Most option grids are written after the decision. You can tell because the criteria fit the winner the way a suit fits the man it was cut for, and because the grid is never seen again once it has done its job of justifying the thing that was already going to happen.
A global staffing and HR services group did the opposite, and the evidence is not that the argument reads well. The evidence is that the same five criteria were taken off the shelf a second time, all but unchanged, and pointed at a question they were never written for. A frame that survives being reused is a frame. A frame that gets rewritten for each decision is advocacy with a table around it.
The challenge
The group is federated by construction. National markets carry their own clients, their own systems and their own numbers, and a market that has run its own operation for years does not experience a central instruction as help.
What the programme wanted to build crossed all of them. Matching, job suggestions, demand prediction, and behind those the machine learning pipelines, the feature engineering, the lineage and the data governance that only pay for themselves if they are built once. None of that is an organisational question until somebody asks who builds it, who runs it afterwards, and who is allowed to refuse a market that wants its own variant. That was the open question.
Five design principles were already on paper from the earlier blueprint work, written before any candidate operating model existed. Service delivery should be good for customers and for talent. The experience should be consistent from end to end, across channels and countries. Improvements should arrive quickly. Data should be the thing driving matching and experience rather than a by-product of them. And the design should build on strengths the markets already have.
The approach
Four candidates went onto the grid, and each cell holds a written judgement rather than a number.
The first concentrates everything centrally: the centre decides, the centre builds, the markets receive. The second keeps the centre as builder and turns its output into a catalogue of shared services the markets can draw from. The third is federated with a centre that coordinates rather than owns. The fourth is a central delivery unit with local representatives seated inside it, executing through the markets under one named leader accountable for delivery and coordination.
The two centralised options failed for much the same reason, written twice: with no engagement in the local markets, the work misses local needs, adoption is hard to win, consistency suffers because of it, and almost nothing travels back up. Both are quick to mobilise and slow to market. Both are hard to run in a group that is not already centralised, which this one is not. The catalogue version does carry two readings the mandating centre does not, that shared central expertise can coordinate investment and that some learning travels back up where the central unit goes looking for it. Neither moved the verdict.
The two federated options record the same reading on three of the five principles. Both put central oversight over local execution, both use skills already sitting in the markets, and both carry practice upward because local people sit inside the central unit. Two cells separate them, and only one of the two did any work. Under the customer and talent principle, the option with local representatives seated in the centre picks up an extra line about shared central expertise strengthening what is on offer. Under the principle about delivering improvements quickly, one option has no single point of accountability, so progress depends on goodwill and central investment is hard to put to work, and the other has a named owner. That second cell is the one the recommendation turns on.
Four reasons were recorded behind the recommendation, and they were recorded as sentences rather than left to the shading of the cells. Central functions can develop guidelines, practice and services quickly, and the markets can execute them. A single shared delivery unit gives close control over development and execution together, so what ships matches the guidelines. Local advocates carry engagement and adoption inside their own markets. And central money goes further when the platform and the skills are grown once rather than in every market separately.
Then the same five principles were used again. The next decision in the same blueprint was where the people would come from, and it had four candidates of its own: retrain existing staff, hire permanent, hire contract, outsource. Same five principles down the side, same convention of a written judgement in every cell. That grid did not produce a winner. It produced a hybrid, three routes run together, with outsourcing held back until the operation was running rather than ruled out. One of the objections recorded against outsourcing was that the data generated by the work would leave the organisation and might not come back in.
Two decisions, one frame, and exactly one adjustment between them: on the sourcing grid the principle about building on existing strengths carries a parenthetical narrowing it to the digital staff already in place. That is the whole edit. It is what makes the instrument worth copying and it is also where its limit shows. Principles written for a customer and data programme do not have equal grip on a hiring question, and it is visible: on the sourcing grid, one cell records that the group already has staff in different countries at different levels of maturity. That is a fact about the group, not a judgement about the option above it. A reused frame carries its origin with it, and the honest way to read a cell like that is as context rather than as a point scored.
The outcome
What this phase produced is a design and a recommendation, not a running organisation. No adoption figure, no throughput figure, nothing measured.
The place the chosen model stops being a diagram is the role tables underneath it, and they are more informative than the grid. Every role in the design is placed against three columns: the centre, a large national market, a small one. Read the data family that way and the operating model is priced. All ten data roles are placed centrally. Five carry a firm yes in the larger markets, and every one of those five is written as execution, not oversight. Two reach a small market at all: the steward, kept as a role rather than a full seat, and the compliance officer, and only where the regulations differ.
That table is the recommendation made concrete, and it transfers directly to any group standing up a machine learning capability across many countries. The centre owns the pipelines, the feature store, the deployment path, the lineage and the standards, because those are expensive once and ruinous four times over. The markets own the use cases, the ground truth and the adoption, because none of the three survives being specified from a distance. The steward is the role that cannot be fully centralised, because a shared definition of a candidate, a placement or a client is argued locally before it is agreed globally, and a definition nobody argued with is a definition nobody uses. That split is what the Platform is built around.
Two lessons were recorded next to the recommendation and both are about timing. Get the governance agreed and operational early, because the moment a real decision has to be made an unagreed governance model works against you. And get the central people on the ground and delivering early, because a centre that has shipped nothing is read as a pipe dream, and a centre nobody believes in has no standing when the first contested call arrives.
The reusable part of this engagement is not the option that was recommended. Another group with a different market structure would reasonably pick a different one. The reusable part is a set of principles agreed before the options were drawn, a written judgement in every cell instead of a score, reasons recorded underneath the recommendation, and the willingness to point the same frame at the next decision without editing it to suit.