An operating model diagram is the page in a board pack that everyone reads and nobody adds up. Each box carries a headcount range, because a box without a number looks unfinished. The ranges get chosen one box at a time, by people who know roughly what a governance function costs or what a delivery squad needs to be viable. Then the page is approved as a picture, and the picture is filed as the target. The sum of the boxes is never printed anywhere on it.
Going back through a target operating model I worked on for a consumer and SME banking group operating across several European markets, I ran that sum. The design is a conventional hub and spoke, and the numbers on it are sensible per box. A central hub of twelve to sixteen people owns AI strategy, governance, shared platforms and specialist expertise. Five embedded spokes sit in retail banking, SME banking, risk and compliance, HR and support, and finance and operations, at four to eight people each. Every one of those figures is defensible on its own. Multiply the second one by five and the picture describes an organisation of thirty-two to fifty-six people.
A few pages later in the same pack, the funded roadmap tops out at twenty.
- 12-16
- People in the central hub, as designed
- 4-8
- People per embedded spoke, across five business units
- 32-56
- Derived total the reference model implies, never printed on it
- 20
- Headcount the funded roadmap reaches at month 24
The reference model prices five spokes and the reader prices one
The gap opens because of how the diagram is read rather than how it is drawn. A hub-and-spoke picture puts one hub in the centre and repeats a spoke shape around it. Attention lands on the hub, which is the box with the interesting job titles in it, and the spokes register as a pattern rather than as five separate hiring plans. The eye takes in "four to eight" once. The budget has to take it in five times.
The split of responsibility is a good one, and the arithmetic is not an argument against it. The hub owns what creates duplication and risk when it is decentralised: enterprise strategy, governance standards, shared platforms and infrastructure, and deep technical expertise no single business unit can justify keeping. The spokes own what dies when it is centralised: use cases built on hub infrastructure, alignment with what the unit is trying to do, and a liaison role that gives the unit a name to call. Risk and compliance gets a spoke of its own rather than sitting in the centre, which puts control expertise inside delivery instead of at the end of it.
That structure is coherent. What it is not is cheap. Five business-facing teams, each large enough to carry a business product owner and enough engineering to ship, is a real organisation, and the design says so in the only place it can, which is the FTE range in the box.
The pack does cite an external benchmark on the same page, attributed to Bain and Company: roughly sixty per cent of financial services firms run centralised governance with hybrid staffing. That supports the shape of the model. It says nothing about its size, and it is somebody else's survey rather than anything measured at this bank.
The funded roadmap is a different object
Beside the target picture sits a twenty-four-month staffing and budget plan in four six-month phases. Three to five people in the first, ten to twelve in the second, fifteen to twenty in the third, twenty in the fourth. Why the First Six Months of an AI Function Cost the Most Per Person works the same curve for cost per head and finds the opening phase roughly two and a half times more expensive per person than the closing one. The point here sits at the other end of the same curve. Month twenty-four, the point at which the plan considers the function mature, lands twelve posts below the cheapest possible reading of the target picture and thirty-six below its top end.
These two pages are not in contradiction so much as they are different kinds of document. The reference model is a shape argument. It answers where capability should sit and who owns what, and it sizes each box at the level where that box works. The roadmap is a funding argument. It answers what a board will approve in six-month increments against an investment envelope. Nothing forces the second to reach the first, and in my experience it almost never does.
The reference model says what good looks like when it is finished. The roadmap says what will be paid for. Neither page is wrong, and the missing page is the one that says which parts of the first arrive last.
The workshop hands out a third number
The same pack carries a ninety-minute session in which executives design the organisation themselves, and the participant material sets its own constraint: fifteen to twenty posts for year one, to be split between hub and spokes, with the first three critical hires named.
Follow the printed numbers strictly and the exercise does not close. A hub of twelve to sixteen, plus five spokes at four to eight each, starts at thirty-two. The canvas asks for that allocation inside fifteen to twenty. Even reading "four to eight" as a single pool covering all five spokes, the hub and the pool together give sixteen to twenty-four, which meets the stated envelope only in the narrow band between sixteen and twenty.
I would not call this a defect in the material. A constraint that cannot be met by copying the reference model is exactly what forces a room to make a choice, and the run-book closes by asking each table what its hardest trade-off was, which makes the choice the output. It flags the classic failure modes by name: leaving risk and compliance out of the organisation, putting everyone into IT and creating a bottleneck, and isolating data scientists from the business users they are meant to serve. It also presses the question most target operating models skip, which is who pays, the centre or the business unit. A group that walks out with fifteen posts allocated has answered a question the diagram was allowed to leave open.
A peer benchmark in the same pack makes the point from the other direction. Another European bank is reported as running a team of about fifteen with bank-wide reach by building on hyperscaler partnerships instead of hiring against every box. That is somebody else's figure rather than anything measured here, and it is enough to show that twenty is not obviously too small.
The number nobody writes down
Take the funded endpoint at face value. Twenty posts at month twenty-four, against a hub the design floors at twelve. That leaves eight people for five business units. Spread evenly, each spoke gets one and a half people, which is not a spoke. It is a named contact with a calendar full of meetings about work happening somewhere else. Concentrated, it is two spokes at four and three business units with nothing.
Both of those are legitimate answers. The decision between them is the single highest-value thing on the page and it is the one thing the page does not carry. A target operating model that shows five spokes and a roadmap that funds two is not a plan with an error in it. It is a plan with a sequencing question left unanswered, and the answer gets made by default, usually by whichever business unit has the loudest sponsor.
Three things fix it, and none of them costs anything to add.
Put a start month against each spoke rather than a headcount. A spoke that opens in month nineteen is a different commitment from one that opens in month seven, and the diagram treats them identically.
Publish a threshold below which a spoke is a liaison rather than a team. Two people embedded in a business unit with a hub behind them is a workable arrangement if everyone knows that is what it is. The same two people labelled a delivery spoke will be held to a delivery promise nobody funded.
Say where the use cases belonging to the unfunded spokes go in the meantime. They do not stop existing because the org chart has not reached them. They go to the hub, or they wait, or they get built by the business unit outside any governance at all, and the third of those is the one the whole model was designed to prevent.
Source: a board deliverable and executive workshop kit I worked on for a consumer and SME banking group operating across several European markets, dated January 2026. It is a design and a proposal. The hub and spoke sizings, the four-phase roadmap and the workshop constraint are all planning figures from that pack; the thirty-two to fifty-six total is our multiplication of the published per-box ranges and appears nowhere in the source. The sixty per cent governance statistic is attributed in the deck to Bain and Company, and the peer team of about fifteen is another institution's reported figure. Nothing described here had been hired or built when the pack was written. The reading that the gap between the reference model and the funded roadmap is a sequencing decision rather than an error is ours.
“A spoke with one and a half people in it is not a spoke. It is a named contact with a calendar full of meetings about work that is happening somewhere else.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.