Every integration programme I am asked to review carries the same hopeful assumption somewhere underneath it. Join the systems, and the people using them will work out how to work together. Wire the campaign platform to the analytics platform, land both on one customer record, and a shared way of operating will follow from the shared plumbing.
It does not follow. What follows is a set of markets, each with a working integration and an unchanged habit, arguing about whose definition of a campaign the joined system should have been built around. The technical milestone is met. The reason anybody wanted it is still open, and now it is open inside a system that is expensive to change.
A plan I keep returning to puts the two in the other order, and appears to do it deliberately. 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 stratified into three labelled layers, technology, content and process, and the layers carry across all three stages.
Read down the process layer and the sequencing shows itself. The integrated global operating model for campaign management and digital analytics is a second-stage outcome. The integrated platforms are a third-stage one. The plan has the humans agreeing how to work across markets a full year before anybody promises to join the systems they will be working in.
The cheap half and the expensive half
Most plans treat integration as the hard part and the operating model as the thing that gets sorted out during rollout. The cost profile says the reverse.
Joining two campaign systems is bounded work. There are connectors, a schedule, a test plan and a date, and when it is late you can see that it is late. Getting every market to mean the same thing by campaign, by response, by audience, by contact is unbounded, political, and invisible when it fails. There is no build server that turns red because two commercial teams have quietly kept their own definitions.
So the unbounded work gets scheduled last, where there is no room left, and gets renamed. It appears on the plan as change management, or as adoption, or as training, which are all words for arguing about definitions after the cost of changing your mind has gone up by an order of magnitude.
The plan I am reading does not do that. It gives the unbounded work its own stage, its own outcome line and its own year, and it puts that year first. Whether the organisation held to it I cannot say. That it chose to draw it that way costs nothing to copy.
What an integration silently decides
An integration is not a neutral act of connection. It is a decision about whose definitions win, and it is made by whoever configures the mapping.
When two systems are joined, somebody has to say what a record means on both sides. Does a campaign end when the last message goes out or when the response window closes. Is a market's own follow-up a new campaign or the same one. When a customer appears in two markets, which one counts the interaction. Every one of those questions has a defensible answer and no correct one, and the joined system needs a single answer to function at all.
If the operating model has been agreed, the mapping implements a decision people made in a room and can point at. If it has not, the mapping makes the decision instead, in a configuration file, on a Tuesday, by whoever had the ticket. Nobody signs it. Nobody is told. Markets discover it months later when their numbers stop matching the numbers they used to report, and the reasonable conclusion they draw is that the new system is wrong.
That is the moment the programme starts spending its second budget. Not on plumbing, on persuasion.
What the earlier stage actually has to produce
An operating model outcome sounds soft until you list what it owes the stage after it. In our words rather than the plan's, it owes at minimum: one definition for each counted thing, with the market variations named as variations rather than left as local truth; one owner per number who can be asked why it moved; a rule for who may start and stop a campaign in a market and who may overrule that; a declared home for the audience list and a declared home for the response data; and a route by which a changed definition reaches everyone who reports on it.
None of that requires a platform. All of it requires decisions and a written record, which is why it can be finished a stage early, and why so few programmes bother.
It also happens to be the raw material for everything that comes later. Lineage is only meaningful once a field has one agreed meaning to trace back to. A shared measure across markets is the only thing that makes a comparison honest. MLOps practice can tell you which version of a pipeline produced a number and cannot tell you what the number means. The meaning is upstream, and it is written by people, in the year that plan gives them. Fixing that meaning before anything is served on top of it is the opening block of a RealAI Platform engagement, for the same reason the roadmap put it a stage early.
Why this reads differently now than when it was drawn
The page was drawn for campaign management and digital analytics. What I am asked about today is retrieval-augmented assistants over commercial content, copilots for market teams, a vector store of approved material, and a few carefully scoped autonomous experiments at the edges. The sequencing problem got worse, not better.
A retrieval assistant answering questions across markets is an integration wearing friendlier clothes. It reaches into several stores of material and produces one fluent answer. If two markets mean different things by the same field, the assistant does not surface the disagreement. It resolves it, invisibly, in favour of whichever document ranked highest, and returns a confident sentence that is right in one market and wrong in another. A joined database at least throws a duplicate key. Retrieval throws nothing.
The evaluation problem lands in the same place. An evaluation set is a list of questions with agreed right answers. If the organisation has not agreed what right means across its markets, the set cannot be written, and what gets built instead is a set that encodes one market's answers and quietly grades every other market as wrong. Teams read that as a model quality issue and spend months on the model.
There is a governance reason to care about the date as well. The European AI Act reached political agreement last month, and whatever the final text says about documentation, purpose and human oversight, the material that satisfies it is operating model material. Who decided, on what basis, who may override, where the record lives. An organisation that scheduled that work into an early stage will be writing it down. An organisation that scheduled it as adoption will be reconstructing it from memory.
An integration is a decision about whose definitions win. Make that decision with plumbing and you have not made it at all. You have let whoever configured the system make it on everyone's behalf, quietly, and without a record.
What I would take from it
Three moves, and only the first is difficult.
Put the operating model on the roadmap as a dated outcome with an owner, in an earlier stage than the integration that depends on it. Not as a parallel workstream, not as a governance track, as a deliverable with a date that something else is allowed to be blocked by.
Let the interim tool be interim on purpose. The plan schedules a stated interim campaign management solution as its first-stage process outcome, ahead of the agreement work rather than instead of it, which is a more honest arrangement than pretending the permanent platform is nearly ready. An interim tool that everyone knows is interim gets replaced. A temporary tool nobody labelled becomes the estate.
Sequence the integration after the agreement and say out loud which half is cheaper. The connectors are the cheaper half. They always were. The reason integration programmes overrun is that the expensive half was scheduled inside them without being named.
Read from an internal digital strategy proposal at a global pharmaceutical company: a three-stage operational roadmap whose outcomes are stratified across technology, content and process. It is a planning document, and it states intent on a date rather than reporting a result. Which layer sits in which stage is the plan's; the argument about why that order is the expensive one to reverse is ours.
“An integration is a decision about whose definitions win. Make that decision with plumbing and you have not made it at all. You have let whoever configured the system make it on everyone's behalf, quietly, and without a record.”
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.