The life sciences sector has always depended on the quality of evidence behind its decisions. Today, the volume of evidence, and the speed at which it can be analysed, is changing dramatically with artificial intelligence.
In September, the National Commission into the Regulation of AI in Healthcare, an independent advisory body set up by the Medicines and Healthcare products Regulatory Agency (MHRA), published its recommendations [1]. Its chair, Professor Alastair Denniston, described AI as an exceptional opportunity for healthcare, comparing its potential impact with breakthroughs seen with the development of antibiotics and MRI scanning [2]. Yet that enthusiasm came with an important caveat. Progress needs to be matched by safeguards that ensure these systems can be completely trusted.
This challenge is particularly acute in the life sciences sector. Pharmaceutical companies can now interrogate vast volumes of scientific literature, clinical data, company disclosures, and market research at a scale and speed that would have previously been impossible. This has the potential to accelerate work across drug discovery, trial design, medical safety and commercial strategy. Yet faster access to information does not automatically mean better judgement.
The Commission’s 44 recommendations, informed by more than 12,000 patients, clinicians and others, underline the importance of accountability when technology is used to inform decisions about health. For pharmaceutical companies, the same principle applies throughout the drug lifecycle, from identifying promising drug targets to monitoring safety risks once treatments reach clinical trials and beyond.
The true cost of intelligence fatigue
Every new AI tool introduces more signals, interpretations, and recommendations from which life sciences teams can gain new insights. While one model may identify a promising therapeutic target, another could highlight regulatory or safety risks. The volume of intelligence grows, but confidence in the final decision is often not linear.
The life sciences sector is well placed to benefit from AI developments because it generates enormous quantities of biological, chemical, clinical, and patient data. It is also one of the least forgiving environments in which to deploy AI. For example, an error can be amplified by a model, creating false confidence in a flawed output. If that output informs trial design or safety monitoring, the consequences extend far beyond an inaccurate spreadsheet and could affect patient health.
Fragmentation furthers the problem as intelligence remains dispersed across different sources and platforms. Research teams may work from one evidence base while commercial, medical, and regulatory colleagues use another. An AI-generated answer can appear authoritative without revealing whether it reflects the complete picture, or rests on outdated or partial information.
The MHRA’s concern about AI systems that continue to learn, adapt and “drift”, where changing data causes outputs to become less reliable over time, illustrates why a one-off assessment is not enough. The Commission’s recommendation that healthcare AI should be monitored continuously, with approval removed if performance deteriorates, points to a wider principle for the sector.
Trust must be maintained throughout the life of a system, not granted permanently at the point of deployment. This is where having a human in the loop is essential, particularly in healthcare. Qualified professionals must be able to scrutinise AI-generated outputs, identify errors, and remain accountable for decisions that could affect the entire health and pharmaceutical lifecycle.
Better decisions start with better workflows
New, powerful AI models are being released weekly, but these models still require the right intelligence workflow built around them. These workflows connect trusted information, automated analysis, and expert insights, allowing users to synthesise intelligence without losing context. Crucially, intelligent workflows are designed around the decision, rather than the novelty of the technology. This is the crucial difference between decision-grade AI and generic AI models.
Businesses operating in the life sciences sector can build more trusted intelligence workflows by focusing on three priorities. First, AI must operate with domain-specific context. High-stakes decisions require an understanding of scientific research, clinical and drug data, and regulatory relationships that general-purpose models may lack. Domain-specific reasoning can produce analysis experts can, and should, properly interrogate.
Second, it’s vital that validation is continuous. In drug discovery, computational predictions should be tested through physical experiments, with laboratory results used to refine the model. AI then becomes part of an evidence-generating process, rather than a detached layer producing recommendations.
Third, governance should reflect risk. A tool summarising market commentary should not be governed like one influencing clinical development. The Commission’s proposed AI “L plate”, allowing new healthcare models to be tested under close professional supervision, recognises that innovation and oversight can progress together. Life sciences companies need the same tiered approach, with human approval for critical applications.
Building trust through transparency
As AI-generated analysis becomes more common, provenance becomes more valuable. Every important insight should be traceable to authoritative source material. Users need to move easily from an answer to the supporting document, examine its original context, and distinguish established evidence from inference.
Provenance allows a scientist, clinician, or other life sciences professional to challenge a conclusion, a commercial leader to defend an investment, and a regulator to understand how a decision was reached. While technology may save time, the responsibility for correcting its mistakes must remain with humans.
Transparency matters for patients, too. The Commission recommends that people should know when AI is involved in their care and be able to access information about the product. Research indicating that some patients may withhold sensitive information when AI processes a consultation shows how deployment can change behaviour [2]. A technically effective system can still make healthcare less effective if people do not wholeheartedly trust it.
Predictive accuracy alone is therefore too narrow a measure of success. Organisations should also assess whether an AI system is robust, transparent about uncertainty, and grounded in verifiable evidence. An impressive or polished answer that is factually incorrect because it lacks verifiable contextual evidence is not only less useful, it can be dangerous.
The real measure of AI in life sciences
AI is changing how life sciences organisations turn knowledge into action, from the earliest stages of discovery to treatment. But realising that potential depends on more than adopting increasingly capable technology. It requires rigorous validation, clear provenance and workflows that keep expert human judgement firmly in the decision making process.
The sector already has an abundance of information at its disposal. The advantage will come from knowing exactly what deserves attention, what can be properly trusted and where human judgement is paramount.
Author Bio
