TwinLadder Weekly

Issue #43 — The BIS Reaches for the Canal Mania

4 July 2026 · Weekly intelligence on judgment, governance, and the boards accountable for both

Editor's Note

From Alex —

On 28 June the Bank for International Settlements published its Annual Economic Report. In the passage describing the AI capital-expenditure boom it reached for three historical parallels at once: canal mania, railway mania, the dotcom boom.

That is the central bank of central banks putting three bubbles in one paragraph, in the document its own members read line by line.

The figures attached to it are these. The five largest hyperscalers are set to spend over one trillion US dollars on AI-related capital expenditure across 2025 and 2026. Those commitments run ahead of what the same firms earn and ahead of the free cash flow they generate, and some are issuing debt to cover the difference. The BIS warns that disappointing returns could turn the boom into a protracted investment bust with effects across the wider economy.

Two easy readings are available and I take neither. The first is that a crash is coming and boards should stop spending. The second is that central banks say this sort of thing every June and the report can go in the folder.

Here is the reading I take. Every board approving an AI budget this year is underwriting a return. Somebody in that room has made an assumption about what the technology will do inside the work, over what period, at what unit cost. The BIS has now put on the record that at the aggregate level the returns case is unproven. Which makes each board's own case something it has to argue in its own room, on its own evidence.

So the question this issue puts on the table is narrow, and it holds whether or not the BIS turns out to be right about the cycle. Who in your boardroom is equipped to interrogate the return assumption underneath the AI line?

Liga has the analysis.

— Alex

What the BIS Put in Its Annual Report

The report was published on 28 June. Its treatment of AI capital expenditure covers three things: the size of the commitment, the way it is being financed, and the historical company it keeps.

On the financing, the BIS is direct. "These commitments," it writes of the hyperscalers' AI capital expenditure, "are outpacing earnings and the free cash flow of these firms, leading some to issue debt to raise additional financing." The spending has passed the point where it is funded out of what the businesses throw off. It also notes that implied long-term earnings growth for the largest corporations sits well above historical benchmarks — which is to say that the price of those shares already contains a forecast about how the AI investment ends.

Then the parallels. Canal mania, railway mania, the dotcom boom, named explicitly and in sequence. The warning attached to them is that disappointing AI returns could convert a capital-expenditure boom into a protracted investment bust with recessionary effect across the economy.

Be precise about what that is. The BIS is describing a financing condition at the level of five firms and the market that prices them. Whether AI pays inside a mid-size European insurer, a regional bank or a manufacturer is a separate question, settled in that company, on that company's evidence. A supervisor's macro warning leaves it open.

What the warning removes is a comfort. Until this report, a board could treat the returns case as something the market had already settled — validated by the scale of the spending itself, by the conviction of firms holding more information than anyone in the room. The BIS has said in public that the aggregate case is unproven and that the largest spenders are borrowing to keep spending. The assumption a board underwrites is no longer inherited. It has to be made.

Three Instruments Got Simpler in the Same Week

The same seven days settled a good deal of the European legal frame around AI, in one direction.

On 29 June the Council of the EU gave final adoption to the amendments to the AI Act, completing the legislative procedure. From that date, the AI Act a policy document cites is the 2024 text as amended, and any internal paper still quoting the original dates describes an instrument that has moved.

On 30 June the data strand went the other way. The Cyprus Presidency withdrew its compromise text on the GDPR side of the Digital Omnibus from the COREPER II approval process, having established that it did not command a qualified majority, and the Presidency passed to Ireland on 1 July with no Council position agreed. Privacy Next's account of the negotiations records that the Council "has made significant progress in recent months, but it has not yet found a position capable of securing Member State support." So the AI rules were simplified and the data-protection rules stayed put, which leaves the constraints on training data and deployment where they were in January.

On 3 July the Commission adopted the delegated act revising the European Sustainability Reporting Standards. Mandatory datapoints fall by more than 60%, total datapoints by more than 70%, with expected reporting-cost savings above 30% per company. The stated goal is to simplify and streamline the previous standards, following through on the changes the Omnibus I package made to the scope of the CSRD. The revised standards enter into force on 20 November 2026 and apply to financial years beginning on or after 1 January 2027.

Three instruments moved in one week. Each governs how a company reports, how a system is classified, or what a deployer must be able to show. None of the three asks what return the company expects on the capital it is committing to AI, or on what evidence. The reporting burden came down. The capital commitment stayed. Whatever discipline attaches to the investment case has to be built inside the building.

The Underwriter's Question

One habit of mind shows up whenever a board is asked to act on evidence it regards as incomplete. The instinct is to wait for the study — to hold the position until somebody produces the number that settles it. In a laboratory that instinct is a virtue. At a board table it is a category error about the job.

