Platform Last

The received order of operations for a first analytics capability has not changed much in the last decade. Consolidate the sources. Build the warehouse or the lake. Establish lineage, cataloguing, quality gates and access control. Hire the scientists. Only then let the business bring use cases to it. Two years is the number people say out loud, and the second year is the one that gets cut.

I have watched that order fail often enough to stop arguing with it in the abstract. More useful is a document that inverted it on purpose and then put a price next to the inversion, because pricing is where sequencing stops being a philosophy and becomes a commitment somebody has to sign.

The document is a proposal for data analytics support put to a European food and consumer goods manufacturer standing up its first analytics capability across marketing, sales and the supply chain. I co-wrote it. It was offered and never delivered, so nothing below is a result. What it is, is an argument about order.

Two prices for the same starting capability

The commercial page is the part I would show to anybody who thinks sequencing is a slide topic.

Option one buys a lean team of analytics specialists and data scientists who start on an already-identified case. Sized at 1.3 full-time equivalents, six to eight weeks, a quoted 7,300 euros per week. The stated outcome is analytics insight and a first commercial payoff. No environment is included, and the offer does not make one a condition of starting.

Option two buys two cases, a larger team at 2.1 full-time equivalents, and a cloud analytics environment with licensing brought by the supplier, so the manufacturer can begin testing an architecture while the cases run. Quoted at 15,000 euros per week. The stated outcome adds a foundation for an analytics hub to the first commercial payoff.

The difference between the two lines is 7,700 euros a week. Some of that is the second case and the extra 0.8 of a full-time equivalent, and some is the environment and its licences. The proposal does not itemise the split and I am not going to invent one. The shape matters more than the arithmetic: the platform sits on the far side of a priced boundary, crossed only if the buyer chooses to cross it. A third option, a scoping session plus a capability review, is offered separately, the session carrying no cost if either paid option is taken.

Everybody says start small. Almost nobody writes an offer where the foundation is optional, because the foundation is the profitable part and the part that lasts long enough to be renewed.

The roadmap comes after, and says so

The covering letter does not hedge this. It proposes to choose one to three areas where a first proof case with concrete results can be delivered in a relatively short window, to make those choices in a prepared workshop, and then, building on the results delivered and the insights gained during that delivery, to explore the roadmap that would help build the right platform and the right skills.

Read that ordering carefully. The platform question and the skills question are not deferred out of timidity or budget. They are treated as questions the first case is expected to answer. You cannot specify what to build until something has failed in a way that tells you what to build, and nothing fails informatively until a real question is put to a real data estate under a real deadline.

The proposal is honest about what that first case runs into. It lists the obstacles in the analysis phase as data sources not being available across silos, or being protected by the people who own them, and the company lacking the combination of data-literate business people and data scientists. Those are the findings a long foundation programme is supposed to prevent and routinely does not, because a silo boundary is a political fact rather than a technical one, and it does not surface until somebody needs to cross it for a reason that has a sponsor attached.

The platform decision is handed back on purpose

There is a detail in the effort planning that I have come to value more over time than anything on the commercial page.

The proposed engagement is scored task by task, supplier share against client share. The supplier takes 90 percent of crunching the data into a usable state and 90 percent of integrating and transforming it to test the hypotheses, 80 percent of understanding the data model and assessing source quality, 70 percent of defining the approach. Then, on the line that reads agree which analytics platform to use, it takes 10 percent and leaves 90 percent with the client. The same 10 to 90 split appears on providing security and technical safeguards before the data is gathered.

A supplier that wanted to sell a platform would not score that line that way. It says the platform is the manufacturer's decision, made with the supplier's evidence rather than the supplier's preference, in a table the client can hold the supplier to. The heavy supplier share sits on the perishable work, which has to be redone for the next case anyway. The client share sits on the durable decisions.

7,300 euros / wk

Quoted, one proof case, 1.3 FTE, no platform

15,000 euros / wk

Quoted, two cases plus supplier analytics environment, 2.1 FTE

10% / 90%

Proposed supplier / client effort on choosing the analytics platform

6 to 8 wk

Quoted duration, against 4 to 6 called typical for an established environment

Why this order is the right one

The argument for platform-first was always economic, and the economics have moved.

When a pipeline was a year of bespoke engineering, it made sense to build it once and route everything through it, because the marginal cost of the second use case was what you were optimising. That has stopped being true. Ingestion, orchestration, model deployment and monitoring are close to commodity in most stacks now. The slow, irreducible part of a first capability is not standing the machinery up. It is finding out which of your numbers mean the same thing twice, which of your silo boundaries are defended, and which of your questions have an owner who will act on the answer. None of those three are discoverable from an architecture diagram. All three are discoverable in six weeks by attempting one question end to end.

A platform bought before the first case is a platform shaped by nobody's workload. The feature store gets built with no consumer to argue with its definitions. The lineage graph gets populated and nobody reads it, because reading it becomes urgent only when a number in front of a director disagrees with a number in front of a plant manager. Then the first real case arrives, wants three fields nobody modelled, and the programme discovers in its second year what it could have discovered in its second month.

Running the case first inverts each of those. The definitional fight happens with a live question on the table, so it gets settled rather than documented. The silo boundary gets crossed, or refuses to be crossed, and either way you learn something true about the organisation. The platform requirement then arrives written in workload rather than in vendor capability, which is the only form of requirement worth buying against.

A platform bought before the first case is a platform shaped by nobody's workload. You do not discover what the estate cannot do by drawing an architecture. You discover it by asking one question of it that somebody upstairs actually cares about.

What the inversion costs when it is wrong

I am not going to pretend the order is free of risk, because the proposal does not.

The failure mode of case-first is a capability that never compounds. Six weeks produce an answer, the answer is applauded, the working data set is abandoned, and the next case starts from the same silo negotiations as the first. That is why the proposal keeps a final week for benefit tracking and states the risk plainly, that the company may lack the discipline to run continuous benefit reporting. Case-first works only if each case leaves behind a piece of the foundation the next case inherits.

The other risk is the choice of case. The proposal's bar is five times return on effort, and it names the absence of a case meeting that bar, with the absence of senior sponsorship, as the two things that sink the framing week. A case chosen because the data happened to be available will clear a technical review and teach the organisation nothing about its priorities.

Both risks are cheaper than the alternative, because both surface inside six to eight weeks. A foundation programme that has picked the wrong shape surfaces that far later, with the sunk cost already argued about in a steering committee.

This was an offer. The sequencing was put to a manufacturer with a price on each half, no engagement followed, and there is no outcome to report. What survives is the structure of the offer, which I would still write the same way: prove one thing, let the proof specify the foundation, and let the buyer decline the foundation entirely if the proof does not earn it.

Figures are as written in a proposal for data analytics support put to a European food and consumer goods manufacturer: its priced options, its effort planning and its stated risks. Prices, durations and effort splits are quoted terms of an offer that was never taken up, not measured outcomes. The reading of the sequencing is ours.