Economics Archives - The AI Prism

The AI Productivity Paradox: Why the Gains Haven’t Shown Up Yet

The AI Prism Admin — Sun, 09 Aug 2026 10:00:00 +0000

We’re using AI everywhere and measuring it nowhere.

Walk into almost any knowledge-work office in 2026 and you will find assistants embedded in the email client, the code editor, the CRM, the support queue and the slide deck. Ask the same organisation what any of it did to output per hour, and the room goes quiet.

That silence is the paradox. Adoption is close to universal at the individual level, spending is enormous, and yet the macro statistics that are supposed to register a productivity boom look stubbornly ordinary. Something has to give — either the tools do less than the demos suggest, or our instruments are pointed in the wrong direction.

The honest answer, laid out well in Bjorn Roche’s essay on the AI productivity gap and in the long Hacker News discussion it kicked off, is that both are partly true — and that neither is a reason to panic. This is what technology adoption has always looked like from the inside.

The Paradox, Stated Plainly

The shape of the problem is old enough to have a name. In 1987 Robert Solow observed that you could see the computer age everywhere but in the productivity statistics — a line the NBER literature on the productivity paradox has been unpacking ever since.

The 2026 version is tighter and faster. US Census Bureau survey work found firm-level AI use climbing from low single digits in 2023 to roughly 9% of firms by late 2024, and the Business Trends and Outlook Survey has kept tracking the curve upward since.

Meanwhile US labour productivity growth has run in the low single digits — respectable, not transformational, and well inside the range you would expect from a normal cyclical recovery, per Bureau of Labor Statistics series.

So the question is not whether there is a gap. It is which of several unglamorous mechanisms is producing it.

Measurement Lag: We Count the Wrong Things

Productivity is output divided by hours. Both halves of that fraction are hard to observe in knowledge work, and AI makes them harder.

If a support agent resolves the same number of tickets but each reply is clearer, output measurement records nothing. If a developer ships the same number of features with fewer defects, the defects that never happened do not appear in any denominator. National accounts are built to count widgets and billable hours, not avoided rework.

There’s a subtler problem: much of what AI produces is free. Generated images, drafts and summaries that would previously have been purchased or skipped entirely show up as consumer surplus, and GDP-based statistics famously undercount consumer surplus, as the Brookings work on measurement and intangibles has argued.

Roche’s essay makes a version of this point from inside a company rather than inside an econometrics paper: most firms deploying assistants never established a baseline, so they have no way to know what changed.

Diffusion Takes Years, Not Quarters

The second explanation is the one economists find most boring and most convincing. General-purpose technologies pay out slowly because the technology is the cheap part and the reorganisation around it is the expensive part.

Paul David’s canonical study of electrification showed factories took roughly 40 years to reap the full productivity benefit of electric motors — not because motors were bad, but because the payoff required abandoning the central-shaft factory layout and rebuilding the floor plan entirely.

Erik Brynjolfsson and colleagues formalised this as the “J-curve” of general purpose technologies: measured productivity falls first, because firms are pouring resources into intangible complementary investment — training, process redesign, data plumbing — that the statistics treat as cost rather than capital formation.

If that model holds, we are currently in the dip. The trough is not evidence of failure; it is the price of the reorganisation.

Task Automation Is Not Workflow Automation

Here is the mechanism that most directly explains why individual enthusiasm doesn’t aggregate into organisational output.

Assistants are extremely good at bounded tasks: draft this, summarise that, refactor this function, name these variables. Almost every credible study measures exactly that kind of task. The Brynjolfsson, Li and Raymond study of a customer support deployment found roughly a 14% average increase in issues resolved per hour, concentrated among less experienced workers.

But a workflow is a chain of tasks with handoffs, approvals and queues between them. Speeding up one link in a chain moves the bottleneck; it does not necessarily move the throughput. Anyone who has drafted a document in ninety seconds and then waited nine days for legal review knows this in their bones.

Amdahl’s law is a useful mental model here. If AI touches 30% of the work and makes that portion twice as fast, the end-to-end gain is about 18% at best — and that’s before the coordination costs of the new step.

The gains show up in the statistics only when the queue between the links is redesigned, and queue redesign is a management problem, not a model problem.

The Quality Offset: Busier, Not Better

The fourth mechanism is the least comfortable one. Some of the output AI enables is work that nobody needed.

When drafting becomes nearly free, the marginal cost of producing a document collapses — so more documents get produced. More documents means more reading, more review, more meetings about the documents. The volume of artefacts rises while the volume of decisions stays flat.

Open-source maintainers have been unusually blunt about this. Several major projects reported a surge in low-quality, AI-assisted submissions that consumed more reviewer time than they saved, with the curl project’s experience with AI-generated security reports becoming the canonical example of generated volume imposing a review tax.

