steps across three teams. Most of the time is not work, it is waiting between hand-offs.
With AI on every step, letters are drafted, documents read and amounts suggested. Each step gets faster; the route stays the same.
Redesigned: photos and policy data arrive with the claim, and AI checks cover and estimates the damage at once. Simple claims are settled in a day.
People move to where judgment matters: complex, contested or suspicious claims, with the whole file in front of them.
Exhibit F A motor insurance claim, from first notice to payment
08 Build the case
Efficiency: the same output for less cost or time. In claims, the driver is handling cost; the lever, straight-through processing.
Growth: more revenue from the same customers and channels. The driver is conversion; the lever, instant and personal quotes.
Quality: fewer errors and less rework. The driver is the accuracy of decisions; the lever, checks on every claim, not a sample.
Risk: fewer losses, frauds and breaches. The driver is leakage; the lever, fraud and anomaly detection running on every claim.
Experience: what customers and employees feel. The driver is time to settle; the lever, same-day decisions and proactive updates.
A use case without an outcome is a demo. Outcome first, then driver, then KPI, and only then the AI.
Exhibit G From outcome to AI lever: a value tree
Efficiency
Handling cost per claim
Cost per claim · cycle time
Straight-through processing of simple claims
Growth
Quote-to-policy conversion
Conversion rate · premium per customer
Instant, personalised quotes
Quality
Accuracy of decisions
Error rate · rework
AI checks on every claim, not a sample
Risk
Claims leakage and fraud
Leakage % · loss ratio
Fraud and anomaly detection
Experience
Time to settlement
Days to settle · NPS
Proactive updates, same-day decisions
09 The playbook
Efficiency, growth, quality, risk or experience: name the one the business case must move, the driver behind it and the KPI that proves it. Then choose the AI.
AskWhich number must move, and by how much?
Give everyone the tools: adoption builds literacy and trust. Then pick one or two domains with a single owner and redesign them end to end, instead of fifty pilots.
AskWhich domain will we redesign first?
Map the work as it really runs, hand-offs and waiting included. Remove the steps that only moved information, and put people where judgment matters.
AskWhich steps only exist to move information?
Models are becoming a commodity. Data, domain experts, skills, governance and access to customers are not: invest there, and make them reusable across domains.
AskWhat do we have that a competitor cannot rent?
Licences and hours saved are inputs. Track the KPI you named at the start, domain by domain, and stop what does not move it.
AskWhat did the KPI do this quarter?
The motor is not the transformation. The factory is.
Adoption puts the motor in every hand, and it is where everyone starts. Transformation rebuilds the factory around it. The model is becoming a commodity; the factory is still yours to design.
References
Exhibits C, F and G are illustrative: they show a method, not measured data.