STRATEGY

The core of the business no longer runs on technology.It runs as technology.

Most of your cost and most of your decisions already happen inside a system. Not in a committee: in an origination queue, in a decision engine, in an approval flow that runs thousands of times a day.

The company still governs those systems as support. It budgets them as technology spend, measures them by availability, and reviews them when they break.

That distance — between where the business happens and where attention goes — is what this page measures. Not with opinion. With four figures that hold together.

THE EVIDENCE

Four figures. One proves nothing; chained, they do.

67%

of the cost of producing a mortgage is personnel expense

US$6,900

top quartile, per loan

US$16,500

bottom quartile, per loan

Product cost is process, not capital.

Between the most and least efficient quartile there is a 2.4x gap. Same product, same risk, same funding, same market. The only thing that changes is the process.

Freddie Mac, Cost to Originate Study 2024, n=203 institutions. Salaries and benefits = 47.2% of US banking noninterest expense (FDIC, full census). Cost-to-income of euro-area significant institutions: 54.95% (ECB, n=111).

17 → 40

systems in scope, between two measurements of the same programme

21% → 17%

automated controls

+44%

programme cost in two years

Digitising is not redesigning.

KPMG measured the same internal-control programme twice, with the same methodology. Systems more than doubled. Automated controls fell. Hours per control rose. Systems were bought and manual work grew.

KPMG SOX Survey 2025, n=146 companies. Only 32.6% of invoices are processed without human intervention (Ardent Partners, n=212). 65% of medical prior authorisations remain manual despite an electronic standard existing (CAQH, 600+ organisations). A worker loses 9% of the day re-orienting between applications, instrumented measurement (HBR / Soroco).

14×

difference in customers per employee inside the same industry

6×

passengers per employee, airlines

3.3×

revenue per employee, fashion retail

Structure does not grow the same everywhere.

Nu Holdings serves around 13,100 customers per employee; Santander, around 907. It is not a quirk of financial services: within the same sector, the top decile produces five times more value per worker in manufacturing and more than seven times in services.

And the case that does not fit

Lemonade serves 3.7 times more customers per employee than Progressive — and generates 2.7 times less revenue per employee. The two axes move in opposite directions when the unit value of the product differs. Anyone using a single ratio to carry an argument is not measuring: they are picking the number that suits them.

The full development of this axis →

FY2024–FY2025 annual reports of the companies cited. Within-sector dispersion: Berlingieri, Blanchenay and Criscuolo, 16 countries, up to 812,850 firms.

is the return on technology without redesign, statistically

>80%

see no EBIT impact from generative AI

21%

have redesigned their workflows

Buying technology without redesigning destroys value.

Firms combining high technology with a redesigned organisation outperform sector peers by more than 4%. Firms doing only one of the two perform worse than firms doing neither. And across more than 30,000 establishments with US census data, the effect of predictive analytics on productivity is never different from zero without the organisational complements.

Before the figure, a warning

The number everyone quotes — that 70% of transformations fail — has no study behind it. A peer-reviewed paper traced the five sources that claim it and concluded no valid evidence supports it. If someone puts it in front of a board, they are repeating, not measuring.

Bresnahan, Brynjolfsson and Hitt, Quarterly Journal of Economics, 2002, ~400 firms. Brynjolfsson, Jin and McElheran, 2021, 30,000+ establishments, US Census Bureau data. McKinsey, March 2025, n=1,491 — correlational, not causal. Hughes, Journal of Change Management, 2011.

THE CROSSOVER

This is not a banking argument.

An airline’s revenue management, a retailer’s pricing and assortment, a logistics operator’s dispatch and a bank’s origination are the same problem: a business decision taken thousands of times a day inside a system, which nobody revisits because the system “works”.

Banking is where the deep proof sits — nine operating models with unit economics — but the thesis is not a banking one.

37.8%

of airline operating expense is labour

16.8% is fuel

44%

of trucking cost per mile is the driver

21% is fuel

The input that defines each industry weighs half of what the human process administering it does.

THREE QUESTIONS

01
What does it cost to originate one unit of your main product, today?
02
What share of that cost is time spent inside a system?
03
When was that process last redesigned?

If you cannot answer all three, it is not a data problem. It is that the process has no economic owner — it has a functional owner, which is a different thing.

The same industry, split by its own operating model

Each row compares two companies in the same sector. The laggard is worth 1. The distance is what the operating model explains.

referencecore runs as software1× = the laggard

The same pair, measured by revenue per employee, inverts: more customers served is not more revenue generated.

Denominators are not identical — customers, passengers, policies, revenue — and fiscal years differ. The gaps are not explained by those differences, but every figure carries its label. Sources: FY2024–FY2025 annual reports of the companies cited.

THE PIECES

A strategy that runs is a circuit, not a list.

Four pieces, and each one feeds the next. The economics change the decision, which changes the process, which changes what is measured. Remove any one of them and the loop stops closing — what is left is a plan.

Click a piece to take it out.

Decision

What the machine decides, what it proposes, what it escalates.

Process

With an owner who answers for the P&L, not a functional owner.

Measurement

Per unit served. Not system availability.

Economics

The unit cost that actually moves.

The loop closes. Every change can be traced to a unit cost, and every unit cost points back at a decision someone owns.

BEFORE AUTOMATING

What has to be redesigned first

The literature is consistent about which complements make technology pay. Translated into what a board can decide:

The process has a P&L owner

Not a functional owner. If nobody answers for unit cost, nobody will redesign it.

The decision is described, not just the flow

What the machine decides, what it proposes, what it escalates. Without that there is nothing to automate: there is a faster form.

The process is measured per unit

Cost per origination, not system availability.

Whoever operates it can change the rule

Without opening a project. If every adjustment costs a budget cycle, the process is frozen on the day it went live.

One figure worth saying out loud: when the effect of changing management practice on productivity is measured, it appears two and three years later. Redesign does not pay inside the quarter. Anyone promising otherwise is selling.

Bloom, Sadun and Van Reenen, American Economic Review, 2012 — quasi-experiment over 21,746 establishment observations.

THREE EXITS