SPEAKER:
Tim Irfan, Commercial Programs Manager, Oracle Life Sciences, Oracle
Marina L. Martinez, Associate Principal, Digital and Data, Oracle
KEY TAKEAWAYS:
· An AI agent closed a revenue gap across eight countries and 24 reports in seconds
· Manual iteration across the same scope produced significantly less growth than the estimated target after days of work
· The AI explicitly requested permission to adjust pricing and showed advanced reasoning before executing
· The value of forecasters now centers on scenario planning, constraint definition and assumption governance, not repeated model execution
· Oracle's competitive differentiation rests on integrated data infrastructure, not algorithmic capability alone
"I've updated all of the files, 24 files manually. But you know what? We didn't get to that 2% revenue increase. We just got 0.6." Marina L. Martinez delivered that line as a roleplay setup, but the number it captures is not theatrical. It is the structural reality of multi-country commercial forecasting: iterative assumption-tweaking across disaggregated files compounds inconsistency rather than converging on targets. The process rewards persistence. It does not reward precision.
Tim Irfan framed the alternative as a single question: "What if we could actually start with the intended goal in mind rather than creeping closer and closer to the goal that we want to see?" What Oracle demonstrated was not a faster way to run the same process. It was an argument that the process itself runs in the wrong direction. The demo's most revealing moments, though, were not the ones where the AI moved quickly. They were the ones where it stopped and asked for permission.
Goal-First Forecasting Changes the Conversation
Conventional commercial forecasting has an epistemological design flaw that rarely gets named. It is structurally backward-looking: forecasters adjust inputs, observe outputs, and iterate until the numbers reach an acceptable range or time runs out. Tim Irfan named this loop without softening it: "The typical cycle for forecasting is you update assumptions, you run the model, you look at the outputs, rinse and repeat as often as you feel comfortable and until you see what you want to see." The phrase "until you see what you want to see" is doing real diagnostic work. It acknowledges that the process embeds confirmation bias by design. Forecasters are not running toward insight; they are running toward comfort.
During the session. Oracle's agent inverts the direction. Instead of projecting assumptions forward and hoping they approximate the target, the system starts with a defined outcome, such as a revenue goal, market share, or upside scenario, and works backward to determine which assumptions must change, in which markets, and by how much to achieve it. The demo executed this across eight countries, three indications, and created 24 reports in a timeframe that made the manual comparison almost absurd. Tim Irfan described what this represents as "not just efficiency by making a model run faster, it is flipping the script on how we're actually working and flipping the script on how forecasting has been done years and years ago."
The more significant implication is one the session didn't fully develop: goal-first forecasting changes what commercial planning leaders can promise in budget conversations. Instead of presenting outputs the model produced, they can present the conditions required to reach a defined target, which lever combinations are most accessible, which markets carry the most sensitivity, which assumptions are doing the most work. The conversation shifts from reactive reporting to structured optionality. That is a different relationship with executive leadership, not just a faster version of the current one.
Tim Irfan was explicit that this is not a roadmap claim. "We are now moving actually to define, simulate, decide. And this is not a wild dream about what might be possible tomorrow if the stars line up. This is possible today." For commercial leaders evaluating alternatives, that present-tense accountability matters. Define-simulate-decide is a named methodology, and any competing platform should be evaluated against the same capability, not merely against modeling speed or interface design.
The Guardrail Problem Is an Organizational Problem
The demo's pivot point came when Marina Martinez imposed hard caps before releasing the AI. "One constraint we are going to put is the diagnostic rate cannot be increased by more than 3% and neither the market share by more than 1%." Without those caps, the agent would have recommended whatever assumption changes were mathematically sufficient to reach 2%, including implausible ones. The guardrails are not a safety feature added to an otherwise complete system. They are the architecture. The AI's strategic credibility is entirely a function of the quality of the boundaries humans define for it.
Marina Martinez also established why this complexity is sector-specific rather than generic. "In pharma, pricing is not just a number, not just a figure. It's a little bit more complex than that. And it depends on the market rules, it depends on the pricing contract." The platform reflects this: its pricing module is pre-filtered to respect contract and regulatory constraints, which is why the AI asked permission before touching pricing and accepted the denial. But pricing is one assumption category. Diagnostic rates, treatment duration, patient share trajectory, market access timing, each carries domain-specific plausibility ceilings that vary by country and therapeutic area. The platform does not know these ceilings. The human must define them, in structured and defensible terms, before the model runs.
