Human expertise combined with AI: The future of drug safety

Introduction 

Pharmacovigilance is entering a new era. The volume, variety, and speed of safety information generated throughout the medicinal product lifecycle continue to increase, creating both opportunities and challenges for pharmacovigilance professionals. Artificial Intelligence (AI), machine learning, natural language processing, and automation can help pharmacovigilance teams process large volumes of information more efficiently and identify patterns that may otherwise be difficult to detect. 

However, the future of drug safety is unlikely to be about AI replacing pharmacovigilance professionals. Instead, it will increasingly depend on combining technological capabilities with human clinical judgment, scientific expertise, regulatory knowledge, and ethical decision-making. 

AI can help identify, organize, prioritize, and analyze information. Human experts remain essential for interpreting evidence, assessing clinical relevance, challenging algorithmic outputs, making informed decisions, and ensuring that patient safety remains at the center of the process. 

The most effective future PV model is therefore not Human vs. AI, but Human expertise + AI. 

1. Why Pharmacovigilance Needs a New Approach

Modern pharmacovigilance operates across a complex network of data sources. 

Safety information may come from: 

• Individual Case Safety Reports (ICSRs) 

• Clinical trials 

• Scientific literature 

• Electronic health records 

• Patient support programs 

• Registries 

• Social and digital sources where applicable 

• Epidemiological studies 

• Regulatory authorities 

• Medical information channels 

• Product quality and complaint systems 

The increasing volume of information creates a practical challenge: pharmacovigilance teams need to identify meaningful safety information without compromising accuracy, quality, or regulatory compliance. 

Traditional manual processes remain essential, but technology can support professionals by reducing repetitive activities and enabling faster access to relevant information. 

The objective should not simply be processing more data. It should be turning more data into reliable safety intelligence. 

2. Where AI Can Support Pharmacovigilance

AI can potentially support multiple stages of the pharmacovigilance lifecycle. 

Case Intake and Triage 

AI-enabled systems can help identify potential safety information from incoming communications and classify information according to predefined criteria. 

For example, automated tools may support: 

• Identification of potential adverse events. 

• Case triage. 

• Duplicate detection. 

• Case prioritization. 

• Data extraction. 

• Routing to the appropriate workflow. 

This can reduce manual administrative workload and allow PV professionals to focus more time on activities requiring scientific judgment. 

Data Processing 

Natural language processing can assist with extracting relevant information from unstructured text. 

Potential applications include: 

• Identifying adverse event terms. 

• Extracting patient characteristics. 

• Recognizing medicinal products. 

• Identifying dates and clinical information. 

• Supporting coding. 

• Highlighting missing information. 

However, automated extraction should remain subject to appropriate validation and human oversight.

3. AI-Assisted Literature Surveillance

Literature monitoring is another area where AI may provide significant support. 

Traditional literature screening can involve reviewing large numbers of publications to determine whether they contain potentially relevant safety information. 

AI can assist by: 

• Screening large volumes of publications. 

• Identifying potentially relevant articles. 

• Classifying publications. 

• Extracting safety-related information. 

• Prioritizing articles for human review. 

The role of the pharmacovigilance professional remains critical. 

An algorithm may identify an article as potentially relevant, but a qualified reviewer must determine whether the publication contains reportable or clinically meaningful safety information. 

This illustrates an important principle: 

AI can accelerate the search; human expertise determines the significance. 

4. AI and Signal Detection

Signal detection is one of the most important areas where advanced analytics can support drug safety. 

Traditional signal detection may involve statistical screening, medical review, and analysis of trends within safety databases. 

AI and machine learning approaches may help identify: 

• Unexpected patterns. 

• Associations between products and events. 

• Emerging trends. 

• Changes in reporting behaviour. 

• Complex relationships across multiple variables. 

However, detecting an association is not equivalent to establishing causality. 

A statistical or algorithmic signal may be influenced by: 

• Reporting bias. 

• Confounding. 

• Indication. 

• Concomitant medications. 

• Changes in exposure. 

• Data quality. 

• Stimulated reporting. 

Human pharmacovigilance experts must therefore investigate the underlying evidence and determine whether the observation warrants further assessment.

5. The Human Expert Remains at the Centre

AI systems can process information at a scale that would be difficult for individuals to achieve manually. But pharmacovigilance decisions require more than data processing. 

Human experts contribute: 

Clinical Judgment 

A physician, pharmacist, or appropriately qualified PV professional can interpret clinical context and determine whether an event is medically meaningful. 

Scientific Reasoning 

Experts evaluate biological plausibility, alternative explanations, consistency of evidence, and the overall scientific context. 

Regulatory Knowledge 

Local regulatory requirements can differ between jurisdictions. Professionals must understand applicable reporting obligations, timelines, documentation requirements, and regulatory expectations. 

