Only 14% of organisations surveyed said they have a clear AI strategy with defined goals and outcomes. The majority are either still developing a strategy (39%) or operating in disconnected pockets with no enterprise-wide direction (32%). Meanwhile, AI keeps moving, embedding itself into workflows, making recommendations and, in some cases, making decisions.

This pattern is visible across industries, but in pharma it carries a specific and acute risk. Too many organisations are treating strategy as a technology choice, selecting a platform provider and calling it a plan. But choosing an AI vendor is not an AI strategy. A genuine strategy requires organisations to go back to first principles: who are we, what do we exist to do and where does AI therefore fit? The organisations that demonstrate the clearest understanding of their own identity and purpose are the ones best positioned to harness AI effectively. Without that foundation, the approach becomes piecemeal.

The consequence is that accountability becomes variable. Ownership of AI-driven outcomes shifts between functions, vendors and platforms, and no single line of sight connects a model’s recommendation to the person answerable for its consequences. For pharma, where regulatory frameworks already require documented decision rights, validated processes and clear lines of accountability, this is not a theoretical concern. It is a compliance exposure that grows with every uncoordinated deployment.

In highly regulated environments, AI governance takes on an even greater level of importance. When AI contributes to a regulatory submission, a pharmacovigilance signal, a clinical trial design or a pricing decision, the ability to explain not just what the system recommended but how it arrived there is not a technical nicety. It is a requirement that existing frameworks will increasingly demand and that internal audit, regulators and, in the event of harm, courts will expect to see answered.

The problem with ‘a human in the loop’ as it is currently practised is that it implies review without specifying the nature, depth or authority of that review. A human may be present in the process, but are they genuinely steering, or simply watching the outputs pass through? The distinction matters enormously. As a way of thinking about it, the concept of ‘a human at the helm’ shifts the frame: you are still doing what you have always done, but now you are doing it with AI at your disposal. A scientist is still going to keep discovering new drugs. A laptop made that faster. AI will make it faster again. But the scientist remains at the helm.

Outright AI hallucinations may have reduced significantly, but what has replaced them may be harder to catch: misrepresentation, miscategorisation and subtle contextual errors that an untrained eye will not spot. These are not failures of the technology so much as failures of the governance model that surrounds it.

When the output looks credible and the reviewer lacks the scaffolding to interrogate it, the loop exists on paper but is hollow in practice.

The research identifies a reinforcing pattern that pharma organisations should recognise immediately: fear makes experimentation risky; limited experimentation prevents judgement from developing; without judgement, confidence does not form; and when confidence stays low, employees defer to AI rather than challenge it.

In pharma, the functions that most need to challenge AI outputs – medical affairs, regulatory, pharmacovigilance, clinical operations – are staffed by people with deep professional expertise. But expertise and confidence in challenging AI are not the same thing. AI’s outputs have become impressively fluent and that fluency is itself a risk. When a model produces a result that looks polished and authoritative, the instinct to defer is natural, even among subject matter experts.

AI should be treated as a capable companion, not a replacement for professional judgment. It can accelerate discovery, surface patterns and handle volume at a pace no human team can match. But it does not yet possess the contextual understanding that resides in people’s heads, the tacit knowledge built through years of working within a specific therapeutic area, regulatory environment or organisational culture.
 That knowledge does not live in documents or meeting notes alone; it exists somewhere in between. Until AI systems can access that full context – and they cannot yet – human oversight must be substantive, not ceremonial.

The research proposes a governance framework built from five foundations, each enabling the one above it. For pharma, each layer has a specific translation:

1. Direction means going back to first principles. A pharma company’s AI strategy should begin not with the technology but with the organisation’s identity: what diseases it is trying to treat; what its pipeline demands and where AI can genuinely accelerate that mission.

2. Authority means defining who has the right and the obligation to override an AI recommendation at every decision point in the value chain, from molecule selection through to commercial launch.

3. Visibility means building the infrastructure for full explainability, not just technical transparency but the ability to trace how an AI-informed decision was made, by whom and on what basis.

4. Capability means investing in the skills that allow professionals to work with AI critically rather than deferentially, developing AI-native workflows where governance is embedded, not bolted on.

5. Accountability means closing the loop: ensuring that every AI-assisted decision has a named owner; a documented rationale and a clear escalation path when something does not look right.

The opportunity for pharma is not to slow AI adoption but to rewire how it is deployed. Treating AI as a bolt-on, a chat interface layered over existing processes, a dashboard enhancement or an incremental efficiency tool will deliver incremental gains. But the organisations that reimagine their workflows to be AI-native, rethinking processes from the ground up rather than optimising each existing step, are the ones that will build a genuine competitive edge. If a drug discovery process currently involves 40 steps, AI should be used to fundamentally reimagine the process as 18 or 20 steps with AI woven in from the start.

The industry’s fundamental challenges are not going to change. Pharma companies will still be measured on the drugs they discover, the treatments they bring to market and the patients they reach. AI does not alter those imperatives. It raises the stakes for getting the governance right before an incident forces the issue.

Pharma has always understood that some decisions are too consequential to leave ungoverned. Patient safety, regulatory integrity and clinical credibility are not values the industry adopted under pressure. They are its foundations. The question now is whether organisations will extend that same rigour to the intelligent systems increasingly shaping those decisions or wait until a failure makes the case for them.