More than a decade ago, a global pharmaceutical company running commercial digital across its markets put a three-year digital strategy and roadmap to its digital leadership and its partner functions in commercial operations, customer experience and IT. One of the four priorities named for year one was indirect impact, and the deck said plainly what indirect meant: guidelines and best practices. The instrument was a publication roadmap of twelve guidance assets covering search, media, social, mobile, email, design and sales-force detailing, each sorted into a status or priority band, all set inside a single calendar year.
The challenge
The stated purpose sat on patients rather than on marketing throughput: digital innovation that improves patients' lives and the healthcare experience. Reaching it required four functions to move as one. Digital strategy, global commercial operations, customer experience and IT each held a piece of the outcome, and none of them held all of it.
Underneath that, markets ran their own digital priorities. The deck nowhere claims the centre could direct them. It says the guidance would be used to inform the markets' own digital priorities, which is a different verb, and it files the whole guidance stream under indirect impact. Central digital had a vision, a platform that year one was committed to stabilising, and a route to the markets that ran through influence.
The roadmap answered by sequencing three years so the unit of value changed each year. Year one was local, measurable, high-impact projects. Year two was global customer impact at scale. Year three was an integrated customer experience at scale. Scaling was deliberately gated on local proof, and integration was gated on scale. Year one carried four priorities and named six high-impact local projects, and one of those four priorities was not a product at all. It was a process for managing the digital demand that would arrive later.
The approach
Every horizon was resolved across three layers: technology, content and process. Nothing was allowed to count as complete because a platform shipped. The clearest expression of that discipline sits in the sequencing. The integrated global operating model for campaign management and analytics was a year-two outcome. The integrated platforms were a year-three outcome. The people agreed how to work across markets a full year before the systems were wired together.
The guidance portfolio was the mechanism for indirect impact, and it was managed like a product backlog rather than a document folder. Twelve assets, sorted into four status and priority bands, with two thematic clusters called out across the set. The bands mattered more than the topics. One band held finished assets. Another held assets that existed but were in the wrong format and needed reformatting before they could travel. That single distinction, content exists against content is usable, is the one most knowledge programmes skip. At least one asset was benchmark-based, built on measured channel performance rather than opinion, so the guidance was expected to carry evidence.
Distribution was designed with the same care as the content. The guidance travelled three ways: published to markets, fed into each market's own digital prioritisation process, and adopted by the partner functions in commercial operations and IT. The strategy group also held a seat at the steering forum. Two dates were attached, and the wording of the two is not identical. A usable draft for the partner function was estimated within 30 days. Final publication to that function and to the markets was committed within 60 days, with no hedge. The rough artefact was estimated; the promise to markets was made.
The plan then closed with a question back to the room, asking what other guidance the consuming functions would need and what format would work best. Format was treated as a first-order design decision rather than a production detail.
When the plan was written, the answer to that question was a format a busy market manager would open. The same question now has a second reader, and the second reader is a model that has to see the slide rather than only the words on it.
Governance was drawn as six functions, each with a stated mandate, and one capability cutting across all of them. Advanced analytics was given two jobs, not one. The first was the usual performance job: insight at group, region and market level through dashboards, measurement and target tracking. The second was strategic conformance and course correction, including content utilisation and market case studies. Analytics was asked to measure whether the strategy was landing and whether published content was actually being used. That is a materially different mandate from campaign reporting, and the design named it explicitly.
The technology portfolio followed the same logic. Candidate projects were mapped onto a single closed customer-experience loop with seven numbered stages, running from insight and planning through campaign design and content strategy to touch point design and channel execution, with response and customer data flowing back into the data layer. The priority order follows the dependency chain rather than the topic list: data foundation first, then touch point integration, then global campaign management, then the market and therapy-area applications. Five high-impact initiatives were set out for walk-through, each with exactly one named owner in the table. One name per initiative, no shared accountability.
- 12
- Guidance assets in one year
- 60 days
- To final publication
- 7
- Stages in the experience loop
The outcome
This was a proposal, and it should be read as one. The deck records a design and a set of commitments, not measured results. What it delivered on the day was a sequenced three-year plan, a guidance backlog with states and a publication clock, a prioritised technology portfolio ordered by dependency, and an operating model in which one cross-cutting capability was accountable for checking whether any of it was being followed.
The reading today is where it earns its place. That guidance library was, without anyone calling it that, an early enterprise knowledge base. Twelve assets holding the company's accumulated judgement about how to run digital work, written for humans to read at the moment they needed it. The failure mode was always the same: the market manager did not know the asset existed, or found it after the decision was made.
Agentic AI removes that failure mode, and it removes it by reading rather than by reminding. The same twelve assets become the grounded corpus an agent works from, and the guidance arrives inside the brief, the agency review and the campaign approval rather than in a portal someone has to remember. The band labelled "to be reformatted" is the tell, and it is not a text problem. What a commercial digital function knows about search, media, social, mobile, email, design and sales-force detailing sits in slide layouts, annotated screenshots of a good detail aid, marked-up banner examples and a channel benchmark that only means anything next to the chart it is drawn from. Computer vision is what gets that out, because a model has to read the annotation on the image and the callout on the slide, not a stripped text dump with the pictures missing. Pointing a pipeline at a shared drive still does not fix it.
The conformance mandate ages even better. The instruments named for that job at the time were content utilisation reporting, market case studies and service enhancement pointers, and all of them arrive after the decision they describe. The artefacts they were describing were visual: an e-detailing screen, a banner set, an email template, a design layout, each checked by a person opening it and comparing it against a guideline held in their head. An agent with the published standard on one side and vision over the artefact on the other checks every asset before it ships, logs what it compared and why, and sends only the exceptions to a named human. That is the autonomous half of the design, in the plain sense that nobody starts it: conformance stops being a look backwards and becomes a gate that runs on its own inside the workflow. The analytics function asked to answer "is the strategy landing" can answer it continuously, per market, per asset.
Put those two together and the seven-stage experience loop stops being a picture of who hands what to whom. Insight and planning, campaign design, content strategy, touch point design, channel execution, response, and customer data flowing back into the data layer were each a queue with people waiting in it, which is precisely why the plan had to agree the operating model a year before it wired the platforms together. Run those stages as interconnected automated pipelines with agents as the workers and the handovers stop being the constraint. A market brief pulls against the guidance corpus, a draft comes back, the conformance gate reads the artefact, the exception goes to its owner, the approved asset publishes to the market, and utilisation data lands in the same place the next brief will read from. Agents carry a real share of how the function runs, and the humans keep the parts that were always theirs: what to promote, in which market, to whom, and whether the standard itself is still right.
Two cautions carry forward. The first is sequencing. Process integration was scheduled ahead of platform integration for a reason, and enterprises wiring agent frameworks together before agreeing who owns the output are repeating the inversion the plan avoided. The second is ownership. A 60-day publication clock made sense for a static library. A corpus that agents read continuously has no publication date, so the discipline shifts to an owner and a review date per asset.
The three-year shape is the part that would change most. That plan spent its first year proving locally because proving anything cost a year of people, and it gated scale on local proof for the same reason. Loop and harness engineering changes that arithmetic. The loop is the part that drafts against the corpus, tests its own output against the standard and comes back round; the harness is the part that fixes which markets, which systems and which records it may reach, under whose identity, with every step written down. Neither is a research problem. Both are build work, and build work does not need a horizon to sit in. Scaling judgement without scaling headcount is now the straightforward half. Keeping that judgement correct, owned and current is the work.