Future-proofing AI use in Medical Affairs

Opinion by Dr. Leia Hazlewood, Senior Consultant in Medical Affairs, Arcondis

AI is already changing ways of working within Medical Affairs, but many organisations are stuck in proofs of concept that are either difficult to scale or have limited use cases.

What will separate the leaders in AI adoption will be the ability to match the right tool with high value use cases, and the implementation of fit-for-purpose governance that is aligned with fast-evolving global regulations. Impact is delivered when technology is tailored to multi-step workflows with appropriate workforce investment that ensures high quality human-in-the-loop verification.

WHY THE RISK-REWARD BALANCE HAS SHIFTED

AI implementation within Medical Affairs workflows can proceed with increasing certainty due to the scale of uptake in the industry and a greater awareness of potential risks and rewards. Regulatory scaffolding and practical guidance are catching up with the technology, which reduces ambiguity about how to implement systems that are both effective and auditable.

The European Union’s (EU) Artificial Intelligence Act¹ was implemented in August 2024, with staggered obligations applying between 2025 and 2030. Crucially, the Act applies to all AI systems that impact individuals in the EU, regardless of where the developer or deployer is located. It provides a risk-based framework to ensure that AI systems placed on the EU market are safe, trustworthy, transparent and preserve fundamental rights.

These principles have been reinforced recently by a joint statement from the FDA and the EMA, defining guiding principles for the use of AI in the medicinal product lifecycle.² This sets expectations for human-centered oversight, appropriate controls, and traceable technical documentation across discovery, development, product information, manufacturing and post-marketing activities.

The FDA’s 2025 draft guidance for the use of AI³ proposes a credibility assessment for AI models that are used to support regulatory decision making. This requires definition of context of use, model risk, data documentation, performance evaluation, lifecycle monitoring and fallbacks.

Regulators are also clarifying how large language models (LLMs) should be handled regarding medicines regulation. The EMA and Heads of Medicines Agencies have issued guiding principles that recognise the utility of LLMs for summarisation and knowledge mining, while highlighting the issues of results variability, hallucinations and data security risks.⁴ They encourage governance, training, and specified use cases for staff, mirroring the requirements for LLM utilisation within Medical Affairs. These regulatory documents provide Medical Affairs with a solid foundation for inspection-ready implementation, allowing leaders to develop plans based on established legal standards rather than assumptions.

MEDICAL AFFAIRS AND SAMD SUPPORT TOOLS

Understanding where AI is used within Medical Affairs workflows and where Software as a Medical Device (SaMD) begins is very important.

Once an AI feature moves beyond internal, non-patient specific support, into functions that can influence individual clinical decisions or patient management, it can be considered a medical device. SaMD has different obligations for validation, post-market monitoring, labelling and change control.

For Medical Affairs, this means designing AI capabilities with a clear context of use and early regulatory triage. Any features that could be used for the diagnosis or treatment of patients should be developed as SaMD with appropriate conformity assessment and oversight.

Within Medical Affairs, a Medinfo chatbot may fall within the SaMD category. In this context, managing scope-creep, where added-on functionalities may introduce clinical decision-making implications, needs to be carefully planned. The Therapeutic Goods Administration (TGA) in Australia, for example, has specifically highlighted this issue, and considers tools that are used for clinical decision-making are SaMD, regardless of intended use.⁵ It is likely that other regulators will follow suit, so designing tools to cover potential future regulatory changes is important.

USE CASES FOR AI WITHIN MEDICAL AFFAIRS

Insight generation

Embedding models for semantic search, clustering and sentiment analysis can shorten the time to identifying a reliable trend from a scattered set of observations. This can change the nature of insight reporting. Traditionally, field medical staff have reported an insight if they concluded that this information should or could trigger an action. With the use of AI, all observations can be collected, and the AI is able to sort these. Once a critical mass of certain observations is reached, the AI can trigger the action, which could involve labelling data as a relevant insight or could go further by generating a suggested response to this insight.

These models can be used to clearly articulate key insight questions, especially when applied to Medinfo logs, field notes, advisory board outputs and congress reports. This approach goes beyond field medical simply reporting an interaction in the CRM, by enabling easier access to insights collected from other functions that interact with HCPs, to gain a better understanding of HCP interactions across the company, helping to generate more holistic and meaningful insights.

A template for this approach is the Medical Information Data Uses for AI Semantic Analysis (MUFASA) tool.⁶ This tool uses semantic clustering and visualisation to reveal patterns in Medinfo data that are useful for Medical Affairs strategy and training. This approach can be implemented as decision support with visible human validation, aligning with requirements for proportional risk controls and traceability.

A MILE-AVAYL-MSL Society position paper⁷ lays out a Medinfo workflow whereby the model drafts the content and a human specialist refines this to approve the final version. In this model, citations and excerpted source passages are presented to the reviewer for verification within a governance framework that ensures records are inspection ready. Scientific response document development can be high-volume, tedious work, strengthening the case for a hybrid approach that combines transformer summarisation with human review.⁸ It is essential that safe input practices and critical cross-checking accompany such assistants, to keep productivity gains aligned with data protection and quality.

