AI
AI transformation in life sciences starts with processes
Why adding AI to an unchanged process rarely moves the KPIs that matter
There is huge potential for AI in the life sciences sector. But that potential is not being realised. Broadly speaking, tasks are not getting completed substantially quicker, efficiency isn't improving – in fact, pick a KPI, any KPI, and AI is unlikely to be moving the needle. At least, not in a transformative way.
“Measured with a stopwatch, people are saving a bit of time, 10 minutes here, 15 minutes there, and they feel more efficient,” says NNIT's Chief AI Officer Sam Laermans. “This has benefits, but it is not showing up in throughput.”
So, what's the issue? Is the technology the problem?
“The technology works, it's getting better by the day, and, when done right, AI can move the needle,” Sam Laermans continues. “So, no, it's not the technology. The problem is how AI is implemented."
“What we see is that there are very few companies that can say, ‘we gave people licenses or a system, now they are working differently, and the result is the process is faster or cheaper."
Sam Laermans, Chief AI Officer, NNIT
“Most companies implement AI, but they don’t change processes. That’s understandable because in life sciences organizations, a process isn’t just how work gets done. Instead, it’s a controlled, documented, and often validated state. When you change one step, you trigger a change control process that includes a formal impact assessment, revised SOPs and templates, documented retraining for every affected role, and the potential revalidation of computerized systems. In some cases, there is also an assessment of whether the change affects what has been filed with regulators. As a result, companies take the path of least resistance and layer AI on top of the process they already have.” Sam Laermans explains.
To realise the potential of AI, to properly move the needle on critical KPIs, and to achieve transformative change, AI implementation issues need to be understood and addressed.
Real-world example: Document drafting
Take regulated document authoring in the pharmaceutical industry, such as clinical study reports (CSRs), protocols, investigator’s brochures, and submission documents. Large language models (LLMs) and AI assistants like Copilot are really good at drafting documents.
“One of the use cases we have seen is a company that gave its clinical writers access to an AI assistant to help with document drafting,” Sam Laermans says, and continues: “The process, however, remained unchanged.”
“The writer opens up Microsoft Word and starts writing. When they need to insert a value, rather than going off to a different system to get that value, they stay in Word and ask the AI assistant to get the value for them.”
“That should speed up the work of the author, right? They don't have to spend time going off to find the values they need for the document, so they can complete the task in a shorter period of time. It doesn't work like that, though, primarily because the process hasn’t changed,” Sam Laermans explains.
The author has an AI tool, but:
Review cycles still run through email and Teams comments, with multiple rounds of SME, medical, regulatory, and QA review
Comment-resolution and alignment meetings are just as frequent as before
Hand-offs between functions and QC/QA sign-off still create the same queues, so cycle time from first draft to approval is unchanged
Authors still write the same number of paragraphs
Every value the AI assistant retrieves still has to be QC’d against the source data – such as the tables, listings, and figures (TLFs) – before the document can be approved.
Some technical jargon is needed to expand on the last point – deterministic and non-deterministic systems. Out-of-the-box AI tools like Copilot are non-deterministic. This means they are not sufficiently predictable; i.e., with the same input, the outputs might differ.
What this means in practice is that an author or another human-in-the-loop (such as a reviewer) has to check the accuracy of the values the AI retrieved. This is an essential step for compliance, but it also quickly eats into the time that was saved by using AI in the first place.
For AI transformation, you need to change the process
“As mentioned, the tech is not the problem,” continues Sam Laermans. “It is entirely possible – excitingly possible – to create AI systems with deterministic components that ensure the same inputs always produce the same outputs.”
“We have done this for other pharmaceutical organizations, where, for example, document authoring is moved out of Word entirely into a tested and validated AI system. The system manages who reviews what and when, greatly reducing meetings, waiting times, and other delays. And because the system is tested and validated, if the AI says the value is 6 grams, you can trust it is 6 grams,” he says.
“Non-deterministic elements remain, such as checking the phrasing that the LLM component of the system used to write the document aligns with regulations and internal communication standards. This is a human-in-the-loop part of the process, but it’s a different type of review compared to double-checking that AI contributions are accurate, as in the first example.”
“The difference is not the AI itself, but what it allows the process to become. Authors no longer look up and re-key values because data flows directly from the source systems. Reviewers no longer need to QC every number, so review processes shift to the scientific and regulatory message. And instead of rounds of emails, Teams comments, and meetings, the document follows a single, controlled path from first draft to approval.”
“Validation is what makes all this possible. It is validation that allows the QC step to be designed out of the process rather than simply sped up.”
Sam Laermans, Chief AI Officer, NNIT
With a new process, supported by a tested and validated AI-powered system, the needle on critical KPIs starts to move.
Join the Webinar
Sam Laermans will be presenting a webinar alongside NNIT's AI Governance Lead Emma Skovsted-Andersen on how to develop AI systems like those described above, i.e., AI systems that deliver transformative change. They will be joined by Nadir Mohamed from Anthropic.
The webinar will show that successful AI implementations are about much more than writing a good prompt.
“Successful AI implementations are about designing systems and processes that are truly trustworthy when they are applied to GxP applications,” says Sam Laermans, and continues: “That’s what we are going to discuss in the webinar, that AI transformation is possible with really smart system design and thinking.”
Sign up below.
