SPEAKERS:
Javier Jiménez, Chief Medical Officer, PharmaMar
Joaquín Labado, Global Managing Director, Healthcare & Life Sciences, Globant
KEY TAKEAWAYS:
• The real bottleneck is not AI generation power but structuring and activating the data that is sitting inside pharma's information systems
• PharmaMar is innovating in R&D through an agentic system, "CombiVista", that narrows 8,000 potential drug-drug-indication combinations to a ranked top 10, 15x faster than the normal process
• In commercial, Globant implemented an agentic system that manages more than 10 brands across two markets, running 15 months in production
• Harnesses that allow AI agents to operate in production create compounding returns competitors cannot easily replicate
"To develop a new drug can take 12 to 15 years with a very low success rate," Javier Jiménez told the audience at Pharma 2026. He put a number to it: "It's estimated around 8% of the drugs that enter in Phase I reach commercial status. This rate could be even lower in oncology, to 5%."
Those figures are familiar to anyone who has sat through a pipeline review. What gets less attention is what they imply for how drug development decisions get made, specifically, what they demand of the humans making them.
The volume of knowledge required to navigate a single development decision has grown beyond what any team can reliably analyse. Preclinical data, competitive intelligence, regulatory precedent, commercial access modelling, biomarker literature — all of it arriving faster than any group of scientists can read, weigh and integrate. Jiménez named the constraint plainly: "The reality is no human can really integrate all the knowledge that all these decisions demand."
Jiménez and Globant's Joaquín Labado arrived with separate deployments, one in R&D pipeline prioritisation at PharmaMar, one in commercial brand management, and a shared architectural conclusion: pharma's most consequential AI investment decision is not which model to run, but what knowledge infrastructure surrounds it.
Generation is solved. The harnessing problem isn't
The dominant AI investment narrative in pharma over the past two years has centred on capability: which foundation model, which vendor, which use case to pilot next. Labado's opening reframe was direct. "Generation is no longer the problem. Harnessing it is. So what is the system that you put into place?"
For organisations running dozens of AI pilots, that question cuts. Most of those pilots share a structural flaw: they query whatever data is available at the time of the run, produce an output, and stop. Nothing is retained. The system knows nothing more after a hundred runs than it did after one. The capability is real; the accumulation is absent.
The data maintenance problem underlies this. Organisations have spent two decades and substantial capital on data lakes, master data management programmes and "single source of truth" initiatives — all of which treated knowledge as a storage challenge. The incentive structure was always the problem: maintaining a knowledge base is invisible work with no immediate return, and it consistently lost priority to activities with measurable near-term output.
Labado acknowledged the chronic failure directly: "The usual problem with data is that nobody has the incentive to keep it organised and to maintain it. But now the thing with AI is that we can really keep our knowledge bases very updated thanks to the agents doing that on our behalf."
The shift is structural. Autonomous agents perform continuous knowledge maintenance as a byproduct of their operational function, not as a separate project requiring dedicated headcount and executive sponsorship to survive budget cycles.
The financial logic for investing in that infrastructure becomes difficult to argue against when the cost of a single development misstep is made explicit. "Every particular clinical trial we implemented in a clinical development programme could cost from tens to hundreds of millions," Jiménez noted. "So, it's very expensive to make a mistake in this area."
At nine figures per trial, the ROI case for knowledge orchestration infrastructure is not primarily about efficiency. It is portfolio-level risk mitigation. A development decision made with incomplete or incorrectly synthesised evidence doesn't just waste the cost of that trial; it consumes the opportunity cost of the years and capital that could have been deployed against a better-ranked candidate.
Most pharma organisations are currently running AI as a capability layer on top of fragmented knowledge. The architecture these two deployments suggest should be inverted: the curated, continuously maintained knowledge base is the asset; AI agents are the activation layer. That inversion changes what a company builds, what it protects, and what it treats as proprietary.
Two different domains, one architectural pattern
PharmaMar's challenge was combinatorial. In oncology drug development, the number of theoretically viable drug combination candidates can reach into the thousands. Evaluating each against the full matrix of scientific evidence, regulatory precedent, commercial viability and patient access considerations exceeds any team's bandwidth, which means that in practice, many candidates don't receive the rigorous evaluation they warrant, and prioritisation decisions carry more uncertainty than they should.
"We have created with Globant what we call the CombiVista agentic system," Jiménez described, "with the human in the loop, with our scientists being fully involved, obtaining information, evaluating this information and providing feedback, identify from all those different potential opportunities what are the top candidates with the higher probability of success from a scientific point of view. But not only scientific — regulatory, commercial, access — so that those drugs can really reach the patient."
The multi-dimensional scope is deliberate. A system that optimises only for scientific promise reproduces the historical failure mode of assets that clear Phase II and stall at market access. CombiVista is architected to surface opportunities that are viable across the full development-to-patient pathway, which requires integrating knowledge domains that, in most organisations, sit in separate functions with separate data systems and separate vocabularies.
