Aspecialty pharmaceutical company with commercial operations across the US and multiple international markets — including the EU, Australia, and Canada — generates promotional content at scale across every channel and every geography. Each piece, regardless of market, must clear MLR review before it runs.
The company was using Jasper, Adobe GenStudio, and Writer across all of its markets. These tools generate US-optimized content by design — calibrated to FDA rules, structured around FDA conventions. In every other market, that content failed regulatory review.
The EFSA claims register, the FSANZ framework, Health Canada’s permitted claims list — none of these constraints were known to any of those tools. Local marketing teams were performing the regulatory adaptation manually, market by market, campaign by campaign.
The structural problem.
The deeper issue is that pharmaceutical advertising operates under rules that generic AI tools are not built to respect. Every promotional claim must satisfy fair balance — risks presented with equal prominence to benefits — and must use indication language drawn verbatim from the FDA-approved label.
When tools designed for persuasive marketing language are applied to drug promotion, the output fails structurally, not stylistically. Consider what those tools produce:
Give your patients more time. This treatment delivers results when it matters most.
The MLR committee — the internal Medical, Legal, and Regulatory reviewers who must approve every piece before it runs — is not editing these drafts. They are rebuilding them from scratch.
At a 23% first-draft pass rate, the committee had become a structural copywriting function rather than a regulatory oversight one — and content that should clear review in two weeks was taking six.
The data constraint.
A further constraint: the MLR submission history, reviewer comments, and approved content archive are among the most sensitive data a pharmaceutical company holds. Routing them through a third-party API is not a viable option, legally or operationally. Any solution would need to run entirely within the company’s own environment.
Coral Tree built a content generation system that produces first drafts structured around FDA compliance from the start — fine-tuned on the company’s own data and deployed entirely within their infrastructure.
The training inputs.
The system is trained on two sources the company already owns:
- The approved prescribing information.
The full FDA-approved label for the drug — containing the exact indication language, the permitted clinical evidence references, and the required safety information hierarchy. This is retrieved at generation time via RAG, ensuring that when the label is updated, generation immediately reflects the new language without requiring a full model retrain.
- The MLR review history.
Every piece of content ever submitted for review, annotated with outcome: approved as-is, approved with specific edits, or rejected with reviewer reasoning attached. This history encodes the practical knowledge of what this specific MLR committee actually requires — knowledge more nuanced than the general FDA rules alone, and that no external vendor has access to.
Fine-tuned on what the company already owns.
The system is trained on two proprietary sources: the FDA-approved prescribing information (retrieved at generation time via RAG so label updates propagate instantly) and the full MLR review history — every submission ever made, annotated with the reviewer's decision and reasoning. This history encodes the practical knowledge of what this specific committee actually requires, which is more nuanced than the general FDA rules alone and unavailable to any external vendor.
Compliance from the first token.
When the marketing team submits a brief — a patient-facing social post, a sales rep leave-behind, a digital display ad — the system retrieves the relevant sections of the prescribing information and generates a draft using the exact approved indication language for the drug's indicated population, presenting risk information at appropriate prominence relative to the benefit claims made, and avoiding the specific phrasings the company's own MLR history shows consistently get flagged. The MLR committee receives a draft that is structurally compliant and requires editorial review — not a structural rebuild.
Self-hosted, with a feedback loop.
The fine-tuned model runs entirely within the company's own cloud environment. The MLR history, reviewer comments, and approved content archive never leave their infrastructure. Every draft produced by the system is logged with the inputs used to generate it, satisfying the defensibility requirement if the content is ever subject to external scrutiny. And every MLR review outcome — approved as-is, approved with edits, rejected with comments — is captured and appended to the training dataset. The model retrains periodically on this growing archive of real review decisions, making future drafts progressively more likely to pass on first submission.
Use the exact approved indication language for the drug's indicated population
Present risk information at appropriate prominence relative to the benefit claims made
Avoid the specific phrasings the company's own MLR history shows consistently get flagged
The MLR committee receives a draft that is structurally compliant and requires editorial review — not a structural rebuild.
The numbers tell part of the story.
The improvement from a 23% to 78% first-draft pass rate is the headline figure — but what it represents operationally deserves equal attention. The MLR committee went from spending the majority of its time rebuilding non-compliant drafts to spending the majority of its time performing genuine regulatory oversight of content that is already structurally sound.
The reduction in review hours per content piece follows directly. A committee no longer performing structural copywriting on the majority of submissions moves faster, clears more content within the same review cycle, and ceases to be the bottleneck that delays campaigns by weeks.
The system that produces these results gets sharper with every review cycle — because every decision the committee makes becomes training signal for the next generation.