Skip to content
Hominis Agentic OS · early access program now openJoin the waitlist
RealAI

Case studiesLife sciences and healthcare

Case study
Life sciences and healthcareA global pharmaceutical company

Every initiative with a named sponsor sits below the shared work it depends on, and that is how you know the order is a dependency chain

A global pharmaceutical company published a prioritised list of digital technology initiatives against a single architecture picture. Read top to bottom, the order runs data foundation, then integration, then orchestration, then the market-facing and therapy-area applications, then platform maintenance. The test of whether that is a real dependency chain or a polite ranking is what happens to the items with the strongest sponsorship: each of the five high-impact initiatives had one named owner presenting it at the governance forum, four of those five reappear on the technology list, and every one of the four sits below the shared data and integration work rather than above it. The list carries eight lines against seven numbered positions on the architecture, because the first line names the outcome rather than a project, so we do not publish the count without saying so. This was a proposal and a roadmap. It reported no results, and its only day-counted commitments concern publishing guidance: a usable draft estimated inside 30 days, final publication inside 60.

8 lines, 7 markersPrioritised initiatives against numbered positions on the architecture
Client
A global pharmaceutical company
Duration
Strategy and roadmap proposal, presented to a governance forum
AI · RIDGE E86.4 N50ρmax 1.00
5 of 5High-impact initiatives with exactly one named owner
4 of 5Sponsored initiatives that reappear on the technology list, none of them above the shared layers
60 daysThe only unhedged day-count commitment in the plan

Almost every prioritised portfolio is read as a ranking. The item at the top matters most, the item at the bottom is next year's problem, and the order between them is a settlement among the executives who asked. Read that way, a portfolio tells you about an organisation's politics and nothing about its architecture, because a ranking has no arrows in it.

A global pharmaceutical company published its digital technology priorities in a way that survives a harder reading. The initiatives were listed beside a single picture of how commercial work flows, from insight and planning through campaign design into channel execution and back again as response and customer data, and each one was placed at the position in that flow it was meant to repair. Then the list was ordered: the data foundation first, integration second, orchestration third, the market-facing and therapy-area applications after that, and a platform upgrade last.

The challenge

The order looks unremarkable until you test it against the thing that reliably corrupts a portfolio, which is sponsorship. Elsewhere in the same session, five high-impact initiatives were walked through as a standing agenda item, and each had exactly one named individual presenting it. No committee ownership, no shared workstream, one person per initiative. Those five are the best-sponsored work in the plan by construction: somebody senior had agreed to answer for each of them.

Four of the five reappear on the prioritised technology list, and not one of them is placed above the shared work. The orchestration layer, a single market's integrated portal, a therapy-area digital programme and a next-generation portal personalisation effort take the four positions between the integration layer and the bottom of the list. Every one of them is placed after the data foundation and after integration. If the list were a sponsor ranking they would be at the top, since they are the only items on it with a named human whose year depends on them. Instead they are not merely low, they form an unbroken run at the foot of the list, which is the strongest evidence that something other than volume set the order. The fifth sponsored item, a content strategy, does not appear on the technology list at all.

The bottom of the list makes the same point from the other end. The final line is a version upgrade of an existing content platform. Upgrades are the most persistent recurring demand a technology function carries, and the easiest thing in the world to justify. Putting one last says it will not change what a customer experiences until the layers above it exist, so it waits behind the work that will.

There is a counting problem worth stating plainly. The architecture carries seven numbered positions; the prioritised list carries eight lines. The reading that reconciles them is that the first line names the outcome the whole picture produces, an integrated customer experience, and the seven beneath it are the initiatives that get you there. We would not publish the figure eight without that caveat, and the caveat is not a nuisance: a portfolio whose first line is the objective is refusing to let that outcome become one competing item among its own components. The section framing agrees, introducing the prioritisation as existing in support of that objective, so technology priority is derived from a business outcome rather than assembled from a backlog.

The approach

The mechanism is simple and almost nobody does it. Draw the flow once. Place every candidate on it. Then order by position in the chain rather than by expected value, sponsor seniority or delivery confidence.

Ordering by dependency has a property that ordering by value does not: it is falsifiable in the room. Ask what breaks if orchestration is built before the shared record it orchestrates against, and you get an answer in a minute: every channel holding its own version of the same customer contact, with no arbiter between them. Ask whether a therapy-area programme is worth more than a market portal and you get two business cases nobody can check for two years, argued by the people who wrote them.

The grouping carries the same discipline. The data foundation appears as one line covering a master data store, an analytics capability and the activity layer between design and execution, rather than as three separately rankable items. Split them and each becomes negotiable alone; held together, the plan admits that none of the three is useful without the other two.

An order like this does not hold by itself, and the plan knows it. One of only four priorities in the first year is not a product at all. It is establishing a process for managing future digital demand. That is what protects the chain, because the threat to a dependency order is never the plan as written, it is the well-sponsored request that arrives in month four and lands at the top of the list. Without an intake process the sequence is a document. With one it is a decision somebody can defend.

The same logic runs along the time axis. Scale is not attempted until local delivery has produced something measurable, and integration is not attempted until scale exists. The third year is specified far more loosely than the first two, four itemised priorities in each of the first two years against a single sentence for the third, because precision should decay with distance. And the first-year campaign management solution is labelled interim, with the scaled platform a year behind it. Writing that word is what stops a stopgap becoming architecture, which is how a dependency order breaks without anybody deciding to break it.

The outcome

This was a proposal presented to a governance forum. It reported no results, priced nothing and claimed nothing delivered, which is what a strategy and roadmap document should look like when tabled. The only commitments in it carrying a day count concern the published guidance: a usable draft estimated within 30 days, final publication to the operating functions and markets within 60. The wording is not symmetrical either. The internal draft is hedged as an estimate; the date the markets hear carries no hedge. Treat all of it as method rather than as evidence.

Years on, the ordering logic holds and the way we would build it has changed in two ways.

First, the dependencies were asserted rather than measured. Each arrow in a chain like this is a testable claim about data readiness: does the field exist, is it populated for enough of the population to matter, does it join to the identifier the next layer needs. Those checks turn a sequencing argument into preconditions with a pass or a fail beside each one, and they usually reorder at least one item. The activity record between design and execution is also the input to process mining rather than only to reporting: once contacts are written through one layer with a consistent identifier and a timestamp, whether the sequence held can be measured rather than surveyed at year end.

Second, a machine learning pipeline is itself a dependency chain with a deployment at the end of it. A model that scores the next best contact depends on features, which depend on a joined history, which depends on the foundation this list puts first. Deciding the portfolio order and deciding the model deployment order are one exercise, and in most commercial organisations we work with they are done twice, by two groups, in two sequences. The lineage trail belongs in that same layer: built in at the foundation it costs little, retrofitted across three channels it is a programme.

The current wave of cautious language-model pilots does not change the arithmetic. What it changes is the pressure on the intake process, because a pilot arrives with more sponsor energy than any other request in the building and lands, almost by definition, at the market-facing end of the chain. Our Platform work starts at the foundation for that reason, and Consult engagements start by drawing the flow and asking where each request sits on it.

Read the list as a ranking and it is a compromise. Read it as a chain and it is an argument that can be checked, made by someone willing to put the loudest sponsors below the plumbing.

NEXT STEP

Ready to make AI real?