Accelerating MLR review with specialised AI agents

Medical, legal and regulatory (MLR) review is a critical part of producing pharmaceutical communications. Materials must be accurate, appropriately evidenced and compliant with the requirements that apply to their audience and market.

For global life sciences companies, this creates a significant operational challenge. Reviewers may need to assess content against approved product information, scientific literature, industry codes, local regulations, and company-specific brand standards. Those requirements can vary between countries and formats, while the volume and complexity of material continue to grow.

Mistakes can be costly. An unsupported claim, inaccurate reference or inappropriate piece of promotional language can result in rework, delayed approval, financial penalties and reputational damage.

We set out to deploy specialised AI agents on some of the most complex documents produced by the client to make the process faster and more scalable, all whilst keeping the controls of qualified human reviewers in the driving seat.

The challenge: reviewing complex Core Visual Aids

Our initial work focused on Core Visual Aids (CVAs): detailed promotional documents used to communicate information about medicines to healthcare professionals.

A single CVA can run to more than 100 pages and combine product claims, clinical evidence, benefits, risks, contraindications and case studies. The content may appear as text, charts, tables, diagrams, images and other visual formats.

Every relevant claim needs to be checked against the appropriate evidence and requirements. Depending on the material and market, that can include:

• Spelling, grammar and editorial consistency
• Alignment with the Summary of Product Characteristics (SmPC)
• Applicable EFPIA and IFPMA codes
• Relevant national or local requirements
• Global and local company guidelines
• Brand and style standards
• Accuracy of claims drawn from scientific literature and case studies

The objective was not simply to find typographical errors. The system needed to identify potentially unsupported, inaccurate or non-compliant content, explain why it had been flagged, and help a human reviewer resolve it.

Why a single-model approach was not enough

This problem cannot be solved reliably by giving one AI model a 100-page CVA, a collection of policies and a prompt to “find every error”.

The supporting material can be considerably longer than the CVA itself. Scientific publications, product information and internal guidance can collectively run to thousands of pages. Passing everything into a model at once increases processing costs and makes it more difficult to ensure that the right evidence is applied to each claim.

MLR review also involves several distinct types of reasoning. Verifying a numerical claim against a clinical paper is different from assessing whether promotional language is appropriately balanced. Checking alignment with the SmPC is different again from applying a company’s terminology and brand standards.

We therefore designed a coordinated network of specialised agents rather than relying on one general-purpose reviewer.

The solution: specialised agents with targeted context

Each agent was assigned a defined area of responsibility. Some focused on objective checks, such as spelling, references and numerical accuracy. Others assessed content against product information, industry codes or company-specific guidance.

A code-driven orchestration layer managed how content was divided, which sources each agent received, and when each review took place. Large documents were separated into context-aware sections and routed to multiple agents in parallel.

This approach provided each agent with the material relevant to its task without overwhelming it with unrelated information. It also allowed the system to compare content against several requirements systematically while controlling processing time and cost.

When the system detected a potential issue, it could:

• Identify the affected claim, passage or data point
• Explain the reason for the finding
• Point to the relevant evidence or requirement
• Show the corresponding information from the source material
• Suggest revised wording for consideration by a human reviewer

For example, an agent could compare a clinical claim with the cited academic paper and show where the source did not support the wording. It could also identify superlative or absolute language that might be interpreted as an unsupported factual claim and propose a more appropriate alternative.

Importantly, these outputs were designed to support and not replace qualified MLR professionals. Final assessment and approval remained with human reviewers whom could be more efficient.

Designing for accuracy and coverage

Two measures were particularly important: how many genuine issues the system detected and how many of its findings were relevant.

Missing a material compliance issue could create risk, while producing too many false positives would increase the reviewer’s workload and reduce trust in the system. The solution therefore needed to balance broad coverage with sufficiently precise findings.

Specialisation helped improve that balance. Giving each agent a narrowly defined task, the appropriate source material and clear decision criteria reduced the likelihood of unsupported conclusions. Requiring findings to be tied to specific evidence also made the output easier for reviewers to verify.

The architecture was designed to reduce hallucinations, false positives and false negatives while creating a clearer audit trail for each finding.

Producing results at operational speed

Accuracy alone was not enough. The system also needed to work at the speed and scale required by a global life sciences organisation.

The parallel design allowed different sections of a CVA and different types of review to be processed simultaneously. Additional capacity could be applied to longer documents or higher volumes without requiring every review to run sequentially.

Within a one-month engagement, we developed a pilot capable of reviewing a 100-page CVA in several minutes. In the evaluation conducted during the engagement, the system detected more than 90% of the known MLR issues in the test material.

The client’s existing externally managed process could take weeks. Although final human quality control remained essential, the pilot demonstrated the potential to move much of the initial review earlier in the content-development process and complete it significantly faster.

Teams could submit draft material, receive detailed feedback, make corrections and test new versions before sending the content for final approval. This creates the potential to reduce external review costs, avoid repeated late-stage revisions and allow MLR specialists to focus their attention on the findings that require expert judgment.

Extending the approach beyond CVAs

CVAs provided a strong starting point because they are among the most complex materials handled by the client’s review team. The same architecture can be adapted for other document types, including:

• Communications for healthcare professionals
• Global value dossiers
• Patient information materials
• Emails and digital content
• Other promotional and medical communications

Each application would need to be configured for the relevant audience, market, source material and approval process. The underlying principle and framework, however, remains the same: divide a complex review into well-defined tasks, give each agent the right evidence, and present traceable findings to a qualified human reviewer.

The wider opportunity

The immediate opportunity is to reduce the time and effort involved in reviewing complex material. Faster feedback can also help teams identify problems earlier, reduce late-stage rework and improve readiness for product launches and campaigns.

More consistent review can reduce exposure to avoidable compliance issues, financial penalties and reputational damage. It can also help organisations apply changing requirements more systematically across markets.

Faster MLR review can also improve launch readiness. By reducing delays in the development and approval of compliant materials, teams can equip local markets and communicate with healthcare professionals sooner after a medicine is authorised. In an industry where effective commercial exclusivity is time-limited, even modest time savings can create meaningful commercial value and help information about new treatment options reach healthcare professionals and eligible patients sooner.

To learn more about this case study and how we can help you with your regulatory and content review challenges, get in touch below.