The Layer You Cannot Buy

Two numbers from the same 300 people

BCG asked 300 chief marketing officers whether AI was driving end-to-end transformation of their function. 96% said yes.

BCG then sorted the same people by what they had actually deployed. 42% use generative AI only as an assistant on discrete tasks, in a handful of workflows. 26% have scaled beyond pilots in one or two areas. 32% are running agents inside redesigned workflows, and within that group only 8% have connected multiple agents to run anything autonomously.

BCG calls this the marketing transformation illusion. The tools are everywhere. The system underneath them was never built.

The prescribed fix is a four-layer stack: data, a brand intelligence layer, agent orchestration, and a single interface for marketers, plus the organisational work of new roles, flatter teams and systematic upskilling. The survey also reports that only 20% of CMOs have actually redesigned their structure. Even among the leaders, the technology is ahead of the organisation.

That is my thesis in BCG's own data, so I am not going to spend an essay agreeing with it. The more useful question is which of the four layers is the one that keeps the 42% stuck.

The layer that is not software

BCG describes the brand intelligence layer as the encoded rules and operating context that let probabilistic agents act like the company rather than like a generic model. It carries the messaging guardrails, the trusted sources, the KPIs, and in BCG's words, how the company's best marketers work.

A month earlier, in May, HBR published a piece by Michelle Taite, John Winsor and Will Fernandez that named the same thing a brand code: a machine-readable knowledge base holding brand strategy, customer insight and business rules in a form both people and agents can act on. They describe it as always-on onboarding documentation for humans and agents alike, and they call it the foundation layer.

Two sources, one month apart, converging on the same object. Neither of them is describing software. Both are describing the company's judgement, written down in a form a system can use.

Three of BCG's four layers can be bought. Data platforms, orchestration frameworks and marketer interfaces are products, and the build arms of the big consultancies will assemble them for you. The brand intelligence layer cannot be bought, because its contents are not for sale. They are inside the heads of the people who already do the work.

What it is called in pharma

In a consumer brand the layer is mostly voice, positioning and audience rules. In pharma commercial it is something harder.

It is the accumulated judgement of medical, legal and regulatory review. Which claims survive. Which references are acceptable behind which claim. How far a market may adapt global material before it needs its own review. What gets bounced, and why it gets bounced in Germany but not in Spain. The claims matrix is the visible part. The rest lives in reviewers' heads, in email threads, and in the comments on the last rejected proof.

It is not a brand intelligence layer. It is a compliance intelligence layer, and it has a structural problem the consumer version does not: nobody owns it.

Marketing does not own it, because marketing is the function being reviewed. Medical and regulatory do not own it as an asset, because they experience it as a workload. The agencies do not own it, because they are on the receiving end. It sits in the seam between functions, which is exactly where enterprise AI programmes stop.

BCG presents one of its findings as good news: marketing has won the mandate, and roughly half of CMOs now own AI investment in their function. I read it the other way. A function can transform itself. It cannot transform the seam it shares with three other functions, and in pharma the seam is where the value is.

I have built this layer twice

The first time was FIDo at Hogarth. The platform underneath was a configurable off-the-shelf product. It grew to 15,000 users across 50 markets, became the operating foundation of the business, and was one of the reasons the sale to WPP happened as fast as it did.

The software was not what scaled. What scaled was the set of rules we encoded into it: which adaptations a market was permitted to make, which approvals were required for which asset types, in what sequence, and who was allowed to sign off. Extracting those rules from the people who held them was the work. Configuring the platform was the easy part.

The second time was Workfront for a global pharma company, rolled out across 56 markets to more than 1,000 users. Again, the software was a product anyone could licence. The value was in encoding how review actually happened in each market so that a step could not be skipped, a reviewer saw only what was theirs to review, and a market could not reinvent the process on its own.

In both cases the tool cost less than the extraction, and the extraction was where the value lived. In both cases the extraction was organisational work, done with people who had never been asked to write down what they knew.

That is what the compliance intelligence layer is. It is not a database. It is the operating procedure of a review system, captured in a form that a workflow could enforce then and an agent can act on now.

What the build arms actually sell

The large consultancies now have build arms selling the agentic stack to CMOs, and the BCG report reads in places as a brochure for one. The engagement-based figures in it, 20% to 30% cost efficiency, a threefold increase in marketing ROI, a tenfold improvement in cycle times, come from BCG's own client work rather than from the survey, so I treat them as claims rather than evidence. The survey figures stand on their own.

The honest version of what a build arm can sell is narrower and more interesting than the brochure. It can bring the engineering for three of the four layers. For the fourth, it can bring an extraction method and the discipline to run it. It cannot bring the contents, because the contents belong to the client's reviewers. The pitch that would hold up in a pharma boardroom is: we will build the stack, and you will have to surface the work.

Most buyers hear the first half. That is the illusion BCG measured, from the other side of the table.

The obvious rebuttal

The rebuttal is that you can simply prompt the model with the brand guide and the SOPs, and it will behave.

The brand guide is the deck. The judgement is the system. A document that says "all efficacy claims must be referenced to the SmPC" tells an agent nothing about why the same claim passed in one market and failed in another last quarter. The reviewers know. Nobody has asked them, in a form that can be captured, because the asking does not sit in anyone's job description.

That is the 42%.

Where this leaves pharma

The four claims this site is built on say that pharma runs pilots that never scale, that most of the sector sits at stage one or two of four, that shared production infrastructure beats a prototype shop per brand team, and that the sector is behind a services economy being rebuilt on this logic now. This essay files under the first.

Pilots in pharma commercial do not fail because the model is wrong. They fail because the agent produces more things to reject, faster, and the function that could fix that does not exist. The layer that would fix it cannot be bought.

The capability to build it can be bought, and almost nobody in pharma is buying that, because nobody owns the seam it sits in.