Pharma is not starting from pages.
Large pharmaceutical organisations already have sophisticated modular-content systems: approved claims, reusable modules, references, version histories and rules governing where content can appear, for which audiences and in which combinations.
That is a considerable head start for an AI-mediated web. But it raises a question: what were those structures designed to survive?
In the last post, I framed part of this as a content-design question: relationships between claims, conditions and evidence should not depend on visual proximity that can disappear when a machine extracts text. I now think part of the problem sits one layer deeper. Some of those relationships may not be expressed in the content at all. They may exist as rules in the systems governing it.
Modular content was largely designed for reuse inside an environment the organisation controls. A module is approved. It enters an assembly system the company owns. Rules govern its use. The resulting asset follows the appropriate review and publishing process.
Generative retrieval is a different kind of reuse.
Approved content is published or indexed. A retrieval system divides and selects it. A fragment may be placed alongside other fragments. A model then uses what it has retrieved to construct an answer.
The content has not changed. The execution environment has.
Approval sometimes belongs not simply to a piece of content, but to that content under particular conditions of use.
Imagine an approved efficacy claim that is valid only for a specific patient population. The claim exists as one content object. Its required population qualifier exists as another. Inside the CMS, a rule connects them and ensures they travel together.
Now imagine an AI index receives both as independently retrievable text.
It can retrieve the efficacy claim without inventing a word. The claim is authentic, current and approved. But if the condition governing its use does not travel with it, something important has changed.
The words survived. The approval boundary may not have.
More than a chunking problem
Retrieval systems losing context is not a new problem. Chunking, document structure and context preservation are established concerns in RAG systems.
But losing context and losing governance are not quite the same thing.
A better chunking strategy may keep a heading attached to its paragraph, preserve a table or stop a sentence being separated from the text that explains it. It cannot preserve a rule that was never contained in the text being chunked.
If a CMS knows that Claim A requires Qualifier B, or that Module C is permitted only for Audience D, that relationship may exist perfectly well inside the governed content environment without being visible in the text handed to the retrieval system.
You cannot chunk your way to preserving an approval boundary that was never expressed in the content in the first place.
Restricting AI retrieval to approved content is sensible and, in regulated environments, probably essential. But it solves a different problem.
Restricting the corpus determines which content is eligible to enter an answer. It does not necessarily determine which conditions must travel with that content.
A retrieval system could contain nothing except approved modules and still retrieve an approved claim without its required qualification. Every ingredient has passed the eligibility test. The resulting use of those ingredients has not necessarily passed anything.
The module may be approved. The retrieval event isn’t.
Metadata is not governance
Modern content systems can hold rich metadata and relationships: market, audience, indication, approval state, version, expiry, references, required companions and permitted combinations.
An AI index may ingest some or all of that information. But metadata that travels with content is not necessarily metadata that governs retrieval.
Suppose an indexed module contains:
Audience: HCP
That tells the retrieval system something about the content. It does not necessarily constrain what the system can do with it.
A rule such as:
Do not retrieve for non-HCP interactions
is different. Now the audience changes retrieval eligibility.
The distinction applies equally to market, indication, approval state, expiry and relationships between modules. Having that information in the index is useful; whether the information actually constrains retrieval is a separate question.
So “we ingest the metadata too” is not quite the end of the conversation.
Four checks worth making
None of these requires a new platform or governance programme. They are questions that can be asked of an existing integration.
What does the index actually receive? Is it indexing rendered text, individual content objects, or the underlying CMS records with their metadata and relationships?
Which of those relationships actually affect retrieval? Audience: HCP in an indexed record is descriptive unless audience changes whether that record is eligible to be returned.
What happens when the source changes? If a module is withdrawn, superseded or its approval state changes, when and how does that propagate to the index?
Can the indexed representation outlive the governed source? Could an older chunk, embedding or cached representation remain retrievable after the source record has changed?
The answers may show that the relevant governance already travels very well. If so, good. The useful thing is knowing where the boundary actually sits.
An assurance question, not a risk claim
None of this demonstrates that pharmaceutical AI systems are currently producing unsafe answers from approved modular content.
Well-designed implementations may already preserve the relationships that matter. Many modules may also be sufficiently self-contained that the distinction rarely matters in practice.
Where approval depends on conditions of use, however, a new retrieval environment creates a straightforward assurance question:
Does the retrieval system preserve what made that content permissible to use?
That is different from asking whether the content itself was approved.
Pharma’s modular-content infrastructure was designed around an important assumption: the organisation controlled the reuse environment.
AI weakens that assumption.
Internally, companies are beginning to connect retrieval systems directly to governed content repositories. Externally, general-purpose AI systems retrieve and reconstruct information from published sources in environments the publisher does not control.
The situations are different, but both expose the possibility that some of the meaning of regulated content lives not only in the content object, but in the relationships and boundaries around it.
From governed content to governed retrieval
Pharma has become good at governing content objects: claims, modules, references, versions and approval states.
Generative retrieval does not make that work obsolete. It makes that structure more valuable. But when content leaves the CMS, there is another question to ask:
What governance travelled with it?
The text, its identifier and descriptive metadata may be enough.
If some of a module’s validity depends on rules about audience, indication, combination, qualification or lifecycle state, it may not be.
Validity established under one reuse regime does not automatically transfer to another.
Sometimes the qualifier is not metadata about the claim. It is part of the claim.