A bank deciding to build an AI function of its own usually asks the two questions in the wrong order. How many people first, then how much money. The plan that came out of this engagement answers them the other way around, and the arithmetic between the two answers is the part worth reading.
A consumer and SME banking group operating across several European markets took a board-level AI programme through an executive workshop, and one of the written outputs is a staffing and investment plan laid out in four six-month phases. Headcount rises from three to twenty. The budget does not rise on the same curve, and the gap between the two curves is the whole argument.
The first phase, months one to six, plans three to five people on a band of €0.8M to €1.0M: an interim lead at half time, a product manager, two or three data scientists, a governance charter written from scratch and two or three quick-win pilots. The second, months seven to twelve, plans ten to twelve people on €1.2M to €1.5M, appoints the lead full time, adds MLOps and data engineering and puts the first systems into production. The third, months thirteen to eighteen, plans fifteen to twenty people on €1.5M to €2.0M, embeds resources in the business units and closes a compliance audit. The fourth, months nineteen to twenty-four, plans twenty people on the same €1.5M to €2.0M and spends them on reusable components and standardised delivery.
Take the midpoint of each band and divide it by the midpoint of each headcount. The first six months come out near €225K a head. The second near €123K, the third near €100K, the fourth near €88K. Same currency, same organisation, same six-month block, and the first one costs roughly two and a half times per person what the last one does.
The challenge
That shape is not an accident of rounding, and the pack says so in three separate places.
The first is a scheduling note buried in the exercise material: critical hires take three to four months to land. In a six-month opening phase, that means the budget is committed against a headcount that is largely still in a recruitment process. The money is spent on searches, on an interim leader, on tooling contracts and on cloud footprint, and the people it was nominally for arrive with one quarter left to use them.
The second is the proposed first-year investment mix, which splits into 45% central build, 35% business delivery and 20% governance and risk. Two thirds of the first year is bought for the centre and the control function. Only the middle third is aimed at the business units where anything visible to a customer will actually happen.
The third is regulatory and it does not negotiate. The EU AI Act is in force, credit scoring sits inside its high-risk category, and the obligations attached to that category are not a later-phase concern. A bank cannot run a cheap experimental year and buy compliance afterwards, because the systems most worth automating are exactly the ones the regulation reaches first. The workshop's own facilitator material puts a floor under this: a group's budget only counts as a good one if at least 15% to 20% of it lands on risk and control functions.
The worked example makes the point sharper than the phase table does. Against a €900K envelope the exercise proposes strategy and leadership at 15% (€135K), governance and risk at 15% (€135K), platforms and technology at 30% (€270K), pilot delivery at 30% (€270K) and a contingency of 10% (€90K). The same page estimates the first three hires at around €320K, with the senior AI leadership role alone at about €120K. Three people therefore cost more than the whole pilot delivery line sitting beside them in the same €900K. That is the front-loading stated in cash rather than in curve shape, and it is a teaching example inside a team exercise rather than an approved budget, which is precisely why it is honest about the ratio.
The approach
The plan does defend itself against the obvious failure modes, and the defences are worth naming because they are the parts most budgets leave out. A contingency of 10% to 15% is written in rather than assumed, on the stated grounds that specialist hiring slips and cloud run costs are forgotten. Release of phase two funding is gated on three conditions that have nothing to do with spend: a validated value case, completed risk tiering, and data privacy sign-off. The success measures hung on the senior AI leadership role are a return above 10% against baseline, a 40% cut in time to production and 80% of projects reaching scale. All three sit in a role profile written before anything shipped. They are the standard the hire will be held to, not a report on anything that has run.
Where we would argue with the shape is not the front-loading, which is real, but its composition. A large part of that opening €225K a head is spent rebuilding a delivery substrate that no longer has to be built from parts. Model deployment with audit trails, versioning and lineage, drift and bias monitoring, and the documentation a supervisor will ask for are all listed in the plan as engineering work to be staffed. Most of it is now harness work rather than platform work: the evaluation harness that decides whether a model may be promoted, the graded autonomy settings that decide what it may do once promoted, and the trail that records both. That is the layer our Hominis Agentic OS work starts from, and buying it rather than staffing it moves money out of phase one and into the phase where there are people to use it.
The second argument is about sequencing the control function against the delivery function. The plan treats governance as a parallel workstream with its own percentage. In practice the gating criteria for phase two funding cannot be answered unless the control instrumentation already exists, because risk tiering and a validated value case are both statements about systems that have run. Governance built as a committee produces minutes. Governance built as instrumentation produces the evidence the next funding decision needs, and it is the same evidence the regulation asks for.
The outcome
What this engagement produced is a plan, a target operating model with a recommendation, and a workshop that put a board-level group through the budget arithmetic themselves. No function was hired, no envelope was drawn down, and the phase bands are planning ranges rather than a record of spend. The exercise material is candid that the point of the session was to force a trade-off, and the facilitator guidance even includes a scenario card cutting the budget by 20% to see what a group gives up first.
Read as a purchase order rather than a staffing chart, the curve says something a headcount plan cannot. The first six months buy permission, plumbing and a hiring pipeline. Delivery capacity is bought later, and it is bought at a third of the unit price. A board that funds the visible pilot first and defers the invisible foundation does not save the difference, it pays it twice, and the second time it pays under a deadline set by a regulator rather than by a plan. Our Consult work on these builds starts by separating those two kinds of money, because they behave nothing alike.