“We use AI.”
You hear it more and more. In the pharmaceutical industry too. But that still tells you very little.
Does the system write a report? Recognise patterns in tissue images? Or help researchers design a molecule?
Looking at a number of practical examples, what struck me most was their variety. The same word covers very different applications. And within each application, not every step involves AI.
Six examples make this more concrete.
A first draft of the Product Quality Review
Generative AI is being used in annual Product Quality Reviews to collect data, draft text and perform initial analyses. Quality professionals use that preparatory work to support their interpretation. This application has been described in public reports of operational use.
This is familiar work for a quality professional: bringing information from different sources together and turning it into a coherent account.
Where is the AI? Mainly in generating text and supporting the analysis. Ordinary software can also retrieve data or populate fixed fields. The AI component is the processing in which a model produces, for example, a summary or draft wording.
The result is a report prepared for review. That does not mean the system independently assesses product quality.
Extracting information from an adverse event report
During adverse event intake, AI helps extract relevant information from text and classify reports. Does a report contain the necessary information? Are there indications of a serious case? Both research and operational use have been described for these types of tasks.
Here, AI is trying to identify relevant information in the submitted text and place it into categories.
Where is the AI? In interpreting and classifying information that is not always described in exactly the same way. Routing a classified report to the appropriate work queue can then follow a fixed process rule.
This example also shows that AI in pharma predates the current interest in chatbots. The historical account does not, however, tell us how the system is configured today.
Helping researchers design molecules
In drug discovery, AI systems are used to support molecular design and the planning of ways to synthesise molecules. One described application uses several AI agents to perform different research tasks. Integration of such a system into the research process has been described in a scientific publication.
Here, AI plays a role across a series of research tasks. Different agents can divide tasks and call on digital tools.
Where is the AI? In the model-driven processing and coordination of those research tasks. An agent might use a tool, but that does not make the tool itself AI. AI, calculations and ordinary software can work together within a single workflow.
I find that a useful distinction. “Multiple AI agents” tells us something about how a system is organised. It does not establish that the system independently makes final scientific decisions.
Measuring patterns in tumour tissue
In treatment research, AI is used to analyse digital images of tumour tissue. The system helps measure biological features, such as the amount and location of particular proteins. Researchers then examine whether those measurements are associated with treatment outcomes. Exploratory clinical research findings have been reported for such an application.
Where is the AI? In the image analysis used to recognise and measure biological features. This is a very different application from a language model that writes a report.
There are also several questions to answer. Does the system measure reliably? And what does that measurement mean for a treatment? An exploratory finding does not yet provide an answer for every patient or every clinical application.
A digital twin: is that AI too?
Digital twins are used in manufacturing and development to investigate processes. Laboratory systems in which AI and machine learning play a role in experiments have also been described. These are different applications, each of which needs to be examined on its own terms.
A digital twin is a digital representation that allows you to investigate a real process or system. That can easily sound like AI.
But is it? Not necessarily. Such a model can be based on known physical relationships and fixed calculations. It may also incorporate machine learning, for example to learn relationships from measured data.
The name alone tells us too little. You need to know how the model works and what it is used for. Nor do a digital twin and an AI-driven laboratory automatically form a single integrated system.
The regulator uses AI too
The US Food and Drug Administration also uses AI itself. In June 2025, it introduced Elsa, an internal assistant for tasks including reading, comparing and summarising documents. The FDA describes applications in scientific reviews and identifying priority inspection targets.
In May 2026, it added capabilities including document generation, data analysis and custom AI agents. The FDA also began connecting the tool to a consolidated data platform, reducing the need for staff to upload documents manually.
Where is the AI? In the language models that process information and generate text, and in agent capabilities that can support multi-step tasks. Bringing data sources together is not itself AI, but it makes that support more accessible.
I find this an interesting development: working practices are changing on the dossier review side too. According to the FDA, experts verify inputs, analyses and how outputs are used. The described deployment therefore supports reviewers’ work; it does not demonstrate independent approval of medicines by AI.
The label tells only part of the story
What I find interesting about these examples is that “how much AI does it contain?” is not easily expressed as a percentage.
An application can consist largely of ordinary software, while a single AI step has considerable influence on the outcome. For example, by determining which information appears in a summary or how a report is classified.
So, at the next presentation about “AI in pharma”, I would ask three questions:
- What exactly does the system do?
- Which step uses an AI model, and what is handled by fixed rules or calculations?
- What is the output subsequently used for?
That makes it clearer what we are actually discussing. A writing assistant, a classification model, a research assistant or a model of a manufacturing process.
And it helps us assess what the application contributes to everyday work.
The cases are based on publicly described applications, available on public websites. They have been simplified and presented without company or product names to focus on how the technology works. The examples include both reported operational use and exploratory research. The explanation of the AI component is an interpretation of the described functions; the underlying sources do not provide the full technical architecture for every application.
For the FDA example, two official publications were also consulted on 28 October 2026: the introduction of Elsa, 2 June 2025, and the expansion of AI capabilities and the data platform, 6 May 2026. These are accounts published by the FDA itself.