A business whose entire product is supplying other people's staff into other people's operations sat down to decide where its own digital people would come from, and concluded that it could not buy the capability in. The reason it wrote down was not price, and not control in the usual contractual sense. It was that the data produced by the work would end up somewhere other than inside the company.
A global staffing and HR services group, standing up an integrated digital platform across national markets that had each run their own systems for years, put four sourcing routes on a page: retrain existing staff, hire permanent staff, hire contract staff, outsource. The five design principles the programme had already published became the rows. Four columns, five rows, every cell a written sentence rather than a score.
The challenge
Outsourcing was not the weak option on that page, which is what makes the result worth reading.
On speed it read well, though not alone. The cell records access to a large pool of talent in a short time frame, against retraining, which carries a lead time to upskill people already on the payroll, and permanent hiring, which carries one to identify and onboard the right person. Contract hiring reads just as fast, and it is the route the page reaches for when it writes down where speed will come from. On geography outsourcing stands by itself. The group already had staff in many countries at different levels of maturity, and both hiring routes meant hiring separately in each. The outsourcing cell offers the potential for global support in one arrangement. Two of the five principles read in its favour, and they were the two the option is usually bought for.
It lost on two others. Against the principle of building on existing strengths, the cell says outsourcing does not bring the key digital skills in house, which will be unsustainable. Against the principle of using data as a key enabler, the cell says data leaves the organisation and may not come back in.
Neither sentence is about money or about quality of work. Both are about what accumulates inside the company while the work is done, and where it sits when the arrangement ends. A third objection, on customer centricity, holds that contract and outsourced staff may be less connected to the group's own values, and it is the softest of the three, since every buyer discounts it. The data sentence is different in kind: a statement about the direction an asset moves in.
That direction mattered more here than it would in most places, because of a decision the same programme had just taken on the same five principles. Three platform shapes were evaluated and an integrated platform chosen, and one recorded reason was that consolidated touchpoint technology gives a conduit to aggregate the data needed to drive the business. Matching, job suggestion and demand prediction all sit downstream of that aggregation. The principle that selected the platform therefore also rejected the route that would have generated the platform's raw material inside somebody else's environment.
The approach
The recommendation names no winner: three of the four routes run together, and a fourth written conclusion covers the one left out. Each is justified by what it supplies and by how long it is needed.
Permanent digital hires come first, for a specific stated reason: digital arrives with a whole new set of data and KPIs, and people who have worked with those bring that knowledge with them. New hires were also expected to push the existing staff, which is a workforce argument rather than a headcount one.
Contract staff come second, sized for speed, with a condition most contracting arrangements never state out loud. There is flexibility to downsize once upskilling is complete. The contractors bridge the lead time the other routes carry, and the bridge has an end date attached at the point of purchase. That condition is also where the third route hides: downsizing once upskilling completes only means something if existing staff are being upskilled alongside, which is retraining running underneath the other two rather than instead of them.
The page then adds something that rarely appears on a plan: some experts are needed only for a short time to conduct skills transfer. That is a role whose deliverable is its own redundancy, and it needs a different acceptance test from a contractor who does the work, because the thing being bought is what remains in the building afterwards.
The last conclusion is about outsourcing itself, deferred rather than deleted. Outsourcing a new operation is risky, the study says, and some elements may be considered once the operation is up and running.
That last sentence is the sequencing rule, and it is the part worth taking away. What makes a new operation different from a running one is that its specification is written by the act of doing the work. The definitions of data elements and entities, the KPI set, the working knowledge of which fields in a candidate record can be trusted and which are decorative: none of that exists on day one. It arrives as a by-product, and the by-product is the thing the capability was being built for. Once the operation runs, the definitions live in the building and a supplier has something to conform to instead of something to invent.
The organisation design underneath corroborates this rather than asserting it. Ten data roles were written out, and the two whose stated job is to create and maintain a common set of definitions for data elements and entities both sit inside the group, one of them scheduled into the first hiring wave. Ownership of the definitions is a line in a role table with a horizon against it, not a policy statement.
The outcome
What this engagement produced was a recommendation inside a blueprint. Four routes evaluated, three of them proposed together, outsourcing set aside with the conditions for revisiting it stated. Nobody was hired against it in this phase, no capability was stood up, and no measurement exists of whether the data stayed. The grid records judgements in sentences and applies no weighting to the five principles, which is how it ends in a blend rather than a winner. What is checkable is that the data objection was written down, and that it was decisive enough to defer the one route that would have covered every country at once.
The sentence now decides a different question. Very few organisations outsource digital staffing wholesale any more. A great many hand their machine learning pipelines to a supplier who builds and runs them in an environment the buyer has never logged into.
The contract in those arrangements names the code and the models. It rarely names the rest, and the rest is where the value is. The cleaned and joined tables. The identity resolution that finally made two systems agree about the same person. The labels, bought or annotated by hand. The evaluation set, worth more than the model because it is the only thing that tells you whether the next model is better. The lineage trail. And the record of everything that was tried and predicted nothing, which is the most expensive artefact of the lot and the one least likely to travel back across a contract boundary.
Ask five questions before signing. Which environment does training run in. Who holds the feature store. Where do the labels live and who owns the annotation guidelines. Who keeps the lineage. And what physically leaves when the arrangement ends. A right-to-data clause is not the same as a copy you can run tomorrow morning, and a supplier's model deployment path is not a capability you have acquired just because it runs on your behalf.
The Platform work we do begins with the data plane rather than with the models, on the same logic the blueprint used: decide what has to accumulate inside the organisation, then decide who is allowed to produce it. Hominis rests on the same assumption, that the evidence trail and the feature history belong to the customer by construction rather than by clause.
The reasoning extends to the cautious language-model pilots now running in document-heavy back offices. The prompts, the corrections and the accept-or-reject record are the material for whatever comes next, and they accumulate wherever the pilot runs. A pilot hosted inside a vendor's environment can succeed on its own terms and still leave the organisation with nothing it can carry to the second one.