All writings

On AI short-termism · Prologue

Everyone is behaving sensibly. That is the problem.

Most firms use AI, few have seen a gain yet, and the gains will probably come. The prologue to a series in 4 movements asks what they are being paid with.
The argument in brief: each decision about AI is locally rational, and together they make the system fragile. This prologue sets out the question, 6 terms and the road ahead.

By Louay Belkhamsa · · 5 minproofHow to read
  1. Prologue
  2. I
  3. II
  4. III
  5. IV
  6. Close
fast sand, slow shell
A moment of sand, an age of shell

Here is a finding that should bother you more than it seems to.

In a study published in March 2026, central banks and a university surveyed nearly 6,000 senior executives in the U.S., the U.K., Germany and Australia. About 70% said their firms actively use AI. Yet about 9 in 10 reported that it had made no difference to their productivity or their headcount over the past 3 years.1

On its own, that is not alarming. New technologies rarely pay off straight away: firms first have to learn how to use them, and the gains tend to arrive later. The same executives expect AI to raise their productivity by about 1.4% over the next 3 years, a modest forecast that sits within the range independent economists project.1,2 Nobody here is wildly overpromising.

So the gains will probably come following the J-curve. This series asks a different question:
is money the only price the companies are paying for AI ?

Fig. 2 · the scale

How long a task AI can handle, measured in expert working time1×

2022
2026

What $1 of AI buys1×

2022
2024

Fig. 2. In 2022, each bar was the whole scale.
Fig. 2. Work: METR’s 50% time horizon (2026: public models, February–March, about 12 hours, range 5 to 61). Price: AI Index 2025, November 2022 to October 2024. Each chart has its own scale.
METR, 2025–2026 · Stanford HAI, AI Index 2025

© 2026 Spiresen

AI handles routine work well. What it cannot supply is the experienced judgement that checks its output, catches mistakes and decide on how to proceed. This process is non-negotiable given governance and accountability requirements. That judgement takes years to build, and it is mostly built in entry-level jobs, where employees are supervised by seniors and the blast radius is limited.

Those jobs are where the cuts are landing. Since generative AI spread, employment of 22-to-25-year-olds in the occupations most exposed to it has fallen 19% relative to their peers.3 Worringly, new and smaller companies are not taking up this workforce: start-ups built around AI are smaller than their peers, and entry-level workers make up a share of their staff about 15% lower.4

Surprisingly, each of these decisions makes sense on its own. The AI that handles routine work keeps getting cheaper,5 and the bill for missing experience will not arrive this quarter. Added together, though, these choices have a cost that doesn’t show up in any budget. The savings come partly from not training the people who would become tomorrow’s experienced staff, the ones a company relies on when something unexpected happens. And nobody, inside or outside the company, is rewarded for training them instead.

That is the argument of the title. Almost everyone involved is acting sensibly to what is required of them, and the sum of those sensible decisions is a system that is quietly getting more fragile. In two phrases:

Each decision is locally rational;

the system as a whole is not robust.

  1. i · observation~70% use AI~9 in 10 see no effect yet
  2. ii · questionWhat are the gains paid with?Not whether they will come
  3. iii · mechanismEntry-level jobs cutWhere experience is built
  4. iv · thesisEach decision: locally rationalThe system: not robust
  5. v · exposureDoing right is not enoughOthers’ choices set your exposure

Risk exposure: when being right isn’t enough

In the 2008 financial crisis, most companies never touched a subprime mortgage. They were exposed to the fallout all the same, because they relied on the same credit markets. Something similar applies here, on a far less dramatic scale and with one difference: there will be no single crash (not for the reasons we’re describing). The shared market this time is the labour market, and the damage arrives one company at a time.

Some companies already see this coming and keep hiring and training juniors. But good practice does not shield an organisation from what its competitors do; it only changes the angle of its exposure.

Take company A, which keeps training and pays for it now. Its competitors don’t. When conditions get harder, they turn to the labour market to strengthen their teams. By then, experienced people are scarce, so company A’s people are the ones everyone tries to hire. One company’s shortage becomes another company’s raid. Your exposure depends on the choices of other companies as much as on your own.

So the question this series leaves you with is not whether your company is exposed. It is how, and when.

Six ideas to carry with you

The series uses the following words precisely. Here they are in plain terms.

Locally rational

Not a figure of speech: this is Herbert Simon’s bounded rationality, real economics on how people and firms actually decide. Nobel in economics, 1978, and a field moving fast again today.

A choice is locally rational when it makes sense from where you sit. At a concert, one person stands up to see better. The person behind them has to stand too. Soon the whole crowd is on its feet, nobody sees any better, and everyone’s legs ache. Every choice was sensible. Nobody planned the outcome.