The compression it achieves is measurable. The system can already identify, for some compounds, "the top 10 priority areas to go from up to 8,000 potential combinations." That is not a marginal efficiency gain. It is a fundamentally different way of allocating scientific attention.
Discover more on this topic at Pharma Commercial Data & Tech Europe 2026 (4-5 November, London) Europe's collaborative home for data and tech pioneers. Visit the website here.
The commercial agentic deployment addressed a different pressure. Labado described managing more than 10 established brands across two large markets with constrained investment — a resource allocation problem that commercial leaders across the industry recognise immediately.
Established brands rarely command the internal priority or budget that pipeline assets receive, yet they represent substantial revenue and require ongoing medical, regulatory and commercial content management across multiple markets and regulatory environments. The combinatorial burden of brands multiplied by markets, content types and approval cycles quickly exceeds what a lean team can manage without systematic support.
"Skills in the end are instructions that we give to the agent aside from just a very monolithic big prompt," Labado explained. "And those instructions are written by our medical experts. So our medical experts can download their brains or their skills to the agent so that the agent later on replicates this behaviour when researching for scientific papers."
This is the design choice that distinguishes both deployments from generic AI implementations. Domain expertise is not approximated by a general-purpose model; it is operationalised by the people who hold it. Medical experts don't occupy a post-hoc review role, they author the behavioural instructions that determine how agents approach problems.
Two organisations with different domains, different data ecosystems and different decision cycles both arrived at the same structural pattern: a curated central knowledge base, specialised agents coordinated by an orchestrator, human experts embedded at defined governance points, and a mechanism for feeding reviewer input back into the system.
When organisations with nothing in common except the problem independently converge on the same solution architecture, that convergence is worth examining as a signal about what the problem actually requires.
Every run should make the next one better
"We do leave the traces on what happened in each content creation," Labado described, "so that in next run the agent can start from understanding what happened, what worked or didn't work, what were the inputs from the human reviewer so that the content that it creates on the next phase is better. That's what we call compound."
The mechanism is specific: each run produces an output and a — what was generated, what the human reviewer accepted, corrected or rejected, and why. The next run starts from that accumulated context rather than from a blank state. The system is not being retrained in the machine learning sense; it is developing institutional memory in the operational sense.
At PharmaMar, a parallel trust architecture reinforces this. Scientists receive recommendation with clear provenance, which sources the information came from and what confidence level attaches to each claim, so they can assess whether a recommendation is reliable enough to act on. That transparency is not incidental to adoption; it is the condition for it. And it is also what enables the system to learn which sources and synthesis pathways produce outputs that scientists actually accept.
The feedback loop runs in both directions: human judgment improves the system's behaviour; the system's improved behaviour earns expanded human trust.
Labado's most forward-looking claim follows directly from this logic: "The most valuable work AI can do is the work nobody thought to ask for." A system that has accumulated sufficient institutional context, enough cycles of human feedback, enough provenance data about what has worked, can begin surfacing intelligence proactively.
Connections between datasets that no one queued up a query for. Risk patterns that emerge from the aggregate of prior decisions Combination candidates that sit outside the search parameters a scientist would have set manually. This is the threshold where the system shifts from executing defined tasks to generating strategic inputs that humans would not have produced independently.
The diagnostic your AI portfolio is probably failing
PharmaMar production metrics offer a concrete baseline for what early-stage compounding looks like: the system achieves "more than 90% accuracy on retrieving the scientific information from internal and external sources" and delivers results "15 times faster than the previous process to review the information."
Those are credible early-stage. The more consequential question is how they change over the next 12 to 18 months of accumulated system learning, because a compounding system's value is a function of time in production, not deployment quality alone.
That dynamic introduces a strategic clock that most AI investment frameworks are not built to measure. Standard ROI models evaluate AI on a per-task or per-project basis: cost per output, time saved per workflow, accuracy against a benchmark. A compounding system's primary output is not the current run it is the improvement in every subsequent run.
The relevant competitive comparison is not "how good is my AI today" but "how much institutional knowledge has my system accumulated relative to competitors who started earlier or later".
Three questions surface the gap. Does the system draw on a curated, continuously maintained knowledge base, or does it query fragmentary data at each run? Does human expertise shape the system's behaviour through structured mechanisms, authored instructions and defined governance points, or are experts relegated to reviewing outputs they had no role in shaping? And does each run make the next one better?
If the answer to the third question is no, the organisation is operating AI as a point tool. It gains efficiency. It accumulates nothing.
The cases Jiménez and Labado presented — self-reported, early-stage and without independent verification — are nevertheless specific enough to carry an argument that theory alone cannot. Two production systems, operating across the industry's most consequential decision domains, built on the same architectural premise.
The model is not the moat. The memory is.
To get you highlights of Pharma 2026 faster, we are using generative AI technology to summarise the transcripts of the sessions. If you have any feedback about the summary, please contact lucy.fisher@thomsonreuters.com.
Discover more on this topic at Pharma Commercial Data & Tech Europe 2026 (4-5 November, London) Europe’s collaborative home for data and tech pioneers. Visit the website here.
