A first data project that works is the most misread result an organisation can collect. It arrives with a start date, a named group of users and a visible change in what those users do, which is nearly everything a steering committee asks for and almost nothing a transformation needs. The temptation is to treat it as a small copy of the destination, so that the plan becomes a matter of repetition. It is not a small copy of anything. It is an on-ramp, and an on-ramp is judged only by what it merges onto.
A global staffing and HR services group ran both halves of that sentence inside a single engagement. On one side, a short-cycle accelerator took a sales-prospecting problem from a standing start to a working prospect-scoring tool in the hands of twenty sales people by week six. On the other, the same programme material lays out a multi-country integrated platform, the operating model to run it, and a first horizon of initiatives it had yet to prioritise. Both live in one document. Only one of them has a delivery date attached to something that exists.
That gap is the subject here, and it is worth naming plainly before anything else. The six-week case is evidence. The platform is a design.
The challenge
The group framed its own situation around three pressures, none of them about technology in isolation. Client and candidate expectations had moved to relevant, proactive, round-the-clock and first-time-right servicing across whichever channel someone arrived through. Social and recruitment platforms were removing the intermediary from parts of the market. New recruitment technology was changing how talent could be attracted at all. The question the group put to itself was how to cope with those at global scale, which is a scale question rather than a use-case question.
Against that, look at what the six-week proof actually settled. The sales force could not predict where placement demand would appear, so it could not tell which employers were worth calling first. Business rules for filtering the right prospects out of internal and external sources became variables, those variables were weighted into an algorithm, and the algorithm wrote a five-star rating onto each prospect. Three inputs fed it: how large the employer was, how quickly it filled its vacancies, and how varied its open roles were. Twenty sales people worked the reordered list, and their outcomes were captured to keep tuning the score.
The record says the highest-scoring companies rose to the top of the calling list, that the hit rate went up, and that satisfaction among the people using the tool was measured and rose. It gives no baseline and no percentage for either. Those are directions, and they stay directions here.
So the pilot answered one question inside one process for one group of users. The end state in the same document is described in five elements: the experience offered to customers and to talent, the data capability underneath it, the technology platform carrying both, the operating model governing it across country markets, and the organisation with the skills to staff it. A prospect-scoring tool is a thin slice of one of those five. Treating it as the first fifth of the programme is an arithmetic error that costs a year.
The approach
The two tracks were run at the same time rather than one after the other, and that ordering choice is the part worth stealing.
The bottom-up track began with an open wish list of ideas collected from the business, plotted by value of the opportunity against ease of implementation, and narrowed to a small number of proof cases. Those were incubated inside the company on a set of reusable components: concepts, algorithms and open-source data infrastructure, with the group's own people doing enough of the work to keep the capability. The advertised clock was as little as six to eight weeks from an unfiltered list to a scored playbook. The point of that track was never the tool. It was a defensible order of work, and people who could run the next one.
The top-down track answers the question the pilot cannot. Its governing argument is that an integrated platform, combining the experience layer and the data capability with changes to the underlying business capabilities, beats both alternatives on the table. Integrating every legacy system was judged impractical and too slow. Running separate systems per country would give up the scale advantage and produce an inconsistent experience between markets. So the design separates the layer customers and candidates touch from the systems of record behind it, coordinates centrally while leaving delivery local, and calls for centres of excellence to carry the programme afterwards. Two business concepts, one built around personalised professional servicing across channels and one around straight-through processing at volume, converge on that same platform.
Between the two tracks sits the instrument meant to make the arc a road rather than a set of ramps: a capability review across the operating companies, covering data management, technology, organisation, people and users. It was offered and scoped, not run. Its purpose would be to benchmark one market against another and make reuse the default. Without something like it, every country repeats the pilot, learns the same lesson at the same cost, and calls the result local knowledge.
The group also named four conditions for itself, and they read as conditions rather than deliverables. Shared platforms and pooled investment in scarce skills. Leadership that works from facts and results. One view of the customer and the candidate across the silos. People with enough ownership to carry an idea past its first demonstration. None of the four is bought, and none is closed by a successful pilot.
The outcome
What this engagement delivered is one deployed proof case and an accelerator method with a clock attached to it. Everything else on the page is design: the platform, the coordinating centre, the centres of excellence, the benchmark instrument, the later horizons, and a first horizon whose initiatives are named at an indicative level and not yet prioritised or costed. Those are specified and sequenced. Sponsorship and ownership are named as things a governance framework would settle, not as things settled. None of it is built. Any reading of this arc that reports the far end as a result is reporting a plan.
Run the same engagement now and three things change, none of them the shape of the argument.
The first is that data readiness stops being a footnote. The six-week clock is only honest because the proof case reached data that already existed in usable form. Where that is not true, the same case takes a quarter and most of it goes on lineage: which system a field came from, when it was last correct, who is allowed to see it. That work is the actual gate on everything past the pilot, and it should be scheduled as its own leg rather than absorbed silently into the first one.
The second is that the tuning loop deserves to be permanent. Outcomes were captured from the sellers to keep improving the score, which is the right instinct, and in current practice that becomes a held-out evaluation set and a retraining cadence under proper MLOps rather than a manual pass. A score nobody re-measures decays quietly while everyone still trusts it.
The third is that the ramp looks different and merges onto the same road. Today the first case in a services group is usually a retrieval-augmented copilot over the group's own documents rather than a scoring model, and it demonstrates just as fast. The question it faces at week six is identical: where does the answer go, who is accountable for it, and can the next market use it without rebuilding it. The same holds for the early and carefully scoped autonomous experiments now appearing around recruitment workflows. Lineage carries extra weight now that the European AI regulation has reached political agreement, because what will eventually be examined is the record a system kept, not the intention behind it.
The arc a group actually travels is short, and the arc it draws is long. The value of the short one is entirely in whether it was built to join the long one. Our Consult engagements start by asking which of the five end-state elements a proposed pilot would touch, because a pilot that touches none of them is a demonstration with a budget.