Leading through the dip: the productivity J-curve and the real work of the AI transition

Overview

This weekly edition is about the strangest pattern in this year’s working life: organisations that are growing, cutting jobs, and spending more on AI than ever — all at once, sometimes in the same announcement. If that sounds like the news you scroll, or the place you work, this is a map for reading it.

Why now: in May 2026, Cloudflare announced the first mass layoff in its sixteen-year history — about 1,100 roles, a fifth of the company — in the same earnings report as record quarterly revenue. Meta is cutting roughly 8,000 roles while guiding to US$125–145 billion of AI capital expenditure this year — guidance it raised again in late April, and roughly double last year’s US$72.2 billion. And the productivity all that spending promises is, for most firms, still hard to find on a dashboard.

There is a well-studied economic pattern that makes sense of the contradiction — and it is more hopeful than the headlines suggest. By the end you’ll have the pattern, the century of history behind it, and the three signs that an organisation — including the one around you — is doing the real work of the transition.

The content

Start with the pressure, because it is real and it is not a character flaw. A leader today answers to a quarterly clock: boards and markets want an AI story now. But the underlying work of absorbing a general-purpose technology runs on a multi-year clock. IBM’s 2025 survey of 2,000 CEOs caught the squeeze precisely: only 25% of AI initiatives had delivered their expected return on investment, only 16% had scaled enterprise-wide — and yet those same CEOs expected their AI investment growth to more than double within two years. Investing ahead of proof, under pressure, with imperfect information is not recklessness. It is the standard condition of leading through a technology transition, and the historical record says so.

The record’s best chapter is electricity. Light bulbs existed by 1879 and generating stations in New York and London by 1881 — yet by 1900 the productivity effect of electrification was nearly invisible, and the surge in US manufacturing productivity arrived only in the 1920s, roughly four decades after commercialisation. The economic historian Paul David, who documented this in 1990, found the reason in the factories themselves: owners initially unbolted the steam engine and bolted a big electric motor in its place. Same building, same shafts and belts, same workflow. Nothing improved. The payoff came only when factories were redesigned around the new possibility — single-storey layouts arranged by the logic of the work, a small motor at every workstation, and a workforce retrained to use the autonomy that created. The first electrified factories were not run by fools. They were run by capable people part-way through a redesign nobody had a map for.

Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson gave the modern version of this pattern a name: the productivity J-curve. General-purpose technologies, they show, “enable and require significant complementary investments” — business process redesign, co-invention of new products and business models, and investment in human capital. Those investments are intangible: they are real, they create value, and they are nearly invisible to standard measurement while they’re being made. So measured productivity understates progress in a GPT’s early years, and the benefits are harvested later. The dip comes before the climb, and the dip is not the technology failing — it is the redesign work showing up in the accounts as cost without output.

Read it that way and the recent past looks different. A restructure that didn’t deliver what the plan promised, a tool that disappointed — these are not proof that the people behind them got it wrong. They are what the invisible stretch of the curve looks like from inside it. The question the J-curve actually asks of an organisation is narrower and more useful: is it making the intangible investments that convert the dip into the climb? Three stand out — and you can see all three from any desk in the building, not just the corner office.

Measure outcomes, not activity. Usage statistics describe the dip; outcomes describe the climb. Klarna’s much-watched experiment is instructive — and to the company’s credit, it ran the loop in public. In early 2024 it reported its AI assistant doing “the work of 700 human agents”; by May 2025 its CEO acknowledged that cost had been “a too predominant evaluation factor”, that quality had suffered, and that “really investing in the quality of the human support is the way of the future for us”. That is not a cautionary tale about AI. It is a working demonstration of the discipline the transition demands: measure the outcome, notice honestly, redesign, go again.

Redesign the work, don’t retrofit the tool. The agentic equivalent of the steam-to-motor swap is bolting AI onto an unchanged workflow and waiting for transformation. The unlock, then as now, is deciding which work changes shape: which tasks become agentic, where human judgement sits by design rather than by leftover, and what the role around the new workflow looks like. That is slower than a tool rollout. It is also where all of the historical gains turned out to live.

Build the people system early, because it has the longest lead time. The World Economic Forum’s Future of Jobs survey projects that between 2025 and 2030 the equivalent of 170 million jobs will be created and 92 million displaced — and that of every 100 workers, 59 will need training by 2030, including 19 who can be upskilled and redeployed within their own organisation. Redeployment at that scale does not happen by goodwill. It requires knowing what skills you actually have, where the changing work is, and how people move between the two — capabilities that take years to build and pay off across every restructure that follows. Employers in the same survey named skill gaps the biggest barrier to transformation. The firms that exit the dip first will mostly be the ones that started this part earliest.

Four decades separated the dynamo from the payoff it promised. Few expect AI to take that long — diffusion is faster now, and the tools improve themselves. But the J-curve is a pattern, not a promise: the dip becomes a climb only where the complementary investments are actually made. That work — redesigning processes, measuring outcomes, moving people deliberately — is the part only leadership can do. The rest of us can learn to recognise it, because it is one of the most useful signals available about where an organisation is headed — and about whether the work around you is being rebuilt, or merely re-tooled.

Additional reading

• The Dynamo and the Computer — Paul A. David, AER, 1990 — the original argument: why electrification’s payoff waited on factory redesign, and what that implied for computing.
• It took decades for the economic impact of electricity to become clear — Tim Harford, FT, 2007 — the most readable account of David’s dynamo history.
• The Productivity J-Curve — Brynjolfsson, Rock & Syverson, AEJ: Macroeconomics, 2021 — the formal model; see also the MIT IDE research brief for the plain-language version.
• Cloudflare says AI made 1,100 jobs obsolete, even as revenue hit a record high — TechCrunch, May 2026 — the clock mismatch in a single earnings report.
• Meta to cut 8,000 jobs with more planned for late 2026 — The Next Web, April 2026 — workforce cuts alongside a near-doubling of AI capital expenditure; guidance since raised to US$125–145 billion at April’s earnings.
• Klarna plans to hire humans again — Fortune, May 2025 — the outcome-measurement lesson, run in public.
• IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles — IBM, May 2025 — 2,000 CEOs on ROI delivered versus investment planned.
• Future of Jobs Report 2025 — World Economic Forum — the displacement, creation, and redeployment projections to 2030.

Editor’s note

My own AI dividend arrived relatively late. First I experienced the challenges of tools that didn’t work for me, workflows that I couldn’t split in a way that made sense, and each month a subscription cost with nothing to show for it. The return only came once I stopped simply adding AI to how I already worked and instead rebuilt the work around the outcome I was actually after. It was true for me at the personal level, and we can see leaders grappling with it at the organisational level. Naming and measuring outcomes is harder than providing tools and measuring use, but that is where success actually lives.