Every programme review I sit in asks the same question: is it built. Almost none asks the two that decide whether being built matters. Whether there is anything worth running through it. Whether anybody's job changes when it produces an answer.
That is not the reviewers' failure. It follows from how the money was split months earlier. Programmes get funded in layers whether or not anyone writes the word down. Technology gets a business case, a supplier and a date. Content, the material the system consumes and produces, gets whatever a delivery manager can borrow. Process, the decisions and handoffs that change when output arrives, gets a slide near the end and no owner.
A plan I keep returning to refused that split, and refused it typographically. It is an internal digital strategy proposal from a global pharmaceutical company, written for a governance forum, setting out an operational strategy and roadmap across three annual stages. Its outcomes are not a flat list under each stage. They are stratified into three rows, labelled technology, content and process, and the rows carry across the stages. Once they exist, a stage cannot quietly be declared finished on one of them.
What the three rows contain
The technology row runs from stabilising a platform already in flight, through a campaign management solution the plan itself labels interim, to those platforms implemented at scale, and to integrated platforms fed by a shared activity record.
The content row runs from guidance on good content for portals and other channels, plus a plan for non-branded external material and a solution to sourcing it, to a scaled internal and external library held globally with a process to localise it. Sourcing appears before library. Somebody had worked out that the hard part is supply, not storage.
The process row carries nothing in the first stage. It starts at an integrated global operating model for campaign management and analytics, and runs to integrated operations with commercial operations to run localised experiences at scale. The first stage's process work sits in the priorities instead, as a commitment to establish a process for managing future digital demand.
An honest reading of the table
The three labels appear once rather than above every column, so let me be exact about what that supports. The two scaling stages resolve cleanly, one outcome per row. The first stage does not, and the tidy version of this story would omit that.
Classifying its five outcomes against the labels is my reading rather than the plan's: a stabilised platform, high impact initiatives begun and an interim campaign management solution sit in technology; guidance on good content and a plan for sourcing non-branded material sit in content; nothing sits in process. Three, two, none. The plan that imposed the discipline let its own first stage go two out of three on the outcome row. What saved it is the priorities: establish a process for managing future digital demand is one of only four things that stage commits to doing. The layer was funded, and produced no stateable outcome that year, which is a different condition from being absent, and a more honest one.
What the rows enforce, then, is coverage rather than symmetry. Demand one deliverable per layer per quarter and teams will manufacture deliverables to fill cells. Demand instead that every layer have a funded owner and a visible state, including the state of not yet.
The starved layer decides the result
The three failure modes are not interchangeable.
Starve content and the technology runs perfectly on whatever happened to exist. The publishing platform is live and half the material in it is the wrong format for the channels that now matter. Search returns three versions of the same guidance with no way to tell which is current. Nothing breaks. Usage is simply lower than the business case assumed, and the explanation a year later is that the markets did not adopt it.
Starve process and the output arrives at a wall. The system produces a ranked list, a recommended action, a classification, and it lands in a queue where the person receiving it has no mandate to act differently from before. Straight-through processing that still routes everything to a human, because nobody agreed what the machine may decide, is an extra step rather than automation. Plenty of programmes sit there without knowing, because the throughput they report counts what the system produced, not what changed downstream.
Starve technology and the programme fails in week three. Everybody sees it, everybody says so, and the money moves. That is why it is the safe layer to underfund and the one nobody underfunds. The two quiet failures take a year, surface as adoption problems, and get an engagement survey rather than a budget correction.
The instrument that catches it
The plan does not leave the layers as an intention. It puts owners behind them: six functions with distinct ownership, covering strategic direction and guidance to markets, service design, engagement, content origination and delivery, operations, and technology. A seventh capability, advanced analytics, cuts across all six rather than sitting beside them.
Two details are worth copying. Content origination and delivery is chartered to source, reuse, manage, monitor and maintain, not only to produce, so content is governed as an asset with a lifecycle rather than as campaign exhaust. And the analytics mandate is split in two. One half is what everyone builds: dashboards, indicator tracking and targets at group, region and market level. The other half is conformance and course correction, which names content utilisation and increased use of published assets among the things measured.
Whether your published material is ever retrieved and used is the only reliable detector for a starved content layer. Everything else measurable about content is production volume, which rises fastest exactly when the layer fails.
- Three rows
- Technology, content and process, held constant across the stages
- 3 / 2 / 0
- Our reading of the five first-stage outcomes by layer
- Six functions
- Distinct owners, with analytics cutting across all six
- 30 / 60
- Days to a usable draft of the guidance, then to committed publication
Why this lands hardest on machine learning programmes
Every layer has a counterpart in the work I get called into, and the imbalance is worse here than in commercial digital. The technology layer is pipelines, model deployment, monitoring, the whole of MLOps, and it has never been cheaper or closer to solved.
The content layer is what the model eats and what it emits. Labels somebody defined in advance. Definitions that mean the same thing in two systems. Lineage that survives a question about where a value came from. A feature store is a content asset before it is a technology one: a promise about meaning, written by people rather than by the store. Almost nobody funds it as a layer with an owner. It becomes preparation inside a modelling project, which is how a model ships on features nobody can reproduce next quarter.
The process layer is who acts on the output, what the system may decide alone, who can override it, and whose number wins when two functions read the same figure differently. Process mining tells you what the work does today, and so whether your intended change has anywhere to land.
The early language model pilots crossing my desk sharpen this rather than soften it. They stand up in days, so the technology layer now costs almost nothing to satisfy and two out of three becomes one out of three, with an impressive demonstration on top. Cheap technology raises the relative price of the other two layers and makes skipping them feel free.
A starved technology layer fails loudly in week three, which is why it is the safe one. A starved content layer and a starved process layer fail quietly for a year, and then get written up as an adoption problem.
How to fund three layers when you can afford two
You narrow the scope until all three fit, and the plan shows how. Its first stage is deliberately unglamorous. Stabilise something already running. Deliver a stopgap and write the word interim beside it, so it cannot become permanent architecture by default. Publish guidance rather than centralise execution. Build the intake for the demand that arrives once the first result works. That is a smaller technology bet, chosen so the content and process work fit in the same envelope.
Then it gives the content layer dates: a usable draft of the guidance estimated within thirty days, final publication to the markets committed within sixty. The wording differs on purpose, an estimate for the internal draft and a commitment for what the markets await, which is the tell of a layer being run rather than hoped for. The word usable does work too. Not final, not perfect, usable.
That is the test we run at the front of a RealAI Platform engagement, before a plan is approved. For each stage, name the technology outcome, the content outcome and the process outcome. If one comes back as a phase two item, the plan is two thirds of a plan with a deferral where the third layer belongs. Cut the technology outcome until the other two become affordable, because the layer you leave out decides what the programme is worth, and you will not find out for a year.
Read from an internal digital strategy and roadmap proposal at a global pharmaceutical company, presented to a governance forum. It is a planning artefact: a proposed sequence, a layered outcome table, a governance operating model and two dates, not delivered results. Row labels, outcomes, function definitions and day counts are as printed there. Classifying the first stage's outcomes by layer, and reading the table as a funding rule for machine learning programmes, is ours.
“A starved technology layer fails loudly in week three, which is why it is the safe one. A starved content layer and a starved process layer fail quietly for a year, and then get written up as an adoption problem.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.