Contextual Understanding 

A safety observation cannot always be interpreted independently from the product’s indication, patient population, exposure, treatment alternatives, and existing safety profile. 

Ethical Judgment 

Ultimately, pharmacovigilance decisions affect patients. Human oversight provides an important safeguard when decisions involve uncertainty, competing considerations, or potential public health consequences. 

6. AI Should Support — Not Replace — Scientific Judgment

A critical distinction should be maintained between automation of a task and automation of a safety decision. 

For example: 

AI can: 
Identify potentially relevant cases. 

Human expert can: 
Determine whether the case meets the applicable criteria and what action is appropriate. 

AI can: 
Identify statistical patterns. 

Human expert can: 
Evaluate whether the pattern represents a credible safety concern. 

AI can: 
Prioritize documents for review. 

Human expert can: 
Determine their clinical and regulatory relevance. 

This division of responsibilities creates a controlled approach in which technology increases efficiency while qualified professionals retain appropriate oversight. 

7. Data Quality: The Foundation of AI-Enabled Pharmacovigilance

AI is only as reliable as the data used to train, validate, and operate the system. 

Poor-quality or incomplete data can result in: 

• False positives. 

• False negatives. 

• Incorrect classifications. 

• Missed safety information. 

• Biased outputs. 

• Inconsistent recommendations. 

Pharmacovigilance organizations therefore need strong data governance processes covering: 

• Data quality. 

• Data completeness. 

• Data traceability. 

• Data provenance. 

• Data integrity. 

• Access controls. 

• Data security. 

• Appropriate retention. 

Before implementing an AI solution, organizations should understand what data the system uses, how the data are processed, and how outputs are generated and documented. 

8. Validation and Quality Management of AI Systems

Introducing AI into pharmacovigilance does not remove the need for a strong Quality Management System (QMS). 

Instead, it creates additional quality considerations. 

Organizations should establish appropriate controls around: 

• System validation. 

• Intended use. 

• Performance monitoring. 

• Change control. 

• Version management. 

• Access management. 

• Audit trails. 

• Error management. 

• Periodic review. 

• Vendor oversight. 

A system that performs well during initial testing may behave differently following changes to its underlying model, data, configuration, or workflow. 

Therefore, AI-enabled PV systems should be treated as controlled components of the overall pharmacovigilance system.

9. Explainability and Human Oversight

One of the major challenges associated with AI is understanding why a system produced a particular output. 

In pharmacovigilance, explainability is particularly important because safety decisions may need to be justified to: 

• Internal stakeholders. 

• Quality functions. 

• Auditors. 

• Regulatory authorities. 

• Inspectors. 

If an AI system flags a potential signal, organizations should be able to understand how the output was generated and how it was subsequently reviewed. 

Human oversight should therefore be designed into the workflow. 

A practical model may include: 

AI output → Human review → Scientific assessment → Documented decision → Quality oversight 

This provides a traceable link between automated analysis and the final pharmacovigilance decision. 

10. Regulatory Compliance in the Age of AI

AI implementation must operate within the existing pharmacovigilance regulatory framework. 

Organizations remain responsible for fulfilling applicable requirements concerning: 

• Adverse event reporting. 

• Signal detection and management. 

• Benefit–risk evaluation. 

• Risk management. 

• Documentation. 

• Data integrity. 

• Quality management. 

• Regulatory communication. 

• Inspection readiness. 

Using an AI tool does not transfer these responsibilities to the technology provider. 

The Marketing Authorization Holder (MAH) remains responsible for ensuring that its pharmacovigilance system is effective and compliant. 

This makes vendor qualification and oversight particularly important when AI-enabled services are outsourced. 

11. AI Governance and Clear Responsibilities

Before deploying AI in PV, organizations should define who is responsible for the system and its outputs. 

Responsibilities may involve: 

Pharmacovigilance 

Defines the intended PV use, reviews outputs, and ensures that the technology supports appropriate safety activities. 

Quality Assurance 

Provides oversight of validation, procedural controls, deviations, CAPA, and compliance. 

IT and Data Teams 

Support system architecture, cybersecurity, data management, access controls, and technical performance. 

Regulatory Affairs 

Assesses regulatory implications and supports interactions with health authorities where appropriate. 

Vendor Management 

Ensures that third-party AI providers meet contractual, quality, security, and performance requirements. 

Senior PV Leadership / QPPV 

Provides appropriate oversight of the overall PV system and ensures that the introduction of AI does not compromise patient safety or regulatory compliance. 

Clear accountability is essential: an automated system may generate an output, but an accountable organization must own the decision. 

12. Managing Bias and Algorithmic Limitations

AI systems may reproduce biases present in their training data or underlying datasets. 