Fig. 3 · the concert

Each person asks: should I stand to see better?

Can see the stage

5 of 5

Sitting comfortably

5 of 5

Everyone sees the stage, and everyone is comfortable.

Fig. 3. Locally rational: each step makes sense to the person taking it. The meters show the sum: the same view as at the start, and nobody comfortable.
Fig. 3. Sightlines computed from the drawing: a view is blocked when the head in front crosses the line from the eye to the stage. Start: 5 of 5 see, 5 of 5 seated. End: 5 of 5 see, 0 of 5 seated.

© 2026 Spiresen

Short-termism

Favouring a reward that arrives now over a cost that arrives later, usually on someone else’s desk. In AI, looking fast pays off immediately. The price of not being ready comes due afterwards.

The readiness gap

The distance between how widely an organisation uses AI and how prepared it is, in its data, its people, its systems and the way it is organised, to make that use pay off. Movement 1 measures it.

how widely AI is usedhow prepared the organisation isdatapeoplesystemsorganisationthe readiness gap

A sketch, not data. Movement 1 puts numbers to it.

Formation

How people become experienced: years of supervised work, usually starting at entry level. A company can buy AI by the month. It cannot buy formation quickly, at any price.

Fragility

Risk is the chance that something goes wrong. Fragility is how badly it goes when it does. A fragile system looks healthy while conditions are kind and breaks hard when they are not, because the slack that could have absorbed the shock was spent earlier. A company can grow its revenue and become more fragile at the same time.

Fig. 4 · fragility
HarmlittlebreakseasyharshConditionsrobust: keeps a reservefragile: reserve spent

Fig. 4. Same risk, different fragility. Drag along the Conditions axis, from easy to harsh.
Fig. 4. A sketch of the idea: illustrative curves, not measurements. Robust: harm rises gently with the shock. Fragile: lower while conditions are easy (the reserve went into performance), then rises steeply past a threshold.

© 2026 Spiresen

The risk register

A running list of what could go wrong, who is watching it, and whether anyone is. Each movement adds an entry. By the end, the entries become one document.

No.What could go wrongWho is watchingIs anyone?
Ientered in Movement 1··
IIentered in Movement 2··
IIIentered in Movement 3··
IVentered in Movement 4··

What this series is not

It is not an argument against AI, and it does not predict that AI will fail to deliver. It assumes the gains will come and asks what they cost. Nor does it accuse anyone of bad faith. It is an argument about incentives: what gets rewarded, what stays invisible, and what that does over time.

It is written for anyone who touches these decisions or lives with their consequences: board members, engineers, investors, and people who simply use these tools at work. Each movement can be read on its own, but they build on one another.

The road ahead

  1. Prologue

    Everyone is behaving sensibly · you are here The argument in brief.

  2. I

    Leadership & the readiness gap Why the market rewards speed over preparation.

    1. II

      The ledger The technical and organisational cost of shortcuts that nobody is tracking.

    2. III

      The pipeline How the route to experience is narrowing inside companies.

    in
    parallel
  1. IV

    The missing correction Why no market force will repair it, and why no company is safe from it, even one that does everything right.

  2. Close

    The risk register, compiled Where the current path leads by 2030, and what to do about it.

The series will go over multiple layers of the organization, their debt situation and risks. Any claim or vendor source cited is cross-referenced with independent data, with conflict of interests cited if needed.

We begin where the decisions are made, with the gap itself.

Sources (5)

  1. Yotzov et al., “Firm Data on AI,” NBER Working Paper 34836 / Federal Reserve Bank of Atlanta Working Paper 2026-3, March 2026. Surveys run by the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University. Shares are of all surveyed firms, including those not using AI: 89% reported no effect on productivity and more than 90% none on employment. ↩
  2. Filippucci, Gal, Laengle, Schief and Unsal (OECD), “Opportunities and Risks of Artificial Intelligence for Productivity,” International Productivity Monitor 48, 2025. ↩
  3. Brynjolfsson, Chandar and Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, revised November 2025. Based on U.S. payroll records; the decline is relative to less-exposed workers and controls for firm-level shocks. ↩
  4. Kim and Koning, “AI-Native Firms,” INSEAD / Harvard Business School working paper, June 2026. Compared with non-AI start-ups in the same industry and founding cohort. ↩
  5. Stanford HAI, AI Index Report 2025. The cost of running AI at a fixed level of capability (that of GPT-3.5) fell more than 280-fold between November 2022 and October 2024. ↩

Next · Movement 1

Leadership & the readiness gap

All writings

Ask the author

What is it?

This goes to the author only. Nothing you write here is shown to other readers. An email, if you give one, is used to reply and is deleted after two years.