Most multi-year plans I am handed are one plan repeated three times at increasing volume. The first year runs a pilot, the second runs more pilots, the third runs those pilots in every market. Nothing changes shape; the only variable that moves is spend.
A plan I keep returning to does something else. It is an internal digital strategy proposal from a global pharmaceutical company, written for a governance forum, setting out an operational strategy and roadmap for three years. The interesting part is not the technology. It is the sequencing move: the plan sets one business objective per year, and the objective does not grow, it changes category.
The first year: deliver local, measurable, high impact digital projects. The second: deliver global customer impact at scale. The third: deliver an integrated customer experience at scale. One fixed statement of purpose sits above all three and does not move at any point.
Three rungs, three different questions
Read the ladder as one ambition getting bigger and it looks unremarkable. Read what each rung is asking and the design shows.
Year one asks whether the thing works anywhere at all, under real conditions, with a number attached. The plan writes measurability into the goal line rather than deferring it to an appendix of indicators: the year's stated delivery is local, measurable digital innovation. Year one is not asking for reach. It is asking for one honest reading.
Year two asks whether the thing survives being operated by people who did not build it, in markets that did not ask for it. That is not a larger version of year one. A project that works because three motivated people are standing next to it does not fail at scale for lack of budget. It fails because the supply behind it was never built.
Year three asks whether the pieces behave as one experience from where the customer stands. Two capabilities can both run perfectly at scale and still produce an incoherent experience, because coherence is not a property either of them owns.
Three questions, three failure modes. Local proof fails on measurement. Scale fails on supply. Integration fails on ownership. That is the argument for why the order cannot be rearranged, and why compressing it relocates the failure rather than saving time.
Technology, content and process, at every rung
The plan refuses to record any stage as complete unless it produced an outcome in three layers. Technology, content, process. The same three at every stage, which is what stops the ladder from becoming a technology ladder with commentary.
The content layer is worth stopping on, because it is treated as a supply chain rather than a creative function. Year one produces guidance on good content for portals and other channels, plus a plan for non-branded external content and a solution to sourcing it. Year two produces a scaled internal and external library held globally with a process for localising it. Then it stops. The year-three content outcome is word for word the year-two outcome: the library is finished in year two and deliberately held constant while integration happens around it.
The process layer carries the plan's sharpest decision. The integrated global operating model for campaign management and analytics is due in year two. The integrated platforms are due in year three. The humans agree how to work across markets a full year before the systems are wired together.
Everyone I meet does this the other way around: integrate the systems, then spend a year discovering nobody agreed who owns the output, who may override it, or which function's number wins when two disagree. The wiring is the cheap part and not the part that fails.
Year one includes the unglamorous work
The year-one priorities are not the exciting ones. Stabilise a platform already in flight and deliver on its outstanding commitments. Run a set of named local projects, six of them. Enable indirect impact through published guidance. And, fourth of four, establish a process for managing future digital demand.
That fourth one is the item almost nobody puts on a slide. It is not a product. It is the intake mechanism for the requests that arrive once the first year works and everybody wants one. Building it before the demand exists is the difference between a portfolio and a queue nobody designed.
Year one also carries something labelled on the slide as interim: a stopgap campaign management solution across key channels, replaced by the at-scale platform in year two. Naming it interim in writing is the discipline. An unnamed stopgap becomes permanent architecture by default, and the year-two platform then arrives to find its job taken by something nobody chose.
The guidance work is the mechanism for reaching markets the centre does not control. Fourteen named guidance assets sit in the year-one publication backlog, spanning channel practice, design craft, field-force materials and supplier selection, sorted into four bands by state and priority rather than by topic. Two of those bands carry the lesson: completed, and to be reformatted. The plan separates content that exists from content that is usable, a distinction most data readiness reviews still fail to make before pointing a pipeline at a shared drive.
The portfolio is a diagram, not a list
The IT half of the deck is where the method becomes concrete. The prioritised technology portfolio is not a list of applications. It is drawn onto a single closed loop: insight and planning, fed by a master data repository, third-party sources and analytics; then campaign design and content strategy; then touch point design and channel execution; then the customer experience; then response and customer data flowing back into the data layer. A shared activity record sits in that data layer and feeds the execution channels, so what was planned and what was actually sent stay on one spine.
Seven numbered markers pin seven prioritised projects to their position on that loop. That is the whole trick of the page: each initiative is placed at the point in the cycle it repairs, so the portfolio reads as a system rather than as a queue of sponsors.
The printed order then encodes the dependency chain. The data and analytics foundation is first. Campaign orchestration is second. The regional and market portals, the content-management upgrade and the therapy-area programme, which are the items every sponsor actually wants, sit in the middle. Touch point integration is last of the seven. Integration is not the enabling move on that list; it is what becomes possible once the foundation and the channels underneath it exist.
One governance detail closes it. The five initiatives reviewed at that session each carry exactly one named individual as owner, walked through in sequence as a standing agenda item rather than as an escalation. Five initiatives, five distinct people, none co-owned and none owned by a committee. A portfolio of thirty pilots owned by a function is how thirty pilots survive without a single one being stopped.
What this changes for machine learning programmes now
Almost all of it transfers, because the failure I see most often in model deployment is not technical. A team proves a model on one process, at one site, with a clean extract somebody prepared by hand. The plan then calls for that model across twelve sites next quarter, it dies, and the post-mortem blames data quality. The real cause is that nobody built the supply layer: no shared feature definitions, no lineage anyone trusts, no owner for the number the model produces when two functions read it differently.
So we take the rungs literally, and a RealAI Consult engagement starts by asking which rung a client is actually on. Rung one is a measured result on one process, and measured means somebody wrote down in advance what would count. Rung two is that capability running where you are not standing, which needs a feature store, published definitions, and retraining and monitoring owned by a named person. Rung three is where several running capabilities are made to behave as one flow, and it is bought with an operating model, not with an integration project.
The early language model pilots crossing my desk sharpen this rather than change it. They stand up in days, so the first rung is now cheap enough to skip past, and skipping it is what produces the thirty-pilot portfolio with nothing to show. Cheap proof is proof only if you measure it.
Local proof fails on measurement. Scale fails on supply. Integration fails on ownership. Three different failures, so three different rungs, and no amount of budget converts one into the next.
- Three years
- One business objective per year, each a different kind of proof
- Four functions
- Digital, commercial operations, customer engagement and IT, steered by one plan
- 4 / 4 / 1
- Priorities specified in year one, year two and year three
- 14
- Guidance assets in the year-one backlog, sorted into four status and priority bands
The last rung is one sentence, and that is correct
Year one carries four itemised priorities. Year two carries four. Year three carries one line: integrate the scaled platforms and partner with commercial operations to operate localised experiences at scale.
That asymmetry reads at first like the deck ran out of time. It is not. A three-year plan that specifies its final year as precisely as its first is a forecast with dates on it, and it will be wrong in a way that is expensive to unwind, because people will have committed against it. Precision should decay with distance. The far rung needs a direction, an owner and a dependency, not a work breakdown, because what determines its shape is what the first two rungs taught you.
Which is the whole point of the ladder. Each rung is not just a stage of delivery. It is the instrument that tells you what the next one should be.
Read from an internal digital strategy and roadmap proposal at a global pharmaceutical company, presented to a governance forum. It is a planning artefact: it produced a proposed sequence, a prioritised portfolio and a target architecture, not delivered results. Counts and stage objectives are as printed there. Reading the sequence as a rule for machine learning programmes is ours.
“Local proof fails on measurement. Scale fails on supply. Integration fails on ownership. Three different failures, so three different rungs, and no amount of budget converts one into the next.”
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.