The Limits of LLMs and Infrastructure Challenges

 

This article is part 2 of a 4-part series on the barriers and solutions to building reliable AI in healthcare.

After identifying in the first part the obstacles related to data fragmentation and the lack of traceability, let us continue the analysis with two new major challenges. Beyond the question of data, the integration of AI in the pharmaceutical industry also faces the intrinsic limitations of language models and infrastructure constraints. Two essential points to understand why generic solutions struggle to meet the needs of such a demanding sector.

 

Barrier #3: The Limits of LLMs for Critical Tasks

 

The problem: insufficient analytical capacity for high stakes use cases

Large Language Models are a type of AI trained to generate fluent text from massive datasets, by predicting the most likely sequences. But in the pharmaceutical sector, strategic tasks rarely rely solely on translating data into text. They require business understanding, explicit rules, precise calculations, and fine contextualization, elements that LLMs do not necessarily master.

Here are some concrete examples of critical tasks healthcare teams may face:

• Assessing the cost-effectiveness of a treatment for a reimbursement dossier,
• Comparing multiple therapeutic protocols based on available clinical data and medical context,
• Identifying a target population using epidemiological and regulatory criteria,
• Calculating a budget impact under different market access scenarios.

These tasks require:

• Up-to-date, verified, and structured data,
• Precise analytical models (with business rules),
• And the ability to integrate product, indication, and regulatory context.

 

What this reveal: AI still poorly aligned with business imperatives

These first three barriers show that while AI can generate answers that appear convincing on the surface, they are often unusable in regulated environments. Far from replacing structured reasoning, it simulates expertise without being able to justify it. In a sector where precision, traceability, and compliance are essential, this illusion of competence severely limits the usability of generalist models.

 

Barrier #4: Infrastructure Challenges Between Security and Cost Control

 

The problem: integrating AI in a constrained environment

Implementing AI in the pharmaceutical industry requires much more than simply connecting an API to a language model. It means addressing strict technical, regulatory, organizational, and economic requirements.

These constraints affect every link in the chain, from data access to the use of results, covering hosting, security, performance, and compliance.

 

Non-negotiable security and compliance requirements

Data handled in pharmaceutical projects must comply with strict rules:

• Secure, localized hosting (often in HDS or ISO 27001-certified environments),
• Protection against data leaks or unauthorized access,
• Compliance with GDPR, particularly in terms of retention, traceability, and purpose limitation.

 

In pricing analysis or market evaluation projects, it is often difficult to justify transferring sensitive data to third-party servers outside the organization’s control.

 

A complex IT landscape to integrate

Pharmaceutical companies operate with heterogeneous IT systems, built around:

• Internal solutions (CRM, BI, monitoring databases),
• External platforms,
• Analytical tools.

In this context, an AI solution must be interoperable with existing tools, modular so as not to disrupt workflows, and intuitive to ensure adoption without heavy training.

An essential additional factor is economic viability. The total cost of an AI solution goes far beyond the license for the model itself, it includes infrastructure, supervision, compliance, content updates, maintenance, documentation, and the ability to adapt to evolving regulatory frameworks.

 

Overcoming the Barriers: What’s Needed for Truly Useful AI in Pharma

 

The four barriers presented so far highlight one conclusion: raw AI still shows significant limitations. Neither the promises of generalist models nor the impressive demonstrations of text generation are enough to meet the specific expectations of healthcare teams today.

For artificial intelligence to become a genuine operational lever in strategic and regulated environments, several conditions must be met:

• Availability of structured, contextualized, and reliable data,
• Integration of semantic and business models to guide analysis,
• Definition of specialized AI agents designed to support precise tasks,
• Reliance on robust, interoperable, and compliant technical infrastructure.

Rather than trying to adapt consumer-grade tools to the pharmaceutical sector, the goal should be to build AI conceived from the outset for the constraints, use cases, and objectives of this industry.