The Washing Machine Problem

“MY WASHDAYS ARE HOLIDAYS NOW.”

So ran the headline of a washing-machine advertisement in 1952. The advert sold a machine but the promise wasn’t really cleaner clothes, it was time handed back.

However, the holiday never quite arrived.

When the historian Ruth Schwartz Cowan went looking for it three decades later, she found a different story. Her study of household technology, More Work for Mother, showed that American women spent roughly as many hours on housework in the 1960s as their grandmothers had at the turn of the century, surrounded by machines those grandmothers could not have imagined.

The washing machine made washing easier, but standards rose to meet it. Clothes once washed when dirty came to be washed as routine. Commercial laundries and paid help fell away, and work that had been spread among hired help and the laundry down the road was gathered back into the home and handed to one person. The task became lighter, but as expectations rose, the workload did not.

We refer to it as the washing machine problem: the saving everyone makes is often the saving no one keeps.  And AI is running the experiment again, this time on the people who design, cost and manage buildings (i) by raising standards (ii) moving the work and (iii) cheapening the signal.

The standard rises to swallow the saving

We all know that AI can make individual tasks significantly faster. A report that once took a day can now be drafted in an hour, at a speed that feels faintly indecent. However, task productivity and organisational productivity are not quite the same thing, and the gap between them is precisely where the washing machine has something to teach us.

The hour saved is not set aside in a jar marked margin, still less one marked ‘spare time’. It is converted, more or less on the spot, into speed, or polish, or optionality. The report that use to take a day now takes an hour, so five are commissioned, each with three scenarios and a better graphic.

William Stanley Jevons noticed, in 1865, that more efficient steam engines increased Britain’s consumption of coal rather than reducing it. Make a thing cheaper and people discover they wanted far more of it than anyone had supposed. Planning has illustrated this paradox for years: the Lower Thames Crossing application, submitted a month before ChatGPT appeared, grew to more than 359,000 pages.

Inside many firms now, a first draft arrives in minutes, fluent and occasionally, confidently, wrong. Generation of plausible answers is cheap; verification they are the right ones, less so. The drudgery has moved from the writing to the reading, much as it once moved from the washtub to the ironing board.

The dashboard sees nothing

CAD and BIM were sold as instruments of productivity, and each made particular tasks unambiguously faster. Yet the ONS has found little or negative growth in output per hour across every part of construction between 1997 and 2020. One reading is that the gains were spent before anyone could count them, on more options, more iterations and client expectations that rose obligingly to meet them. The return arrived as complexity rather than cash, in buildings more serviced, more regulated and more examined than their predecessors.

Economists have a name for gains that take time to show: the productivity J-curve. The effort of reorganising around a new technology is intangible, so the accounts record the cost and miss the asset. Any complaints that the cost of AI licence have produced nothing may be looking at the left-hand side of the J.

But the curve turns upwards for firms that do the reorganising. A firm that adopts AI and changes nothing else will find its people doing more of everything, to a higher standard, for the same fee. The saving does not exist until someone redesigns the work, the pricing or the offer to capture it.

The work moves

Some firms are redesigning the work, though not always where it matters most. The easiest saving to find sits at the bottom of the organisation chart, where the tasks are most routine and the machine is most competent. A survey for Lancaster University’s Work Foundation, published in August, found that more than a third of UK employers had cut entry-level jobs in the past year. Around six in ten larger firms said AI or automation had reduced the roles on offer. The saving is being taken after all, from the graduate intake. The larger loss may be harder to see: the work through which juniors used to become useful.

The graduate QS or second-year engineer drafting the first cut of anything is sometimes producing work that needs rescuing. But the rescuing is the point. Judgement has historically been built by doing the rough version, watching it get pulled apart and going again with a little less misplaced certainty. Much of what a professional knows is absorbed by standing next to someone who has it, on tasks unimportant enough to be safely fumbled.

While the junior’s first draft does two jobs, the efficiency measure sees only one.[1] It is inefficient as production and highly effective as training. Some inefficient work is genuine waste; some is the training system wearing the costume of waste (often referred to as ‘character building), and the two look identical on a spreadsheet. Strip out the rough draft without rebuilding the learning loop and you’ve not eliminated inefficiency; you’ve deleted the apprenticeship and booked it as a saving.

When everyone has a clean shirt, the shirt says nothing

For firms like ours, the third effect may reach furthest. Polish has long done useful work as a signal. A literate proposal and a coherent deck told the client something real about the organisation behind them, because producing them took scarce time and scarcer talent.

Michael Spence won a Nobel prize for showing that a signal carries information precisely because it is costly to send. When everyone can send it cheaply, it stops being one. AI makes polish cheap, and in doing so it weakens one of the old proxies for competence. When every proposal reads well, fluency stops being evidence of very much and the hard question becomes who is competent, and how anyone can tell.

Spence shared the 2001 prize with George Akerlof, whose market for lemons describes what follows. When buyers cannot tell good from bad, prices drift towards the average and the genuinely good struggle to prove they are worth more. Quality ultimately fades because the market cannot see it clearly enough to pay for it, however much clients want it.

Scale can widen the gap. A large consultancy can combine brand recognition, polished materials and a confident lead into a convincing proposition, even where the specialist knowledge behind it is relatively thin. That is admittedly a familiar complaint from an SME, but it is also a genuine client problem. When scale and fluency are easier to observe than expertise, procurement can select the firm that looks safest rather than the one that best understands the work.

AI lowers the cost of looking credible faster than it lowers the cost of being credible.

Where the saving goes

None of this makes a higher standard a bad thing. The clothes did get cleaner, and a better building is worth having. The trouble is output nobody asked for, produced because it was cheap rather than because anyone needed it. Firms should decide how to use the saving before competition and habit decide for them. Three questions help, one for each effect.

• Whose saving is it? A firm that sells hours and adopts a machine that shrinks them has automated its own invoice. Unless it reprices around what the machine cannot do, competition will hand the gain to the client, as it handed the washing machine’s to the household.
• Where will judgement be formed? If the machine takes the rough draft, juniors need another way to fumble safely. That is a training cost, and it is cheaper budgeted now than discovered in a decade as a shortage of people able to check the machine.
• How will a client see competence? When polish is free, firms have to show their working. Procurement are likely to place greater emphasis on testing the people who will do the work, not the document that describes them. Expect interviews, case exercises and named-team assessments to count for more, and written submissions for less

The holiday that never arrived

AI will save time. So did the washing machine. It changed who did the washing and what counted as clean.

Everyone will soon have a clean shirt. The firms worth hiring will be the ones who can still tell you how it got that way.

[1] Our observation that apparent inefficiency often serves a purpose the efficiency measure cannot see is supported by a few, not least Rory Sutherland.