In 1895 Professor Crocker of Columbia University travelled to Taftville, Connecticut, to see a new hydroelectric plant power the Ponemah textile mill. He found the electric motors in the basement, next to the engines they had replaced, with one of them driving 1,200 looms through the same shafts, belts and pulleys the old engines had turned.
New power, old factory
Most of the AI programmes I see look like Ponemah. A portfolio of use cases. A handful of pilots that work, and a budget line renewed if the pilots look good. The same models are available to everyone. Used in much the same way, they produce much the same results: a process runs a bit faster here, a team saves a few hours a week there, and the profit and loss account looks much as it did before anyone signed a licence. It is tempting to blame the models.
The ceiling sits in the factory: in how the business is designed, funded and governed. A century ago, investment in new factories created the opportunity to reorganise production around electricity. AI comes with no new factory. The decision has to come from the leadership team.
We have been here before
Electricity. By 1919, more than half the mechanical power driving machines in American factories was electric, up from under 5% in 1899. Productivity had barely responded.
Firms had bolted motors onto the old shafts and paid for both. In US manufacturing, labour productivity grew around 1.5% a year from 1899 to 1914. Suppliers had sold electricity on savings, promising 20 to 60% off the coal bill. Coal use did fall, typically by about a fifth. But fuel was only 0.5 to 3% of production costs. The larger gains came when factories moved towards one motor per machine: the shafts disappeared, machines were placed in the order the work flowed, and buildings could become lighter and single-storey. Some manufacturers reported 20 to 30% more output from the same floor space, machines and people. From 1919 to 1929, US manufacturing labour productivity grew 5.1% a year. Capital became more productive too.
The Nordic evidence points the same way. Swedish industry was about half electrified by 1914–15, ahead of the US. Yet between 1920 and 1930, Swedish industries that used electric motors most intensively had labour productivity growth of 4.3% a year, compared with 4.1% in less intensive users. The difference wasn’t statistically significant. In Finland, manufacturing output per worker grew by just 0.3% annually from 1901 to 1920, then by 4.2% from 1920 to 1938.
Computers. In 1987 Robert Solow remarked that the computer age showed up everywhere except in the productivity statistics.
He had a point. Computers’ share of the services provided by US business equipment rose from about 0.6% in 1979 to 18.4% in 1992, while US nonfarm business labour productivity was growing at roughly 1.5% a year through the early 1990s. The acceleration came later. Firm-level researchers found that technology worked better with complementary organisational changes: companies ranking high in both IT investment and decentralisation were, on average, about 5% more productive than companies ranking high in only one. Buying computers without changing work practices was no guarantee of improvement.
A case study published in 1997 shows how that happens. A large medical products manufacturer, given the pseudonym MacroMed, invested heavily in flexible computer-integrated manufacturing to offer more product variety. The results fell well short of expectations. Workers kept the habits of the old line, such as avoiding stops for product changeovers, and ran the flexible new machines much like the old ones. Management eventually installed the system at a greenfield site with a hand-picked, younger workforce, and there it delivered.
Internet and mobile. Both made it cheap for incumbents to add a channel or a product while someone else rebuilt the business underneath. Blockbuster launched an online subscription service in 2004, the year it had about 9,100 stores and .1 billion in revenue, but its own annual report presented the service as a complement to its rental stores, with free coupons to bring subscribers back through the door. Blockbuster filed for bankruptcy protection in 2010. In mobile, value moved to the platforms with the largest app ecosystems, iOS and Android. Nokia stayed with Symbian, switched to Windows Phone in 2011 and sold its handset business in 2013, by which time its share of global smartphone sales had fallen from 49% in 2007 to 3%.
None of these stories proves a single cause. The electricity figures span a world war, a Finnish civil war and an American immigration stop, part of the computer gains came from the industries making the computers, and Blockbuster and Nokia made many decisions besides the ones above.
What the three shifts share is narrower, and still useful: adopting a technology was never the same as building a business that could win with it.
Every shift came down to three answers
The three stories suggest three choices, and I would put them to any leadership team today:
- Do we expect AI to change how value is created or costs are structured in our industry, beyond making individual tasks more efficient?
- Do we want our AI efforts to give us an advantage over competitors, rather than keep us level with them?
- Are we willing to change how we decide, fund and organise to get there?
The first response to a general-purpose technology is often to extract savings from existing work: cheaper power, faster paperwork, a new sales channel. Those improvements can be worthwhile, but adoption alone doesn’t answer the second question. A durable advantage depends on what a company builds around technology: its processes, skills, information, decisions and incentives.
The third determines whether the first two answers have any chance of becoming more than intentions.
Each no is a legitimate answer. If you expect only local efficiency gains, automate where it pays and don’t overbuild. If keeping level with competitors is enough, follow the market and save the money.
The expensive answer is yes, yes, no.
Yes, yes, no is playing to play
In Playing to Win, A.G. Lafley and Roger Martin distinguish between playing to win and participating without making the strategic choices winning requires. That is what yes, yes, no amounts to.
Playing to win is no guarantee of winning. The value of AI is still a hypothesis, so a strategy for it is a set of bets: where to place them, how much to stake on each, and what evidence would make you stop. Some will fail. Playing to play avoids that risk, and with it most of the chance of winning.
Nobody sets out to produce AI theatre, but an organisation built for predictable returns produces it anyway once you ask it to make bets. Budgets demand predictable returns. Targets are set on EBITDA. A failed bet ends up on someone’s record, while a missed opportunity never appears anywhere. Caution is rational for each leader and often wrong for the company.
AI has no new factory to force the decision
The analogy has limits, and they matter. AI spreads much faster than electricity did, reaches office work as well as physical operations, and the models keep improving. History won’t tell us how long any productivity lag will be.
Two differences make the risk sharper. New plant investment gave manufacturers an opportunity to redesign production around electricity. AI often arrives as a subscription layered on top of the existing organisation, and nothing in that subscription forces anyone to change how work gets done. AI is also relatively cheap to start using. That makes it possible to spend years adding copilots to existing workflows while competitors redesign the economics of an entire service or value stream.
Three mistakes are worth avoiding:
- Buying the coal savings. Licence counts and cheaper tasks may be worthwhile, but they aren’t the whole opportunity.
- Mistaking adoption for advantage. Using the same models isn’t a strategy for building something competitors will struggle to reproduce.
- Funding AI inside the process it replaces. Paying for AI out of an existing process budget makes it difficult to justify changes whose benefits accrue elsewhere. The investment boundary becomes the limit of the ambition.
Answer the three questions before the 2027 budget is locked
Budgets for 2027 are being set. That is your decision point. A budget built like the 2026 one risks buying another year of disconnected pilots.
Put the three questions on the agenda of your next leadership team meeting and ask each member to answer them alone, in writing, before anyone compares notes. If the answers differ, that is your first finding.
If you land on three yeses, you have decided to play to win, and that calls for a strategy: explicit choices about where AI should change your business and how you intend to win there. A longer list of use cases won’t do that job.
If you land on yes, yes, no, stop funding pilots without a path to changing the business. Choose a value stream or capability where you expect AI to change customer value or cost structure, then ask what you would have to change in how you decide, fund and organise to make that possible.
Otherwise 2027 will be another year of new power in an old factory.
Read also Snurre’s last blog post: Deciding what AI agents may decide is an old problem. We keep treating it as new.