Why a Correct Decision is More Important Than Saving Time in Drug Discovery

Stavroula Ntoufa spent 14 years in academia before joining Causaly. She argues that the most valuable thing in pharma R&D is the judgment a scientist applies at every decision gate. But this expert decision-making isn't written down anywhere. AI agents are changing that.

Stavroula Ntoufa did not leave research because she stopped loving it. Fourteen years in, and despite collaborating with big pharma on various projects, she had never seen her own research reach the clinic. And the higher she climbed in academia, the less time she spent on science and the more time she spent on grant drafting and fundraising, because experiments grew more expensive.

She joined Causaly six years ago, before AI had gained major traction in the field. The Causaly team was engineering knowledge graphs and working on natural language processing. What drew her to the company was a set of solutions to problems she had faced many times, quick access to information, answers backed by evidence, and a trail she could follow to decide what to trust.

Six years later, she has watched the questions from customers evolve. Scientists asked about hallucinations, then moved past it. They asked how to phrase a good prompt, then moved past that too. Now they ask about sequencing AI agents and combining products.

What has stayed constant is the harder problem underneath. Standard operating procedures capture the steps of scientific work, not scientific judgment. And as experienced people leave big pharma in waves of layoffs, that judgment leaves with them. She shares how AI agents are solving this problem.

You've said the core problem for scientists hasn't really changed since you left academia. What is it?

Information overload that exceeds human scale. The volume of biomedical knowledge grows faster than any person can absorb, and that absence of coverage introduces bias into research, because you cannot read everything. Important signals get buried in the noise, and contradictory evidence becomes hard to surface even though a decision often depends on it.

There's a second layer that people underestimate. When I worked on large European projects with multiple teams, insights got lost between them. Information sat in silos with no easy way to share it. That still happens, and in big pharma it happens at a much larger scale.

The result is that scientists spend their time hunting for information instead of building scientific hypotheses. The hypothesis is the thing that eventually becomes a target, then a trial, then a drug. It's the most important work in R&D, and right now it's constrained by time that gets spent elsewhere.

Even when a scientist finds the evidence, you say the harder step comes next.

Yes. You have the evidence; now you have to translate it into a decision. To do that you have to understand which evidence is most reliable, whether the literature conflicts, whether the evidence is mature enough to support the decision you're about to make, and where the uncertainty sits.

That last part matters more than people admit. AI tools will hand you a great deal of information. The scientific decision is still yours.

Talk about the gap with standard operating procedures (SOPs).

In R&D there are workflows and there are SOPs, and they tell you to do step one, step two, step three. The scientific judgment is not captured there.

You can read the SOP and still have no idea how a senior scientist with twenty years of experience thinks through those steps. What they look at, what makes them stop, what makes them proceed. That intelligence is not codified anywhere in big pharma. It lives with each individual.

This creates a challenge when people leave companies. There have been major layoffs across the industry, and when experienced scientists walk out, the knowledge walks out with them. Since it was never written down, it's gone. I don't hear anyone talking about this, and I think it's one of the most expensive things happening in pharma right now.

So what does codifying that judgment actually look like?

A good scientific workflow produces outputs that support a decision, showing what the evidence supports, what remains uncertain, what contradictions exist, which risks and gaps matter most, and what decision all of that enables.

To build one, you need input from a scientist who has run the process for year and makes these decisions for years. And you need to codify the decision gates. Which sources am I looking at? What standard am I applying? What completeness do I need before I move to the next step?

We started doing this with customers, and I found it difficult for them. You build a workflow backwards, from the result you want. Scientists are trained to think the other way around. They've never written down what they're thinking when they make these decisions, and now we're asking them to. It takes a push.

The payoff is that a senior scientist can build a workflow that junior scientists then use. Everyone gets the right data, and everyone makes the decision with the same standards.

If you could change one thing about how pharma evaluates AI, what would it be?

Most organizations still evaluate AI in terms of efficiency. Faster searches, faster reports, faster summaries, time saved. Those gains are real. But it's also a smaller opportunity.

The real value in AI is decision quality. Say you identify a target, you fail to find the contradictory evidence in the literature, and you push that target toward clinical trials. You've just committed enormous money, time, and resources. Now say you catch that contradictory evidence and pause the program instead. You've saved all of it.

A correct decision is more important than saving time. AI can support that, because it can go through everything and expose the uncertainty. The same holds for safety and surfacing a hidden signal early, before patients ever receive the drug.

That's how pharma should judge these tools. Better decisions, with more evidence behind them.

As AI becomes entrenched, where does this leave the scientist?

Science moves from execution to orchestration. The manual work, including searching, gathering evidence, organizing findings, monitoring the literature, and writing summaries, gets automated. The scientist starts operating like a director of those systems.

And a director makes the important decisions. You define the critical question, which is deeply scientific work, because the right question leads to the right answer. You set the evidence standards. You build the hypothesis and evaluate it. You challenge assumptions.

Human expertise becomes more valuable, not less. The more information you can access, the more judgment you need to interpret it, prioritize it, reason across domains, and interrogate the output critically. Deciding whether what's on the screen makes sense still belongs to a person, and it remains the critical work of a scientist.

‍