Tim Irfan argued that the platform is "not going to replace the brainpower of the many people touching a forecast but is there to support you and your teams to make evidence-based decisions quicker and with more confidence." That framing is accurate as far as it goes, but it understates what the transition requires. The platform does not simply redirect existing expertise toward higher-value tasks. It demands a different kind of expertise. The skills that made a forecaster valuable in a spreadsheet-driven environment, model architecture, formula logic, file version management, are largely displaced. The skills the platform requires; constraint articulation, assumption validation across markets, output interrogation, are not the same competencies, and most forecasting organizations have not yet built them systematically.
The teams most likely to struggle here are the ones where assumption governance is informal, distributed across email threads, quarterly review meetings, and institutional memory that lives in the heads of two or three senior people and has never been codified. The platform requires that tacit knowledge be made explicit and machine-readable before the first scenario runs. For organizations that can do this, the AI compounds expertise. For organizations that cannot, it accelerates the production of plausible-looking forecasts that lack strategic grounding. Speed applied to an undisciplined input set is not an improvement.
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.
Data Infrastructure Settles the Build-vs-Buy Question
Most AI forecasting conversations center on algorithms. Oracle's competitive argument rests on a different foundation. Tim Irfan described the asset base directly: "We have one of the largest electronic health record databases in the world, claims data and other data. So, with all of that knowledge and expertise, we build the data platform where this agent sits in." EpiDatabase, CancerMPact, EHR and claims databases, these assets enable automatic epidemiological funnel population across markets without manual sourcing or data procurement.
For commercial leaders running a build-vs-buy evaluation, this is likely the decisive variable. Modeling capability can be approximated by internal teams or point-solution vendors. The data infrastructure connecting epidemiology, real-world treatment dynamics, and claims evidence into a single forecasting substrate cannot be replicated at equivalent breadth or speed by teams starting from scratch. The asymmetry is not marginal.
The adoption model reflects an equally deliberate strategy. Marina Martinez addressed the most common institutional objection directly: "If you already have models that work well, it's just that it's uncentralized in a bunch of Excel files, we will take those files and make those work the exact same way that you already work in a cloud-based platform." The Q&A confirmed three pathways; new collaborative build, existing model validation, or as-is migration, indicating a service engagement calibrated to organizational starting point rather than a uniform product deployment. The architecture reinforces this. As Tim Irfan noted, the forecasting agent "sits in our Oracle Life Sciences AI Data Platform," which means adopters can access the full data ecosystem rather than a forecasting interface operating in isolation. The distinction carries real integration implications: the agent draws on the platform's underlying data assets without requiring separate procurement or connection work.
The Organizational Question the Demo Left Open
The interaction design embedded in Oracle's demo, the AI requesting permission before adjusting pricing, accepting hard caps without workaround, surfacing scenario options rather than issuing recommendations, is a change management artifact as much as a technical architecture choice. Oracle has pre-answered the question every forecaster in the audience was privately running: does this eliminate my role? The product's answer, built into its behavior, is no. But the product simultaneously demands a role that most forecasting organizations have not yet defined, trained for, or staffed.
The platform's stated trajectory extends well beyond forecasting. Tim Irfan described a clinical trial optimization agent targeting recruitment speed and per-trial cost reduction, and a separate label expansion agent that identifies next indications, estimates peak revenue, and calculates probability of clinical success. Neither was demonstrated live, so those claims carry less evidentiary weight than the forecasting demo. They do, however, signal the platform's intended scope: a full-lifecycle commercial intelligence stack in which the forecasting agent is the entry point, not the ceiling.
The demo closed with a clean quantified outcome. Marina confirmed: "The tool already updated all of the forecasting exercises for the eight countries, 24 reports in total, and we just achieved this new global growth rate of 2%, just like we expected at the beginning." Eight countries. 24 reports. Target achieved. Within the demo's controlled parameters, the technology performed as claimed.
The remaining question for commercial leaders is not whether the platform works. It is whether their organizations are ready to supply what the platform requires. Can your forecasting team articulate, in structured and defensible terms, the maximum plausible change to each assumption category — by country, by indication, by time horizon? Can they do so under budget-cycle time pressure, and defend those constraints when AI-generated outputs challenge existing forecasts or executive expectations?
If yes, define-simulate-decide is operationally available today. If no, the bottleneck is not Oracle's platform. It is the forecasting organization's capacity to encode its own expertise into rules an agent can actually respect — and most organizations have not yet begun building that capacity, let alone measuring it.
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.
