Beyond the hype: an insider’s view of where Generative AI could work in pharma

What If You Could Ask Your Pharmaceutical Company a Question?

There is a small problem with the way we talk about Generative AI in pharmaceutical companies. We often talk about it as if the big opportunity were to ask AI to write things for us:

“Summarise this report.”

“Draft this document.”

“Make this presentation more executive.”

Useful? Certainly. Transformative? Probably not.

Pharma has been producing information for decades. SOPs, change controls, regulatory submissions, quality records, manufacturing plans, artwork, product databases, supply-chain data, enough documentation to make a small forest nervous. The problem isn't that we don't have information. The problem is that our information doesn't always know that the other information exists.

And this is where I think AI gets really interesting. What if AI understood how the company works? Imagine asking: “I need to introduce a new supplier for a critical raw material. What do I need to do?”. Today, that might mean searching through several SOPs, asking Quality, checking a change-control procedure and eventually finding out that someone in Regulatory also needs to be involved. An AI connected to structured, controlled SOPs could instead identify the relevant procedures and explain the required steps, with links back to the approved sources.

The AI isn't replacing the SOP. It's making the SOP usable. And that distinction is important in pharma.

But the real opportunity starts when systems connect. Consider a much more interesting question: “Can we use the new carton for the next manufacturing campaign?”. That's no longer an SOP question. The answer could depend on:

·        Regulatory approval by market

·        Change-control status

·        Artwork approval

·        Packaging-material availability

·        Manufacturing planning

·        Quality status

All the information probably exists. Just not in the same place.

Today, someone may need to ask Regulatory, Packaging, Supply Chain, Manufacturing and Quality. Or, more realistically, schedule a meeting. Imagine instead asking the AI “Can we implement the new carton in the October campaign?”. The AI checks the relevant systems:

Regulatory: approved in four of five markets.

Change control: implementation phase.

Artwork: approved.

Packaging material: available.

Manufacturing: campaign scheduled for 15 October.

And gives you the conclusion: “Implementation is feasible for the four approved markets in the October campaign. Regulatory approval for Spain is still pending”.

Now we are talking about something much more interesting than a chatbot. We are talking about an intelligent interface to the company's operational knowledge. The AI doesn't need to know everything. In fact, in pharma, we probably shouldn't want it to. The AI should know where to look. The controlled systems remain the sources of truth. The AI retrieves the relevant information, connects it, explains it and helps the user understand what happens next.

That is very different from asking a generic LLM: “What is the regulatory status of this change?”. The first approach says: “Go and find the answer in our controlled sources”. The second says: “Give me your best guess”. Only one of those should make a Quality organization comfortable.

There is a catch. Data.

Everyone wants an AI assistant. Fewer people want to clean up the databases that the AI needs to use. But if the same product has three different names in three different systems, markets are represented differently, document versions are unclear and nobody knows which database is the source of truth, AI will not magically fix it.

It will simply produce a very articulate version of the mess. So the foundation is not the LLM. The foundation is: structured data + controlled documents + connected systems + clear ownership + permissions + traceability.

Then you put the AI on top. From answering questions to understanding what happens next. This is where things become really interesting. Today we might ask:

“What does the SOP say?”

Tomorrow we could ask: “What do I need to do?”

And eventually:

“What is blocking this change?” or,

“Which changes could affect the next manufacturing campaign?” or,

“We have a new regulatory approval. What needs to happen before we can implement it?”

The AI could connect Regulatory with Manufacturing. Quality with Supply Chain. Packaging with Regulatory. Not by replacing these teams, but by removing the enormous amount of organizational archaeology required to get information from one team to another. We don't think the biggest opportunity for Generative AI in pharma is generating more information. We already have plenty. The opportunity is to make the information we already have usable.

A Regulatory professional should not need three emails and a meeting to find out when the next packaging campaign is.

A Quality professional should not have to search five systems to understand whether a change is ready for implementation.

And a Manufacturing colleague shouldn't have to become a detective to find out which regulatory changes affect the next campaign.

The human still makes the decision. AI makes sure the human has the right information at the right time. Perhaps the future pharmaceutical AI is therefore not an AI that “knows pharma”. It is something more useful: An AI that knows where the company's truth lives, and can bring the right pieces together when you need them. And that might be a much bigger transformation than asking it to write another PowerPoint presentation.