This is a genuine negative externality: the producer captures the speed gain and the reviewer absorbs the cost. In an aggregate measure the two cancel, and the statistics record nothing.

Who’s Actually Faster — and Who Only Feels Faster

The most instructive recent finding is one that cuts against the tools’ own users.

A 2025 randomised controlled trial by METR on experienced open-source developers working in their own large repositories found that participants were roughly 19% slower when using AI assistance — while believing they had been about 20% faster. The METR study writeup is careful about its limits, but that perception gap is the single most important number in this debate.

The pattern across studies is consistent rather than contradictory. Novices and people working outside their expertise gain the most. Experts in familiar, complex codebases gain the least and sometimes lose, because verifying a plausible-looking suggestion costs more than writing the line yourself.

Self-reported productivity is therefore close to useless as evidence. Everyone feels faster, because the friction of the blank page disappears and the friction of review is diffuse and unmemorable.

The policy implication is uncomfortable for vendors. The people most likely to renew a licence are the ones who gained the least, because the speed they felt was real to them even when the throughput was not. The people most likely to churn are the experts who quietly lost time and noticed. Adoption metrics and value metrics point in different directions, and most dashboards only show the first.

If your organisation’s AI ROI case rests on a survey asking employees how much time they saved, you do not have evidence. You have a mood.

What Would Actually Close the Gap

Strip out the extremes and a moderate picture emerges from the credible studies. The strongest gains cluster in tasks that are text-heavy, low-stakes and easily verified. The weakest cluster in tasks that are context-heavy, high-stakes and expensive to verify — which describes most of what senior people are paid for.

Adoption itself is lopsided. Anthropic’s usage analysis found delegation-style use growing relative to collaborative use, and the Anthropic Economic Index shows adoption concentrating heavily in software and technical writing rather than spreading evenly across the economy.

The widely cited finding that a large majority of enterprise generative AI pilots produced no measurable P&L impact should therefore be read as a claim about pilots, not about the technology. Pilots without process change rarely move a P&L, whatever the technology. For a broader read on how this pattern plays out in employment rather than output, see The AI Prism’s look at what’s actually happening to jobs.

The pattern repeats across every general-purpose technology before this one. Electricity, the computer, the internet, each spent years as a disappointment in the aggregate statistics before the reorganisation caught up. The mistake was never to expect gains; it was to expect them on the deployment timeline rather than the absorption timeline. AI is behaving exactly as its predecessors did, which is the most reassuring and most ignored fact in the whole debate.

None of this means the tools are useless. It means the gains are conditional, and the conditions are organisational rather than algorithmic. That is good news, because organisations can be changed faster than physics.

None of the fixes are technical, which is precisely why they are slow.

Measure before you deploy. A two-week baseline of cycle time, defect rate and rework volume on one workflow is worth more than a year of adoption dashboards. Seat counts and token spend measure input, not output.

Pick one end-to-end workflow rather than sprinkling assistants across ten. Gains are only visible when the whole chain — including the approval steps and the handoffs — is redesigned around the new capability.

Then account for the review tax explicitly. If generated volume increases, reviewer capacity has to increase or throughput standards have to tighten; otherwise the saving quietly migrates from producer to reviewer and disappears.

Finally, be patient with the macro data. Electrification took decades and computing took roughly twenty years to register clearly. Expecting a general-purpose technology deployed at scale from 2023 to show up in national statistics by 2026 was never a reasonable timeline.

The Management Question Nobody Asks

Underneath all the measurement debate sits a simpler, more awkward possibility: the gap persists because most organisations never treated AI adoption as a change-management project in the first place.

Rolling out assistants is treated as a software rollout — buy licences, send a training email, watch the dashboard. Reorganising work around a new capability is a different activity entirely, and it is the one the productivity literature says actually moves the number. The tool arrives; the workflow does not.

This reframes the paradox in a useful way. The lag is not only a statistical artefact or a diffusion delay. Some of it is simply the ordinary cost of deploying a general-purpose technology without the complementary investment the theory predicts. The good news is that this is the one mechanism on the list a manager can actually do something about on a Tuesday.

The Bottom Line

The productivity gap is not proof that AI doesn’t work, and it isn’t proof that the statistics are broken. It is what the middle of an adoption curve looks like when the tooling has outrun the organisational plumbing around it — task-level speed with workflow-level friction, real gains for novices, unproven gains for experts, and a measurement apparatus that was never designed to see any of it.

The interesting question isn’t when the numbers will finally arrive. It’s whether the organisations currently counting seats and licences will have built the baselines they need to recognise the gains when they do — or whether they’ll still be measuring adoption and calling it impact?

References

The post The AI Productivity Paradox: Why the Gains Haven’t Shown Up Yet appeared first on The AI Prism.

