A rubric is only worth publishing if you are prepared to lose an argument to it. A global staffing and HR services group wrote five design principles for a digital programme before it had any candidate operating models to test, then reused those same five, unchanged, as the criteria when the candidates arrived. Four options, five principles, twenty written judgements in twenty cells.
Read as a scorecard the grid looks like a formality, because the option the study recommends is also the option that scores highest. Read cell by cell it says something harder and more useful. Two of the four options were interchangeable. The two that survived read alike on three principles out of five, and of the four points that separated them, three sat in a single cell.
The challenge
The group is federated by construction. National markets carry their own clients, their own systems and their own results, and a market that has run its own operation for a long time does not experience a central instruction as help.
The programme sitting on top of that had two propositions to serve. One is throughput work, where the target was straight-through processing on efficient technology. One is professional work, where the target was a candidate experience joined up across channels. They pull in different directions, and they converge on the same requirement: matching algorithms, job suggestions and demand prediction, all of which need data pipelines, lineage and governance that cross market boundaries rather than stopping at them.
None of that is an organisational question until you ask who builds it, who runs it afterwards, and who is allowed to say no. That is the question the blueprint had not answered.
Five principles were already on paper from the earlier blueprint work. Service delivery should produce a good outcome for customers and for talent. Experience should be consistent across touchpoints, channels and geographies. Improvements should land early, with a steady pipeline behind them. Data should drive matching, experience and everything downstream of both. And the design should build on strengths the markets already have rather than replace them.
The discipline worth copying is the sequence. The rubric was published first and the options were tested against it afterwards. A rubric written after the options is not a rubric, it is a justification.
The approach
Four candidates were put on the grid, and each was written up with its failure mode named rather than implied.
The first concentrates everything centrally: the centre decides, the centre builds, the markets receive. The second keeps the centre as the builder but turns its output into shared services the markets can draw on. The third is federated, with a centre that convenes rather than owns. The fourth is a central unit with local representatives seated inside it, executing through the markets, with one named leader accountable for delivery and co-ordination.
The two centralised options failed on the same four principles for the same reason, stated four times: without engagement in the local markets, the work misses local needs, adoption is hard to win, consistency suffers as a result, and almost nothing travels back up. Both were quick to mobilise and slow to market. Both are hard to run in an organisation that is not already centralised, which this one is not. The only thing separating them was that the shared-services variant could at least co-ordinate investment centrally, and might notice a good local idea if somebody went looking for it.
Counted that way, the four options total minus four, minus three, plus two and plus six. The gap between the two centralised options is a single point across all twenty cells, which is the first honest finding: they are one option written twice.
The second finding is the one that matters. The two federated options record the identical reading on three principles. Both combine central oversight with local execution, which is where consistency actually comes from. Both use skills already sitting in the markets. Both seat local representatives inside the central unit, so practice travels upward instead of only downward. On those three principles the choice decided nothing at all.
Four points separate them in total. One of those points is thin: both cells say the local markets are engaged so the work can be tailored, and the winning cell adds a second statement, that shared central expertise improves the offer. The other three sit in one cell. One option has no single point of accountability, so progress depends on goodwill and central investment is hard to put to work. The other has a named leader accountable for delivery and co-ordination. Three of the four points between the two live options are in that one judgement.
This is where an operating-model debate usually goes wrong. It is argued as centralisation against local autonomy, loudly, while the live options on the table are already both federated and the real question is narrow: does one person own delivery, or does the centre co-ordinate by consensus. That question is not about org charts. It decides who can retire a model, who can refuse a market's request for a bespoke variant, and who signs off a shared data pipeline that every market downstream depends on.
The outcome
What was delivered is a design and a recommendation, not a running organisation. The grid selected the hub-and-spoke shape, and four reasons were recorded behind that choice rather than left to the colour 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 both development and execution, so what ships adheres to the guidelines. Local ambassadors carry engagement and adoption inside their own markets. And central investment is easier to put to work when the platform and the skills are grown once rather than market by market.
Three conditions were attached: high levels of engagement, tools that actually enable the markets, and early mobilisation. Two lessons were recorded, and both are about timing. Get the governance agreed and operational early, because the moment you step into real decision-making territory, an unagreed governance model works against you. And get the central resources on the ground and delivering early, because a centre of excellence that has not shipped anything is seen as a pipe dream, and a centre nobody believes in has no authority when the first contested decision arrives.
The people plan sits underneath and front-loads the cost. A blend of creative, technology, data, operations and management skills, most of it needed in the first horizon, to detail the strategy and execute it at the same time. Sourcing is hybrid on purpose: upskill existing staff, hire permanent roles, and use contractors to move while the other two catch up. Hire the key roles early enough that they help define the vision, and do not overlook retraining people already inside the organisation.
Two things travel from this artefact to any group standing up a machine learning or data capability across many markets.
The first is the shape of the split. The central unit owns the pipelines, the feature store, the model deployment path, the lineage and the governance, because those are the things that are expensive to build once and ruinous to build in every market separately. The markets own the use cases, the ground truth and the adoption, because those are the things that do not survive being specified from a distance. That division is what the Platform is built around, and it is also why the group's own advice, to stick to market standards and explain any deviation rather than invent, is not a technical preference but an operating-model decision.
The second is the accountability cell. A shared delivery unit without a named owner degrades into a co-ordination forum, and a co-ordination forum cannot retire a model. The pipeline keeps running, the numbers drift, and everyone waits for consensus. Write the owner into the design before the first release, not after the first argument.
What this grid does not contain is any weighting. Five principles, treated as equal, with no statement anywhere that they are. And the cell doing most of the separating widens only because the losing option collected two written drawbacks in it while the winner collected a strength. Three points of rubric, produced by how many sentences somebody wrote in one box. That does not make the recommendation wrong. It makes the grid a record of the judgement rather than the source of it, and the four stated reasons underneath it are doing the work the cells appear to be doing. Check which one your own rubric is before you rely on it.