A board occupies the underwriter's position, not the scientist's. The scientist asks whether a claim has been proven. The underwriter asks three other questions: what is our exposure, how would we know it was forming, and what do we provision against it before the loss event. A director who imports the laboratory's standard into the boardroom has mistaken which chair they are sitting in.

Turn those questions on the AI capital line and they read like this. What did we assume the technology would do inside the work? What observation would tell us the assumption was wrong? By what date would it arrive, and who is obliged to bring it?

An investment case that cannot answer the second question is a mood with a spreadsheet attached. And the second question is answerable. Three readings can be commissioned this quarter from data the institution already generates. The first is the share of routine, formative cases that junior staff still work unaided, against the share arriving pre-drafted or pre-flagged by a model. The second is the override rate: how often the humans in a function challenge or reverse what the system produced. The third is decision quality in the functions most saturated by AI, set against comparable functions riding the same conditions, so that a general improvement in the business does not get credited to the tool. Each is a proxy. Each moves before the outcome does, and each lives in work-allocation systems and audit logs already running.

Two things from the same week show why the assumptions underneath a business case need re-reading each quarter.

The first is a price. On 30 June Anthropic released Claude Sonnet 5 at $2 and $10 per million tokens, with a one-million-token context window. Further down the release announcement is a sentence that moves every budget built on the headline: "Claude Sonnet 5 also uses a new tokenizer that produces approximately 30% more tokens for the same text." The same contract, the same claims file, the same set of board papers, now consumes roughly a third more tokens to process. Unit price and unit cost have come apart. Anyone budgeting AI from the price card is budgeting a number that no longer describes their bill.

The second is a matter of sight. On 25 June the General Board of the European Systemic Risk Board adopted a formal warning on the systemic cyber risks created by frontier AI models. Inside a document about attack capability sits an observation about self-knowledge: "establishing and maintaining visibility of AI deployment within any financial institution – including fully understanding where and how AI is deployed – is challenging, taking into account that it is used in customer interfaces, internal agents, business processes, and integration with third parties." A European systemic-risk body, writing to supervised institutions, records that a bank finds it hard to see where its own AI is running.

Hold those two beside the return question and the difficulty becomes concrete. The benefit is claimed on a population of uses that nobody holds a complete list of, priced from a card that has stopped matching the invoice. That is the state of the evidence a board is being asked to underwrite.

Then there is the fallback. When the return fails to arrive, the standard plan is to hire the capability in. At the level of one firm that works. At the level of the system the arithmetic fails to close: lateral hiring reallocates a stock of senior judgment formed by training decisions taken a decade ago. It does not create one. If every institution defers formation at the same time — which is what a tool everyone adopts at once encourages — the market becomes a bidding contest for a fixed number of chairs.

Who in the Room Can Interrogate It

On 25 June Deloitte published an analysis of the last six leadership roles held by every director on a Fortune 100 board. CEO experience is present on 100% of those boards. Finance executives are on 96. Operations on 89. Marketing on 67. Senior HR on 56 or more. Nonprofit, military or government backgrounds on 65. The study is framed as a baseline for refreshment discussions, and it offers no equivalent count for technology or AI leadership. "Having broader functional experience in the boardroom," the authors write, "could help enterprises adapt more effectively to shifting market conditions."

Picture what that composition means when the AI paper reaches the table. The former group finance director takes a discount rate apart at sight, because she has spent thirty years doing it and would notice a terminal-growth assumption carrying too much of the case. The former chief executive knows what a market-size claim looks like when it has been reverse-engineered from a target. Between them they can dismantle almost any proposal that arrives in the room.

The AI proposal is a different animal. Its load-bearing claim is about what the technology does inside the work — how much of a claims adjuster's judgment the model carries, how much of a credit memo it drafts to a standard, where it is confidently wrong and who would catch that. Testing that claim takes someone who has shipped a model into a live process and watched what it did to the people around it.

Two further findings from the same fortnight fill in the picture.

On 27 June, Pearl Meyer reported on roughly 2,500 public-company proxy statements filed in 2026. About 2% include a formal AI metric in executive incentive programmes. Of that 2%, only 12% use an explicit AI metric; 60% fold AI into a broader technology or transformation objective and 28% handle it through individual performance assessment. Seventy-seven per cent of respondents report they have not yet scaled AI enterprise-wide. The question the analysis puts to compensation committees is "not simply whether to incorporate AI into incentive plans, but whether organizations are prepared to measure and govern it in a way that supports sound pay decisions."

Read that against the capital. The AI line is among the largest new discretionary commitments many of these companies have made in a decade, and in ninety-eight proxies out of a hundred nobody's pay depends on what it returns.

