A digital portfolio usually arrives as a list. Eight or ten initiatives, a sponsor beside each one, a rough sequence that reflects who asked most recently and how loudly. Reading it tells you about the organisation's politics and almost nothing about its architecture, because a list has no shape and therefore no dependencies.
A global pharmaceutical company presented its commercial digital portfolio a different way. It drew one closed loop, put seven numbered positions on it, and placed each committed initiative at the position in the loop it was supposed to repair. The list of projects still appeared on the slide, but it was subordinate to the picture. That inversion is the whole argument, and it is worth taking seriously because the loop it draws is the same loop every organisation now wants a machine learning model to sit inside.
The challenge
The loop has three stages and a return path. Insight and planning comes first, fed by a master data store, third-party sources and an analytics capability, and resolving into research, insight and brand planning. Campaign design and content strategy comes second, with campaign management and its own data and analysis feeding it. Touch point design and channel execution comes third, and it lands in three execution channels: a field engagement platform, a web content platform, and a bucket labelled other channels. What comes out the far end is the customer's experience.
Then the loop turns. Campaign response data, service data and customer data flow back into the data layer, drawn from the operational and third-party systems that hold them. The diagram spends real space on the return path, which is unusual. Most target architectures of this kind are outbound only: design, publish, send, and then a reporting box off to one side that nobody wires back into planning.
The problem the picture exposes sits at the join between stage two and stage three. Three execution channels means three systems that can each record what was sent, to whom, and what came back. Each of them is perfectly capable of keeping that record, and each of them keeps it in its own vocabulary, on its own clock, against its own identifier for the same person. Nothing in the diagram forces them to agree. If you let that happen, the return path does not close the loop. It delivers three partial accounts of the same activity into a planning stage that now has to reconcile them before it can plan anything, and reconciliation of that kind is never done once. It becomes a standing quarterly exercise that consumes the analysts who were hired to find something new.
So the interesting element on the slide is not any of the three channels. It is the layer drawn between campaign design and channel execution: one activity store that every channel writes through rather than around. One place that knows what was said, to whom, through which channel, and what came back. The channels remain different. The record of activity does not.
The approach
Calling that layer a spine is not decoration. It carries load in two directions. Downstream it is what makes three channels one campaign, because a plan expressed once is executed three ways without being retyped three times. Upstream it is what makes the return path arrive as a single history rather than three feeds that have to be matched. Remove it and both directions fail at once, which is why it cannot be deferred to a later phase without quietly deferring everything that depends on it.
The portfolio ordering follows from that. The data foundation comes first: the master data store, the analytics capability and the activity layer, grouped as one line rather than three because none of them is useful alone. Touch point integration comes next, since a spine that no channel writes through is a database with good intentions. Global campaign management comes after that, because orchestration is only meaningful once there is a single record for it to orchestrate against. Only then do the market-specific and therapy-area programmes appear, and platform maintenance sits last. The order encodes what has to be true before the next thing can work, which is a different exercise from ranking initiatives by expected value.
One honest caveat on the counting, because the slide invites a mistake. There are seven numbered markers on the loop and eight lines in the portfolio box. The first of those eight lines reads as the name of the outcome the whole loop produces rather than as a discrete project, which is the reading that makes the numbers reconcile. Seven initiatives placed on a loop, under one umbrella objective. We would not publish the figure eight without saying that.
The same discipline runs through the years. The plan sets three business objectives over three years, one per year, and the unit of value changes at each step: a delivered project, then a capability operating across markets, then the customer's experience across channels. In every year the same three layers are tracked, technology and content and process, so nothing is allowed to scale on technology alone. A campaign platform without a content model to feed it, or without an operating process to run it, produces a licence renewal rather than a capability. Four functions were named as joint owners of the roadmap, which is the organisational version of the same point.
The outcome
This was a proposal. It priced nothing, reported nothing, and made no claim to have delivered anything, which is exactly what a strategy and roadmap document should look like at the point it is presented. The only dated commitments anywhere in it concern the published guidance: a usable draft estimated inside 30 days, final publication to the operating functions and markets inside 60. Everything else is a plan across three years and a target architecture drawn against it. Treat it as method rather than as evidence.
What holds up, the better part of a decade later, is the spine. What has changed is that the case for it is now arithmetic rather than aesthetics.
Three things we would do differently. First, the activity record is the natural input to process mining, not just to reporting. Once every channel writes through one layer with a consistent identifier and a timestamp, the actual sequence of customer contacts is already in the log, and the difference between the campaign as designed and the campaign as executed can be measured instead of surveyed. Second, that same layer is where lineage belongs. A model that scores a customer for the next best contact has to record which features it saw and which decision it wrote, and in a regulated commercial setting that trail is not optional. Building it into the spine costs little; retrofitting it across three channels costs a programme. Third, the spine is the only sane place to put a feature store. Features derived from contact history are the ones every model in this loop wants, and deriving them three times from three channel exports is how two teams end up disagreeing about what a customer did.
The early language-model pilots now appearing in content operations do not change this arithmetic either. A model that drafts channel-ready copy well still needs the plan it is drafting against and the record of what was already sent. Without the spine it drafts into a vacuum and a human reconciles the output against three systems by hand, which is a demonstration rather than a capability. Our Platform work starts at the activity layer for that reason, and our Consult engagements start by asking whether one exists, because the answer determines whether anything downstream is worth sequencing yet.
Seven positions on a loop is a good picture. The loop closes on one of them.