Every week, another brand team asks which AI tool to plug into their content workflow. Which prompt library, which vendor, which use case. It is the wrong question.
The right question is what the AI is actually reading. If your claim library is spread across 14 PowerPoints in three SharePoint sites, generative AI will confidently surface the wrong claim. If your design system has hex codes but no semantic tokens, it will apply the wrong colour to the wrong audience. If your knowledge hub is a Teams channel plus a shared drive, the model will hallucinate on top of your own inconsistency.
A 2025 MIT study found that nearly 95% of enterprise generative AI pilots failed to deliver measurable business impact, most often because systems stayed disconnected from real workflows and data foundations. In pharma, where MLR sign-off is not optional, that disconnect is not a productivity issue. It is a compliance exposure.
Preparing pharma content for AI is a content architecture project, not a tooling project. Before any prompt engineering, three layers need to be in place.
The first is a validated claim repository with sources visibly attached to every statement. The second is a structured knowledge hub the model can actually query, with clear taxonomy, product boundaries, and market scope. The third is a design system built on semantic tokens that carry meaning, not just colour values.
Promedia's work on the MLR process and modular content shows why the pre-submission stage is where the real bottleneck sits. Approved content is an asset. A modular content system lets teams reuse validated claims and building blocks instead of starting from zero every time, and pre-marking sources inside the document saves reviewers hours before the cycle even starts. AI on top of that base compounds the gain. AI on top of scattered decks compounds the risk.
Regulators are already moving. In March 2025, the EMA and FDA jointly defined ten guiding principles for good AI practice, stressing risk-based validation, data quality assurance, and continuous monitoring. The EU AI Act's general-purpose obligations took effect on 2 August 2025, with high-risk provisions applying from August 2026, and both directly affect pharma content workflows.
A validated claim repository is a single, structured source of truth for every promotional and non-promotional statement a brand can make, with the supporting reference, the approval status, the market, and the expiry date attached to each claim.
It is not a Word document. It is not a shared folder called ClaimsFinalv3. It is a queryable database where each entry carries:
Without this, generative AI has nothing reliable to retrieve. It will confidently generate plausible-sounding claims that no medical reviewer has ever seen. With it, retrieval-augmented workflows can draft copy that already sits inside approved boundaries, and MLR flags surface before submission rather than during it.
As Promedia's work on pharma-specialist versus generalist agencies makes clear, no one outside the discipline flags an unsourced claim, knows the difference between promotional and non-promotional content, or has heard of an SmPC or a PIL. The claim repository encodes that discipline so the AI layer inherits it.
Content architecture for AI readiness has three pillars: claims, knowledge, and design. Each needs to be structured, versioned, and queryable.
On the design side, semantic tokens are the missing layer for most brand systems. A semantic token is not a colour. It is a role, such as colour/content/secondary or brand/warning/background, that maps to a value. When the token changes, every instance updates. When AI generates a new asset variant, it reaches for the role, not the hex.
Promedia's audit of a multibrand email system showed what this unlocks in practice. Two token changes fixed 90% of all findings. An accessibility score of 33 out of 100 reflected the same two token values repeated hundreds of times across instances, not deep architectural problems. Updating colour/content/secondary and the red brand token was projected to raise the accessibility score above 75 and push overall system health past 90. The architecture was solid. The issues were targeted, fixable, and only visible because the system was structured enough to audit at scale.
Co-building the token layer with AI shortens the setup. Drop a PDF or Figma file, let AI extract colours and typography as primitive tokens, generate the ramps, and cover the brand states. Designers then shape the naming layer: primary, secondary, states, a shared language for design and development, which can be replicated across brand variants automatically.
Hallucinations in pharma content workflows are almost never a model problem. They are a retrieval problem. When the model has no clean source to pull from, it fills the gap with statistical plausibility, which reads as confident, well-formatted, and wrong.
The three most common causes are the same three foundation gaps: unsourced or duplicated claims, unstructured knowledge repositories, and design systems without semantic meaning. Fix those, and the model has somewhere real to ground its output. Skip them, and every prompt becomes a coin flip that a medical reviewer has to catch.
By 2026, the question is no longer whether pharma teams will use generative AI, but whether they will use it inside a workflow that compliance, legal, and MLR can actually defend. Governance is the moat.
At minimum, three things: a validated claim repository with linked sources and approval metadata, a structured knowledge hub with clear taxonomy and market scope, and a design system built on semantic tokens rather than raw hex values. Without these, generative AI amplifies existing inconsistencies rather than resolving them.
No. Every regulator moving on AI, including the EMA and FDA joint principles issued in March 2025, treats human oversight and continuous monitoring as non-negotiable for regulated content. AI can accelerate drafting, surface MLR flags earlier, and reuse approved modules, but final medical, legal, and regulatory sign-off remains a human decision.
A semantic token is a named role, such as colour/content/secondary, that maps to a specific value. It carries meaning, not just a hex code. For pharma brands running multiple products across markets, semantic tokens let one update cascade through every asset, keep accessibility scores auditable, and give AI tooling a reliable vocabulary when generating brand-compliant variants.
A modular system lets teams reuse pre-approved claims and building blocks instead of drafting from scratch, so reviewers see fewer novel statements per submission. When sources are pre-marked inside the document, reviewers no longer spend hours tracking down references. The result is fewer review rounds, fewer compliance flags, and faster approval, because the MLR cycle stops being where the whole process backs up.
Do not start with the AI tool. Start with an honest audit of where claims, knowledge, and design tokens actually live inside your organisation. If any of the three sits in a shared drive or a Teams channel, that is the first project, not the second.
The teams currently getting value from generative AI in pharma are not the ones with the cleverest prompts. They are the ones who spent six months cleaning up their content architecture first, so the AI layer had something structured to read.
If you are mapping out where to start, Promedia's team works with pharma brand, medical, and regulatory leads on exactly this foundation, before any generative layer goes on top.
One short email when we publish. No spam, no sales pitch. Unsubscribe whenever.