On 30 June, EY published a board effectiveness study drawing on more than 100 directors and interviews with veteran board leaders. Directors identify board composition as an area of weakness. They report that the quality and flow of information from management is uneven. And they want more time on AI, talent and geopolitics — with the better boards, the study finds, restructuring how information reaches them "instead of just adding hours".

Those three findings describe one condition from three angles. The people in the room were formed in finance, operations and general management. The information reaching them about AI arrives through a channel management controls. The pay of the executives bringing it is, in almost every case, unaffected by whether the investment works. And the directors themselves, asked what their board's weakness is, name composition.

Which leaves the question for the next meeting. Our AI capital allocation assumes a return. Who in this room can interrogate that assumption?

The Line That Does Not Appear in the Capital Budget

A second decision sits folded inside the first, and it never reaches the capital paper.

Every hour the tool frees arrives unlabelled. Spent one way, it is booked as saving, and the saving is real and appears this year. Spent the other way, it is reinvested: harder cases moved to hands that are still forming, a senior kept beside them, feedback engineered into work that would not otherwise supply any. The saving shows up in a number that already exists. The reinvestment shows up in a bench that nothing measures until the day something depends on it.

The saving also has a ceiling. Every competitor can buy the same tool from the same vendor and book the same reduction, which moves the whole field to a new floor and leaves the relative position where it was. What survives as advantage is whatever the freed hours were spent building.

The economics of spending them on formation are old and unfriendly. The capacity to tell a sound piece of work from a fluent one travels with the person who holds it. A firm that funds the years in which that capacity is built cannot fully capture the return, because a rival can hire the formed professional the year they mature and pay nothing toward the forming. Left to the ordinary incentives, the bench thins by the logic of the market, without anyone making a mistake.

Management is measured on one to three years. The erosion runs on ten. One body in the building holds a mandate longer than every career reporting into it, and it exists to carry costs that fall outside everyone else's horizon. So the board approving the AI capital line and the board that would approve a formation line are the same board making one decision twice. Half of it appears in the budget pack.

The BIS wrote about a financing condition among five firms. The version that reaches a European board is smaller, closer and more answerable. Capital is being committed on an assumption about what the technology does inside the work. The people who could test that assumption in 2036 are the juniors whose formative work the technology is absorbing this year.

What This Means for Boards Right Now

One. Put the return assumption in writing, and put the falsification condition next to it. A single page: what we assumed the technology would do inside which processes, over what period, at what unit cost, and what observation would tell us we were wrong. Then commission the three readings that move early — the share of formative work still done unaided, the override rate, and decision quality in AI-heavy functions against comparable ones. A case that cannot name what would disprove it has not been assessed.

Two. Ask for the deployment inventory before you ask for the returns report. The ESRB has told supervised institutions that seeing where AI runs inside their own house is hard, across customer interfaces, internal agents, business processes and third-party integrations. A benefit claimed on uses nobody has enumerated cannot be audited. And the unit economics under those uses are moving in ways the price card does not show: a model released on 30 June adds roughly 30% more tokens to the same text.

Three. Treat the answer to the composition question as a finding with a date on it. Name the person at your table who has put a model into a live process and watched what it did to the work. If the seat is empty, that belongs in the refreshment discussion alongside the finance and audit seats every board fills without argument. Directors themselves name composition as the weakness. Two per cent of proxies attach an AI metric to executive pay. Both numbers describe one missing capability, seen once in the boardroom and once in the incentive plan.

Reading List

The longer treatment of what happens when growth assumptions outrun the question nobody wants to ask, worked through a sector where the multiples are public: Legal AI Valuations: When the Numbers Stop Making Conventional Sense
Why so much enterprise AI spend produces no measurable P&L effect, and the organisational reading that explains more of it than the skills-gap story does: Why Your AI Pilot Is Stuck: It's Not the Skills Gap
What the entry-level work was actually buying, and what closes when it is automated without the learning function being replaced: The Broken Learning Ladder: AI Is Removing the Work That Built Expertise

What We Are Watching Next

Whether the Irish Presidency brings the data strand of the Digital Omnibus back to the Council, and on what text after the Cyprus withdrawal
Whether any European issuer discloses an AI return assumption or a payback period in its reporting, now that the revised ESRS have cut the mandatory datapoints by more than 60%
Whether the ESRB warning converts into a supervisory request for a complete AI deployment inventory from national competent authorities
Whether any board this season adds a director whose stated experience is deploying and governing AI systems, and which committee that director joins

The next issue goes deeper into one of these. If you want a specific function or sector covered, reply to this email.

— Liga

TwinLadder Weekly is a weekly intelligence report on judgment, governance, and the boards accountable for both. Subscribe at twinladder.ai/newsletter. Forward this issue freely.