The AI Debt Binge: $1.65T of Hidden Borrowing and a Demand Bubble

The AI Prism Admin — Fri, 07 Aug 2026 10:00:00 +0000

Everyone’s counting the revenue. Nobody’s counting the debt.

The artificial intelligence sector has spent two years training the public to watch one number: how much money the models earn. Quarterly earnings calls lead with cloud growth, and headlines celebrate record funding rounds. Yet a different ledger is growing quietly in the background, and almost no one in the mainstream conversation measures it. The companies building the AI infrastructure are borrowing on a scale that has no precedent in the history of the industry, and they are doing it through channels that don’t show up cleanly in the headlines.

Two recent reports frame the problem from opposite ends. A Fortune investigation places the total of hidden AI-related borrowing at roughly $1.65 trillion, spread across hyperscaler bond issuance and structured financing that sits off the obvious balance-sheet lines [Fortune, 2026]. A separate analysis at wheresyoured.at argues the demand side is itself a bubble, with revenue that depends on the same firms renting capacity back to one another [wheresyoured.at]. Put the two together and a single question sits underneath the entire AI trade: what happens when the borrowing stops being cheap and the demand turns out to be circular?

The $1.65T Number

The headline figure is large enough that it is easy to dismiss as a rounding error of the cloud era. It isn’t. The $1.65 trillion estimate aggregates the debt raised by the largest technology companies and the financial vehicles they use to fund data-center construction, GPU purchases, and power agreements [Fortune, 2026]. Much of it is denominated in investment-grade bonds, which is precisely why it draws little alarm: the issuers are rated well enough that the market treats the borrowing as safe.

What makes the number useful is less its precision than its direction. Capital spending by the hyperscalers has climbed from a meaningful line item to the single largest use of cash on their books. When a company’s capex grows faster than its operating income for several consecutive years, the gap has to be filled somewhere, and equity investors rarely fund that alone.

Debt fills it. The firms have issued bonds at a pace that would have been unthinkable for a software business a decade ago, because the asset they are buying — compute — is treated as a long-lived, defensible moat rather than a depreciating expense [Fortune, 2026]. The bet is that the revenue arrives before the interest comes due.

How the Borrowing Hides

The reason this debt escaped scrutiny for so long is that a meaningful share of it never appears where an ordinary reader looks. Traditional balance sheets capture bonds and bank loans, but the AI buildout leans on structures that sit a step away from the parent company: special-purpose vehicles, sale-leaseback arrangements, and power-purchase agreements that move the obligation to a financing partner [Fortune, 2026]. The capacity still gets used by the same firm, but the liability lives elsewhere.

This is not fraud. These are legal, long-established financing techniques, and they are used across every capital-intensive industry. The distinction matters because the techniques are opaque by design. An analyst who reads only the headline debt figure sees a healthy balance sheet; an analyst who traces the lease obligations and the off-balance-sheet vehicles sees a very different picture of leverage.

The opacity compounds the risk. When the true scale of borrowing is hard to measure, the market cannot price it correctly, and when the market cannot price it, the first sign of strain arrives as a surprise rather than as a gradual repricing. Surprises are what turn a manageable correction into a forced one.

Capex vs Revenue

The cleanest way to see the tension is to put the spending next to the income. The hyperscalers are guiding capital expenditure upward by sums that dwarf the incremental revenue those investments are expected to produce in the near term. The gap between the two is exactly the slice that has to be financed, and financing means borrowing [Fortune, 2026].

Defenders point out that cloud infrastructure has always been built ahead of demand. The difference now is the slope. The spending is not a gentle curve that smooths out as customers arrive; it is a near-vertical line justified by the assumption that AI workloads will absorb every dollar of new capacity. That assumption deserves scrutiny rather than deference.

Revenue, meanwhile, is real but uneven. The firms report strong cloud growth, yet the portion of that growth that traces directly to generative AI remains a smaller fraction than the capex suggests it should be. Until the two lines converge, the financing gap is a standing liability that accrues interest every single day it remains open.

The Demand Question

This is where the second report lands. The wheresyoured.at analysis argues that the demand we celebrate is thinner and more concentrated than the headlines imply [wheresyoured.at]. Inference traffic is growing, but the customers paying for it are themselves a small set of well-funded incumbents, and a large share of the usage is the labs and platforms consuming their own output to train the next model.

A market where the buyers and the sellers are the same handful of companies is not necessarily fake, but it is fragile. Real demand is measured by entities that could walk away; demand that is captive to the firms doing the building cannot be relied on to persist if the financing environment tightens. The question is not whether anyone uses the models, but whether enough independent buyers use them at prices that justify the buildout.

