The capability gap that matters most in pharma AI is not a shortage of engineers. It is a shortage of leaders who can challenge an AI output, govern an AI-supported decision and design a workflow around AI without handing their judgement to the model.
When an executive committee asks where its AI capability gap sits, the answer it often expects is technical: more data scientists, more machine learning engineers, a larger platform team. That answer is partly right and largely distracting. Technical capacity matters, but in regulated industry the constraint on adoption is usually found at the leadership layer. Programmes stall because nobody senior is confident deciding when an output can be relied on, who owns the decision it informs, and how the surrounding work should change.
The leaders who make the most progress with AI are rarely those with the deepest technical knowledge. They are the ones with the sharpest judgement about when to trust a model, when to challenge it and when to design around it. That judgement is a capability, and like any capability it can be built deliberately. Three areas matter more than any amount of technical depth: model judgement, governance fluency and workflow design.
Model judgement: knowing when to trust the output
The first capability is knowing when an AI output can be relied on and when it cannot. The weak version of this is to accept what the system produces on the assumption that the model has seen more data and therefore has the better answer. The strong version is more demanding. Question every confidence score, test the edge cases yourself and ask where every answer came from.
Confidence scores are easy to misread. A model that reports high confidence is describing its certainty relative to the data it was trained on, not the likelihood that it is right about the case in front of you. If the inputs it now receives differ from those it learned from, the score can look reassuring while meaning very little. A leader who asks whether the current inputs resemble the training data, and what happens at the margins, is far less likely to approve a costly mistake.
Consider an illustrative example. A medical information team pilots a tool that drafts responses to enquiries from healthcare professionals. The drafts read well and the team is impressed. A leader with model judgement asks three questions before wider use: which enquiry types was the tool tested on, how does it behave with rare or ambiguous questions, and can every statement in a draft be traced to an approved source. None of these questions needs technical expertise. All of them change the quality of the decision.
Questions that build model judgement
- What was this system tested on, and how closely does that match the work we will give it?
- What does it do with incomplete, unusual or contradictory inputs?
- Can each output be traced to its source, and who checks that it has been?
- What does a wrong answer cost here, and who would notice it?
Governance fluency: owning the decision rights
The second capability is owning the decision rights for AI use within your own function. The weak version points elsewhere: there is an AI policy, compliance approved the use case, the risk team owns AI risk. The strong version is personal. Know the decision points yourself, own the escalation path and be able to read the audit trail when an AI output has influenced a regulated decision.
Governance is not paperwork. It is a clear specification of who decides what, when and on whose authority. A leader who has handed AI governance entirely to a separate team has not removed the risk; they have made it harder to see. When an AI-influenced decision is examined by an auditor, an inspector or a regulator, the question is not which committee approved the tool. It is who took the decision, and whether they had the authority and the information to take it. The answer should be a named person.
In pharma this sits alongside obligations leaders already understand: medical, legal and regulatory review of promotional material, data protection under GDPR, validation expectations for systems that affect quality, and the risk-based requirements of the EU AI Act. Governance fluency means connecting AI use to those existing frameworks rather than treating it as a separate topic owned by someone else.
Workflow design: deciding where judgement stays human
The third capability is designing the work around AI. The weak version is a set of general expectations: AI will improve productivity, tools will support people, processes will get faster. The strong version is a design discipline. Map the workflow, identify every decision point, name whether a person or the system owns each one, design the handoffs between them and measure the result.
The most common mistake is to celebrate speed without measuring quality. A process that produces a decision in an hour rather than a day is only better if the decision is at least as good. Speed without quality is a regression that looks like progress. Leaders who design workflows well decide in advance how decision quality will be checked, by whom and how often, and they treat review effort as a cost to be tracked rather than an afterthought.
Good workflow design also protects people. When it is clear which judgements remain human, staff know what they are accountable for, reviewers know what they are checking and the organisation can explain its process to anyone who asks.
Why a training course is not enough
Capability of this kind is not built in a one-off workshop. Leaders learn to exercise judgement by practising it on realistic decisions, with feedback, over a sustained period. A briefing can introduce the vocabulary, but it rarely changes how someone behaves when a confident-looking output lands on their desk under time pressure.
The organisations that build this capability well treat it as part of leadership development rather than as separate technical training. They use their own use cases, their own governance requirements and their own workflows as the material. They also make the capability visible, so that leaders can see their own progress and the organisation can see where support is still needed.
A practical sequence
A sequence that works in regulated organisations has four steps. First, assess honestly where the leadership team stands on the three capabilities, using real decisions rather than self-rating alone. Second, define what each role actually needs, because a medical director, a commercial lead and an operations head will exercise these capabilities differently. Third, pressure-test that capability with scenarios that force judgement under uncertainty, drawn from the organisation’s own pipeline of AI use cases. Fourth, embed the practice in leadership development, with regular review rather than a single event.
An assessment of AI readiness is often a useful starting point, because it surfaces both the use cases and the decisions leaders will need to make about them. Our AI readiness assessment method is designed to do exactly that, and it gives a leadership team concrete material to practise on.
Why this matters now
Capability in AI is not the same as technical depth, and it is not a certificate from a course. It is the judgement to recognise, in the moment, when a model is helping and when it is hiding a problem. Leaders who build that judgement improve the decisions taken in their own function, and they become the people an organisation can trust with larger AI programmes.
Where to start
Choose one AI use case that is already under discussion in your function and use it to test the three capabilities: what would you need to know to trust its outputs, who would own the decisions it informs, and how would the work around it change? Glyz supports pharma and other regulated organisations with this through our capability and change programmes and our AI strategy and governance work, designing the practice, embedding it in leadership development and making progress visible.
