Every pharmaceutical analytics team has someone making the case for artificial intelligence right now. The pitch sounds appealing. Feed a model enough historical launch data, clinical readouts, and prescribing trends, and it returns a forecast in hours instead of weeks. The promise is speed, and speed is hard to argue against when a portfolio decision is waiting on the answer.
Stelios Tzellos of the UK has watched this argument build across the industry, and his response is more careful than most. As a professional working in business insights, analytics, and oncology marketing at AstraZeneca, with earlier roles at GlobalData and IQVIA, he has seen what happens when a fast answer turns out to be the wrong one. The question is not whether AI can build an oncology forecast. It can. The real question is whether the people running it understand the biology well enough to recognize when the output makes no sense.
The Speed Problem
AI is very good at finding patterns in data. That strength becomes a weakness in oncology, where the patterns shift every time a new mechanism of action enters the market. A model trained on how previous drugs performed will assume the next one behaves like its predecessors. Sometimes that holds. Often it does not. A first-in-class immunotherapy does not follow the uptake curve of a late-line chemotherapy, and a model that has only ever seen chemotherapy launches will say otherwise with complete confidence.
That confidence is the trap. A spreadsheet built by hand carries visible assumptions that a reviewer can challenge. A model that produces a single clean number hides its assumptions inside layers of training data nobody in the room has examined. The output looks authoritative precisely because it arrived without a human argument attached to it.
Where the Models Learn the Wrong Lessons
Training data reflects the market that existed, not the market that is coming. In oncology, those two things diverge fast. Combination regimens reshuffle treatment sequencing. Biomarker testing changes who counts as an eligible patient. A companion diagnostic with low adoption in community settings can stall a drug that looked unstoppable on paper.
A model does not know any of this unless someone teaches it, and teaching it requires a person who understands why these forces matter. Tzellos studied Biochemistry and Molecular Biology at Imperial College London before moving into healthcare analytics. That foundation lets him look at an AI-generated forecast and ask the questions the model cannot ask itself. Does this assume a testing rate the clinic will never reach? Does it treat a future combination backbone as a standalone competitor?
What the Science Background Adds
During his doctoral research, Tzellos studied Epstein-Barr virus gene regulation and the molecular basis for why type 1 EBV transforms cells more efficiently than type 2. Work like that trains a person to treat biological systems as variable and context-dependent rather than fixed. The same input produces different outputs depending on conditions that are easy to overlook.
Bring that mindset to an AI forecast and you stop accepting the number at face value. You start testing it. You ask what the model assumed about diagnosis rates, about line-of-therapy distribution, about how quickly oncologists actually change their prescribing habits. The model gives an answer. A trained analyst gives the answer a reason to be trusted or rejected.
Using the Tools Without Trusting Them Blindly
None of this means AI has no place in pharmaceutical forecasting. It means the tool belongs in the hands of people who can check it. At GlobalData, Tzellos built epidemiology models and competitive assessments for oncology and haematology indications including Hodgkin’s lymphoma. At IQVIA, he worked with global clients on forecasting and evidence-based strategy through the Analytics Center of Excellence. Both roles taught the same lesson: a model is only as good as the judgment applied to its assumptions.
AI can draft the forecast faster. It cannot decide whether the forecast is right. That decision still belongs to someone who has read the science, watched the market behave, and learned where the easy answer falls apart. At AstraZeneca, Tzellos leads cross-functional projects that depend on exactly that kind of judgment. The technology will keep improving. The need for people who can tell a plausible forecast from a correct one is not going anywhere.