The optimistic case says enterprise adoption is early and will compound. The cautious case says we are watching a self-referential loop where each new data center is justified by the revenue from the previous one. The truth is probably between the two, and the debt does not care which story wins — it comes due either way.

Circular Revenue Worries

The circularity is the part that should make a careful reader pause. A model lab rents GPUs from a cloud provider, builds a product, sells access to developers, and some of that developer activity flows back to the same cloud provider as inference spend. The dollars move in a loop, and at each step a revenue figure is recorded [wheresyoured.at]. Growth that emerges from a closed loop looks identical to growth that emerges from a genuine market until the loop is stressed.

None of this means the products are worthless. Developers are shipping real software on top of these models, and enterprises are finding genuine use cases. The concern is one of proportion: if a meaningful share of the recorded revenue is simply the same money circulating among a small group of giants, then the multiple the market assigns to that revenue is built on a base that is smaller than it appears.

Circular revenue is tolerable when it is a small fraction of the total. It becomes a problem when the entire financing case depends on the total continuing to grow, because the loop has no external force pushing it outward once the participants have saturated their own needs. At that point the growth has to come from someone new, and new buyers are exactly what the cautious reports say are missing.

Who’s Exposed

If the borrowing and the demand are both overstated, the exposure is not limited to the firms doing the building. The lenders who bought the bonds are exposed to the credit risk, and a wave of downgrades would land on insurance companies, pension funds, and asset managers that hold the paper as safe [Fortune, 2026]. The utilities that signed power agreements are exposed to cancellation risk if projects stall.

The AI labs themselves sit in the most precarious spot. They are the smallest balance sheets against the largest ambitions, and many of them depend on the hyperscalers both for compute and for the cloud credits that show up as revenue. A tightening in one relationship propagates quickly through the others, because the ecosystem is more interconnected than its separate branding suggests.

Sovereign wealth funds and other large allocators that poured capital into the theme are exposed as well, though their size gives them patience that a levered startup does not have. The point is that this is not a contained trade. The borrowing was syndicated across the global financial system, which means the bill, if it comes, is shared broadly rather than borne by a single careless actor.

For readers tracking how this could resolve, see The AI Prism’s take on the post-bubble landscape, which maps which parts of the stack are likely to survive a repricing.

Historical Parallels

The pattern has a precedent that is uncomfortable to revisit. In the late 1990s and early 2000s, telecommunications carriers borrowed enormous sums to lay fiber and build network capacity, convinced that internet traffic would grow without bound. The traffic did grow, but not fast enough to service the debt, and the resulting defaults reshaped the industry [Federal Reserve, macro context]. The assets were real; the timing of the payoff was wrong.

AI infrastructure is not telecom, and the firms involved are far more profitable than the carriers ever were. But the structural similarity is the part worth holding onto: when capacity is built on borrowed money against a demand curve that is assumed rather than proven, the discipline is supplied by the credit market, and the credit market is patient only until it isn’t. History suggests the turn is sudden, not gradual.

Another parallel sits closer to the present. The 2021–2022 correction in speculative technology showed how quickly capital that was abundant becomes scarce, and how valuations that looked durable were propped up by a cost of money that changed. The AI debt load is being issued in a different rate environment than the easy-money era, which cuts both ways: the borrowing is more expensive, but the caution is also more warranted.

What a Repricing Would Actually Look Like

It helps to be concrete about the failure mode, because abstraction invites complacency. A repricing does not require a dramatic default. It can begin with a single rating agency placing a cloud provider’s off-balance-sheet vehicle on negative watch, which raises the cost of the next bond, which narrows the spread between borrowing and returns, which quietly slows the next build.

From there the feedback is gentle until it isn’t. The firms most exposed are the ones that borrowed against the most optimistic demand curve; a small downward revision in expected inference growth can turn a comfortable coverage ratio into a strained one. The debt does not need to become unpayable for the financing environment to tighten, and a tighter environment is exactly what stalls the next wave of capacity.

The safeguard is not optimism but optionality. Firms that can slow spending without stranding assets, that have real external demand, and that financed with maturities matched to hardware life will absorb a repricing. The rest will discover that the $1.65 trillion was less a war chest than a timer.

The Bottom Line

The revenue story is real, and no honest account of the AI sector can dismiss the genuine productivity the models have unlocked. The debt story is also real, and it is the one almost nobody is counting. A buildout financed by $1.65 trillion of borrowing, much of it hidden in structures that sit a step away from the balance sheet, only makes sense if the demand underneath it is broader and more independent than the cautious reports allow [Fortune, 2026] [wheresyoured.at]. The two narratives cannot both be comfortably true at the same time, so which one gives first when the credit window narrows?

References

The post The AI Debt Binge: $1.65T of Hidden Borrowing and a Demand Bubble appeared first on The AI Prism.