Score Every Platform Against One Capability Map

Every stack decision I am asked about arrives the same way. A shortlist of vendors, a meeting each, a month gone. Every one of them opens with a picture of the buyer's business, and on that picture the vendor's product covers almost all of it. Nobody is lying. They are simply different pictures, drawn by parties who each had a reason to draw them, and a buyer comparing coverage claims across shifting surfaces is not comparing anything.

The fix is older and duller than any of the products. Draw the surface yourself, once, before you meet anybody. Then put each candidate on it.

The cleanest version of that discipline I have worked on sits in the enabling-technology chapter of a multi-country platform blueprint for a global staffing and HR services group. Five candidate components were on the table, each with its own vendor set and its own persuasive story. The chapter refused to let any of them supply the picture. It drew one capability map of the business, then reprinted that same map five times, once per candidate, with that candidate's coverage marked on it.

The picture belongs to whoever draws it

A capability map is not an architecture diagram. It does not say what talks to what. It says what the business does, at a level a chief executive would recognise, and it survives a re-platforming, because the business still has to source, screen, match, onboard, pay and bill whatever software is doing it.

That stability is the point. The map here numbers nine stages down the page and runs both sides of a two-sided business through them at once, the client journey and the talent journey side by side in every stage. Three sub-activities are printed twice because they genuinely happen twice, once facing a client and once facing a candidate. A vendor drawing your business would flatten that, because a duplicated activity is awkward on a slide. Held on your own map, the duplication is a finding: two teams, two systems, two definitions of the same verb.

Beside it sits the platform tier, forty-two capability labels in three families of fourteen. One of the fourteen under integration reads as managing business rules for straight-through processing. One under experience reads as building segmented and personalised profiles. Those are not vendor features. They are things the business needs done, written before anyone knew who might do them, which is what makes them a scoring surface rather than a shopping list.

Five candidates, one surface, and the two things it shows

Lay five candidates over the same map and two facts appear that no individual evaluation produces.

The first is the gap. A region that stays dark under all five overlays is a capability nobody in the shortlist covers, and finding it during procurement costs a paragraph. Finding it after signature costs a project. The second is the double-buy. A region that lights up under several of the five is capability you are about to pay for more than once, and the overlap is invisible in any single vendor conversation, because each of them is telling you the truth about itself.

Neither fact is available from the evaluations alone. They are properties of the set.

What the individual evaluations did contribute, once the surface was fixed, was an unusually honest set of weaknesses. Each candidate got the same four boxes: what the component is, what it is strong at, what it is weak at, and examples of who sells it. Because the boxes were identical, the weaknesses could not be written as marketing hedges. They came out as preconditions.

The data component's single recorded weakness is not performance or cost. It is that separating local data to satisfy country-specific legislation and data protection law is often a problem. The experience component's decisive weakness is not content or design, which it is good at. It is the integration into the candidate and job data needed to make a recommendation or a match mean anything, plus the precondition that local sites have to be consolidated onto it before any of the benefit arrives. The customer relationship component manages client relationships well and, in the chapter's own words, will not cater for the candidates, which on a two-sided map is not a gap in a product but half a business left uncovered. The workflow and integration component earns its place by letting locally tailored systems survive, and pays for it with the consolidation benefits it forgoes and the number of integrations somebody now maintains forever. The bought-in sector component brings hard-to-build functionality quickly and cheaply and brings it in a shape that is not unique to the buyer and not easy to bend to the buyer's existing governance.

Read those as vendor criticism and they are trivia. Read them against the map and each is a statement about a specific region of your own business.

A vendor will always cover your business, because the vendor drew the business. Take the pen away and the coverage argument becomes a coverage measurement.

What the map cannot settle, and says so

Read the five weakness boxes end to end and the same class of problem keeps surfacing in different clothes. Where local data is allowed to sit is a legal and governance question. Whether local sites get consolidated onto one platform is a question about who has authority over local marketing. Whether a bought-in component can be bent to fit the buyer's own rules is a question about which of the two bends. Keeping locally tailored systems alive is a decision to employ integration maintainers indefinitely. Those sit in the boxes alongside ordinary product complaints, and they are the ones a shortlist meeting is worst at surfacing.

That is the residue the map leaves behind on purpose. It shows coverage, and coverage is genuinely what a shortlist argues about, but the moment coverage is settled the remaining questions are all about who runs the thing afterwards and under whose rules. None of the five vendor conversations covers that, because none of the five vendors is going to be the one doing it.

5

Candidate components, each laid over the same map

9

Business capability stages, client and talent sides together

42

Platform capability labels, fourteen per family

4

Identical boxes per candidate evaluation

Why this belongs in a machine learning conversation

Nothing in the method is specific to a staffing business, and the shortlist a data leader faces today has the same shape as that one. A cloud data platform. A feature store. A pipeline and deployment layer, sold as MLOps. A workflow tool that carries the straight-through processing rules. A packaged point solution for one high-value use case, bought because building it looks slow. Five candidates, a demo apiece, and every deck opens with a picture of your business.

Lay them over one map and the same two facts appear. The dark region is usually lineage: several of the candidates will read data and none will own the question of where a number came from and what it means. The overlap is usually storage and transformation, bought with the data platform, bought again inside the deployment layer, and bought again inside the point solution, which quietly keeps its own copy of everything it scores. That is the pass RealAI's Platform team runs before a shortlist is written, and it is cheaper than any of the licences it cancels.

There is a third thing the map does that matters more this year than last. The modelling end has got cheap. Pipelines are assembly work, deployment is close to solved in most stacks, and the first cautious language-model pilots crossing my desk stand up in days rather than quarters. When the model was the expensive part, its scope got argued over on the way in. Now that it is not, a pilot lands wherever a team happened to feel enthusiastic, and enthusiasm is not evenly distributed across a business. A pilot on a map has a location, which means it has neighbours, dependencies and an owner. A pilot without one is a demonstration.

What to do with this on Monday

Draw the nine or so stages of your own business on one page, both sides of it if you have two. List the platform capabilities you need done, in your words, before you look at any product. Write down beside each one who will be running it a year after signature, and refuse to delete that column. Then take the same four boxes to every candidate and hold the boxes still, because a template that flexes per vendor is a template that has been captured.

You will not get a winner out of this. The staffing chapter did not. It recommended a combination, with the overlaps and the gaps drawn rather than argued. That is worth more than a ranking, because a ranking hides the two questions you actually have to answer: what you are buying twice, and what nobody is selling you at all.

Drawn from the enabling-technology chapter of a multi-country platform blueprint for a global staffing and HR services group: one business capability map, five candidate platform components, each evaluated on the same four-box template and overlaid on the same surface. That chapter produced a recommendation and a set of preconditions ahead of a procurement decision, not delivered results. The reading of it as a method for assembling a machine learning stack is ours.