Evidence generation

Machine learning and Natural Language Processing (NLP) can be used for several tasks involved in evidence generation. It can be used to accelerate literature screening, automate retraction surveillance, support cohort identification in Electronic Health Records (EHR) and claims data, and strengthen pharmacovigilance signal detection. These activities improve the speed and reproducibility of postmarketing evidence generation. Both the EMA and the FDA expect the documentation of model credibility and visible human oversight whenever outputs inform regulatory or safety decisions, so technical plans need to be embedded within regulatory plans from the outset.^3,9

Field Medical enablement

AI assisted pre- and post-call briefs can strengthen scientific dialogue in Health Care Professional (HCP) engagement. It can also improve the quality of the insights that are generated in each interaction by tagging and grouping insight themes.

Pre-call, this may take the form of summarising a clinician’s publications and likely interests to help guide discussion points that are aligned with the HCP’s interests and needs.

Post-call, note tagging can make insight generation easier and faster (whether done by a human or an AI-assisted tool). Omnichannel engagement strategies can be enhanced with the use of AI to target personalised communication with HCPs. For example, integrating CRM systems with predictive analytics can optimise the timing of interactions as well as the communication channel. This can help to build relationships and improve trust with HCPs.

Using AI to combine information from publication networks, trial participation, guideline contributions and social or digital footprints could also help reveal influence structures and emerging experts who may not be visible through volume metrics alone. This can be used to improve stakeholder planning and content targeting. It can also be used in automated monitoring of abstracts, posters and sessions across priority journals and congresses to produce regular intelligence digests.

These can be grouped into a specific therapeutic area to reduce the time spent on searching and collating this data. It is key, however, that field medical staff can trust the completeness and accuracy of the information produced by any AI tools. If trust has not been adequately built-up, there is the risk that they may choose to perform these searches manually to feel confident in their work output, thus rendering the AI tool redundant. Engaging with field medical staff to highlight how AI can strengthen their role is important.

Internal medical education

Generative video and conversational agents can aid internal training of staff. These can provide tailored, interactive learning modules on both internal and external topics. Depending on the learner’s role and needs, information about products, therapeutic areas, objection handling and compliance can be delivered in a manner that best suits the learner (and their preferred language). For example, a new Medical Science Liaison can simulate conversations with different HCPs prior to going into the field, allowing them to make mistakes in a safe environment, and to learn from these prior to real-world interactions.

AVOIDING PILOT PURGATORY

Our key implementation lessons:

• Start small with clearly scoped, low risk use cases
• Define context of use, success criteria and required controls upfront
• Plan validated datasets, evaluation metrics, drift checks and fallbacks early
• Run brief pilots to measure impact before scaling
• Ensure strong data foundations and system integration to avoid failure
• Invest in skilled workforces to maintain transparent human oversight

In our experience, many pilots have underdelivered because the data foundations were not prepared for machine consumption and because the tools were added-on, rather than embedded in daily workflows.

Fragmented data systems and poor interoperability are recurring obstacles that can slow adoption and blunt impact. Hallucinations and inconsistent results have also eroded trust when teams have used general purpose LLMs. Copyright and licensing considerations have held back programmes due to concerns about inadvertently including content without the right permissions.

The safest path here has been to limit data sets to validated, licensed sources. We have found that siloed data repositories with inconsistent taxonomies and weak integration with existing Medinfo and CRM platforms are a risk to success. If the underlying data being used by the AI is not properly considered, initially strong models can prove underwhelming for the end-user. We have been able to combine our experience in data management with our knowledge of medical devices and Medical Affairs workflows to help businesses along the path to success with AI adoption.

KEY STEPS TO SUCCESSFUL PROGRAMMES

In our work with Medical Affairs teams, we have observed that successful programmes usually follow these steps:

INVESTING IN PEOPLE

Transparency in AI systems is highly valued by human users, but can be easy to underdeliver, both in terms of systems design and in communication to end-users. Human oversight needs to remain visible, particularly wherever content or analyses can influence clinical discussions or external decisions, with named reviewers and sign-offs captured in the system of record.

Investment in the human workforce is also critical. While AI will increasingly be able to perform basic, entry-level roles, skilled human oversight will remain essential.

Currently, experienced professionals have had years of on-the-job training, providing them with the skills necessary to oversee AI systems, because they understand basic tasks well.

However, if too many entry level roles are lost to increasingly efficient AI systems, there is the real risk of a resultant training shortage. In the future, there may be fewer experienced workers who have been able to develop critical oversight skills without needing to rely on AI. Employers need to plan for this now by investing appropriately in the workforce of tomorrow with the skills to both use and adequately critique AI outputs.

FROM EXPERIMENTATION TO IMPLEMENTATION AT SCALE

Now is the time for Medical Affairs to take a decisive leadership role in shaping how AI within their organisations is developed and deployed. Medical Affairs professionals need to be involved in defining the right use cases to guide implementation that aligns with their workflow needs as well as regulatory expectations. If not, others will move ahead, and the tools developed may fall short of the scientific rigor and real-world requirements of Medical Affairs teams.

The technical capabilities, use cases and regulatory clarity now exist, so organisations that act decisively will shape the future of how this technology is implemented. This is a clear inflection point to set the direction for how AI will transform Medical Affairs in the future, to deliver measurable impact without eroding trust.

HOW ARCONDIS HELPS?

We help our clients future-proof AI in Medical Affairs by combining deep domain expertise with end-to-end capabilities, from strategy through to implementation. We enable organisations to build strong data foundations, generate high-quality evidence, and deploy AI-driven insights that enhance decision-making and scientific engagement. By integrating medical strategy, digital innovation and regulatory alignment, we ensure that AI initiatives are not only impactful, but also scalable, compliant, and built for the long term. Contact me for more information.