In pharmacovigilance, this can create important challenges. 

For example, differences in reporting patterns between countries, populations, healthcare systems, or product types may influence algorithmic outputs. 

Organizations should therefore monitor AI performance for: 

• Systematic errors. 

• Population bias. 

• Geographic differences. 

• Language-related limitations. 

• Under-reporting patterns. 

• Changes in performance over time. 

Human experts should challenge unexpected results rather than automatically accepting algorithmic outputs.

13. The Future PV Professional: From Processor to Safety Strategist

As automation takes over more repetitive activities, the role of the PV professional may continue to evolve. 

Future PV professionals will increasingly need a combination of: 

• Clinical and scientific knowledge. 

• Regulatory expertise. 

• Data literacy. 

• Critical thinking. 

• AI awareness. 

• Risk assessment skills. 

• Communication skills. 

• Quality management knowledge. 

The ability to ask the right questions will become increasingly important. 

Instead of simply asking: 

“What did the system find?” 

PV professionals may need to ask: 

“Why did it find this?” 

“What evidence supports it?” 

“What might the system have missed?” 

“Does this change our understanding of the product’s benefit–risk profile?” 

This shift could make pharmacovigilance increasingly strategic and evidence driven. 

14. Building a Human + AI PV Operating Model

Successful AI implementation should begin with the PV process, not the technology. 

Organizations should first identify: 

1. Which activities are repetitive and resource-intensive? 
2. Where can automation reduce operational burden? 
3. Which activities require expert judgment? 
4. What risks could automation introduce? 
5. What level of human review is required? 
6. How will performance be measured? 
7. How will the system be validated and monitored? 

The goal should be to create a balanced operating model. 

Automate where appropriate 

Use technology for high-volume, repetitive, and structured activities. 

Augment where valuable 

Use AI to provide insights, prioritization, and analytical support to professionals. 

Retain human control where essential 

Maintain expert oversight for clinical interpretation, signal assessment, regulatory decisions, and significant safety conclusions. 

15. Measuring the Value of AI in Pharmacovigilance

The success of an AI implementation should not be measured simply by how much manual work has been eliminated. 

Organizations should consider a broader set of measures, including: 

• Accuracy. 

• Sensitivity. 

• Specificity. 

• Processing time. 

• Quality performance. 

• Compliance. 

• False-positive and false-negative rates. 

• Human review requirements. 

• System reliability. 

• User adoption. 

• Inspection readiness. 

Most importantly, organizations should assess whether the technology improves the quality and timeliness of safety surveillance. 

Efficiency is valuable, but patient safety remains the primary objective. 

16. What the Future May Look Like

The future pharmacovigilance environment is likely to become increasingly connected and data driven. 

AI may support a continuous safety ecosystem in which information from multiple sources is processed and analyzed in near real time. 

A future workflow could look like: 

Data collection → AI-assisted processing → Pattern identification → Human scientific review → Signal assessment → Regulatory evaluation → Risk management → Continuous monitoring 

Rather than replacing the PV professional, technology could allow professionals to spend less time searching through information and more time interpreting evidence, evaluating uncertainty, and making scientifically informed decisions. 

The organizations that benefit most from AI will likely be those that combine technological innovation with strong governance, quality systems, regulatory expertise, and human oversight. 

At Baupharma, we see this human–technology collaboration as an important part of the future of pharmacovigilance. Through our pharmacovigilance services, we combine scientific expertise and regulatory knowledge with technology-enabled processes to help organizations strengthen their safety operations and adapt to evolving PV requirements. 

Because the future of drug safety is not about choosing between human expertise and technology, it is about making them work better together. 

Key Takeaways 

• AI has significant potential to transform pharmacovigilance by supporting data processing, literature surveillance, case management, signal detection, and analytics. 

• AI should be viewed as an augmentation tool, not a substitute for qualified pharmacovigilance expertise. 

• Human professionals remain essential for clinical judgment, scientific interpretation, regulatory assessment, and ethical decision-making. 

• High-quality data and robust data governance are fundamental to reliable AI-enabled PV. 

• AI systems should be appropriately validated, monitored, controlled, and documented within the PV Quality Management System. 

• Organizations should maintain clear accountability for AI-supported pharmacovigilance decisions. 

• Vendor qualification and oversight are essential when AI-enabled PV activities are outsourced. 

• Bias, explainability, system limitations, and performance changes should be actively monitored. 

• The future PV professional will increasingly combine scientific expertise with data and AI literacy. 

• The most effective model is not Human vs. AI, but Human expertise combined with AI. 

• Ultimately, technology should serve the fundamental purpose of pharmacovigilance: identifying and managing risks effectively to protect patients.