In 2026, “token” names two different things: the form in which finance is moving value on-chain, and the unit in which machine intelligence is bought. BCG calls competition on the second token-based competition. For financial institutions the two are one story — because tokenized markets scale workload without limit, and token-priced intelligence is the only input that scales with it.
This paper argues one thing: in tokenized markets, competitive advantage will be decided by how productively an institution converts purchased intelligence into operational capacity. Three arguments carry the conclusion.
First, intelligence has stopped being scarce. BCG’s research frames the shift: when AI supplies expert-level knowledge on demand, advantage moves from having intelligence to applying it — and in its sample of 107 public technology companies, the heaviest users of AI tokens grew revenue at roughly three times the median rate of the lightest users.
Second, the two tokens meet in the back office. Tokenized assets generate around-the-clock, block-level workload that no staffing plan can track; token-priced intelligence is elastic to that workload in a way headcount can never be. The institutions that operate tokenized value will be the ones that buy intelligence well.
Third, the binding constraint is organizational, not technical. BCG finds fewer than one employee in ten delegates multi-step work to AI agents; the capacity is purchased and then left idle. The metric that disciplines all of this is return on intelligence — value produced per unit of combined human and machine cost — and it is modeled below on your own numbers.
For as long as finance has existed, the scarce input was expert judgment — and institutions competed by hoarding it: analyst classes, credentialed hierarchies, the war for talent. BCG’s June 2026 research argues that era is closing. As AI models reach expert-level competence across most knowledge domains, intelligence becomes abundant, and advantage shifts from possessing it to deploying it against business problems more productively than the next firm. Because that deployment is metered and billed in tokens, BCG names the new contest token-based competition — a deliberate echo of the time-based competition it introduced in 1988, when speed became the basis for winning.
The early evidence is directional but striking. Across 107 public technology companies with more than $500 million in revenue, BCG found the highest-usage quintile of AI-token consumption growing revenue at a 16.5% median year over year, against 5.1% for the lowest — an association observable first in software engineering, where token use can be measured consistently across firms. Association is not causation, and technology firms are the early cohort, not the whole economy. But finance executives have seen this movie: the firms that learned to buy computing well in the 1990s, and data well in the 2010s, did not wait for the causality paper.
What BCG frames for the whole economy lands with particular force on financial institutions, because finance is the industry where the other token lives. Stablecoins settle tens of trillions a year; tokenized funds and Treasuries compound quarterly; licensing regimes from MiCA to VARA now supervise the perimeter (The Autonomous Digital Asset Back Office prices this in full). Tokenized value produces operational workload — reconciliation, custody evidence, screening, reporting — 24 hours a day at block-level granularity. Headcount is produced by the hiring calendar. Only one input on the market scales with the workload itself: intelligence bought by the token.
The asset side of finance is tokenizing value. The cost side is tokenizing intelligence. Institutions that read these as one balance sheet will run tokenized markets at a cost and control standard the rest cannot match.
This is why treating “AI spend” as an IT line item is a category error. Token expenditure is capacity — the operational capacity to absorb tokenized-market workload without linear hiring. The right question is never “how do we cap it” but “where does a token of intelligence return the most,” which is a capital-allocation question, and capital allocation is the CFO’s home game.
BCG’s first prescription translates directly into finance’s own vocabulary: treat token spending as capital deployment, not utility consumption. A firm with capital does not only ask how to spend less; it asks where the next unit earns the most. Two properties make the token line behave like unusually good capital. AI-executed processes get faster and cheaper as underlying models improve — the same workflow, re-run next year, costs less and performs better without a change request. And machine capacity flexes with demand in both directions, which no other line in the operating budget does.
The measurement discipline matters as much as the allocation. Institutions that score AI purely on labor savings will systematically fund the narrowest use cases and starve the compounding ones; institutions that celebrate raw token consumption — BCG’s “tokenmaxxing” — invite the gaming of a vanity metric. ROInt holds the line between the two failure modes because it prices both the human and the machine in the denominator, and only value in the numerator. Put it in the monthly pack next to headcount, and the conversation changes shape.
BCG’s second prescription is organizational: move accountability for token spending out of IT — where it reads as a cost to be contained — and into strategic planning or finance, where it can be steered as investment. This publication has argued the finance-native version of that position (Paying for the Machines): finance is both the payer and, increasingly, the operator of the machine workforce, and the function that owns showback, unit economics, and the rate card should own the token budget’s return. A CFO who can state the institution’s ROInt by process — and defend the token line in the same breath as the headcount line — is planning the operation, not just accounting for it.
The tempting reading of abundant intelligence — cut the people, keep the tokens — fails on its own arithmetic, and BCG’s third prescription explains why. In most AI-integrated workflows, humans start, steer, and sign: they choose the problem, shape the work, evaluate the output, and take accountability for the result. Remove them and the institution loses the capacity to convert machine output into trusted, deployable value. The market is already correcting for over-substitution: Gartner, cited in BCG’s research, predicts that half of the companies that cut customer-service staff for AI will be rehiring by 2027.
Finance has a structural advantage here, because its target state was never full substitution. The thesis of this publication draws the line precisely: machines own execution, evidence, and vigilance; people own policy, judgment, exceptions, and attestation (The Last Org Chart designs the roles). The institutions getting this right are redeploying, not deleting — the pattern BCG illustrates with Cloudflare, which cut supervisory “measuring” roles while accelerating hiring in building and client-facing ones.
The gap the second model prices is behavioral, which is why BCG’s final prescription is organizational honesty. Its research puts the reshaping of work at 50–55% of jobs, against 10–15% displaced — a redesign problem an order of magnitude larger than a reduction problem. Adoption stalls for human reasons: people cannot see where the value is, retreat to familiar routines, or resist tools that press on professional identity. The institutions that move first treat the change as the program itself — BCG’s examples run from JPMorgan’s phased advisor rollout, which its leadership credits with making client research dramatically faster, to Reckitt’s function-by-function redesign of marketing and R&D. The finance translation is this publication’s standing sequence: redesign the close, the reconciliation, the investigation around the machine workforce — then let the token line carry the volume it was bought to carry.
First, put return on intelligence in the monthly pack — by process, next to headcount. One metric, both workforces, and the capital-allocation debate the token line deserves.
Second, move the token budget into finance’s planning cycle now, before it calcifies as an IT cost center. Whoever owns its return will own the operating model it funds.
Third, staff the judgment layer before thinning the volume layer. The rehire wave of 2027 will be populated by institutions that read abundant intelligence as a substitution story instead of a redesign story.
The Lights Out Maturity Index: six questions, two minutes, no scales to interpret. Your anonymous result joins the inaugural Lights Out Finance Survey — the benchmark this publication reports on.
Take the Maturity Index Browse all papersFounder & Editor of Lights Out Finance. Big 4 partner in CFO Advisory & Finance Transformation with two decades across the Americas, EMEA, and APAC; DEng in AI (George Washington), MBA in Finance (Cornell), Master in Financial Engineering (Queen’s Smith), MEng/MBA (École des Ponts ParisTech); US CPA, CGMA, FRM, CQF, CTP, CDAA.