2026 Symposium - Sponsor

Why did Brand X lose 3 points of share in the Southeast — and how much of it is a formulary access problem versus a field execution problem? Today, that question gets split across two teams, four data sources, and a three-week turnaround. In this workshop, you'll answer it yourself — hands-on, in minutes.

Working in a live sandbox with realistic (anonymized) Rx, claims, payer, and CRM data, attendees will conduct a complete multi-source root-cause investigation using natural language: decomposing the share change, separating access-driven from execution-driven effects, reconciling sources with different grains and definitions, and tracing every number back to where it came from. Then you'll take it a step further — turning the finding into a finished, decision-ready work product, and converting the one-off investigation into a reusable app and an always-on monitoring agent, so the next occurrence is caught the morning it starts.

This workshop explores:

• Natural-language root-cause analysis across sources that don't naturally line up — Rx, claims, payer/formulary, and CRM data
• AI-augmented approaches to separating payer-driven share shifts from field execution effects
• Producing finished work products — the exposure quantified, the response drafted — rather than charts that still need assembly
• Converting a one-off investigation into a reusable analytical app and an always-on monitoring agent, hands-on
• Real-world pitfalls and learnings from applying AI analysis to pharma commercial data

The format is a mix of live demonstration and guided hands-on exercises using the Tellius platform. Whether you're in brand analytics, market access, field excellence, or commercial data science, you'll leave with practical root-cause skills, an agent you built yourself, and a concrete sense of what changes when a three-week question takes minutes.

Pharmaceutical forecasting has traditionally relied on a fragmented, multi-step process spanning epidemiology assessment, market research, assumption development, and scenario evaluation across disparate tools, making consolidation and traceability difficult. This workshop explores an agentic AI system that translates a forecasting request into a structured, reviewable plan and orchestrates specialized agents to gather evidence, evaluate epidemiology, assess market dynamics, identify analogs, and quantify market-event impacts, with every input tagged by provenance and confidence for full transparency. By separating agent-driven reasoning from deterministic forecasting engines, the system produces an integrated epidemiology-to-revenue forecast that remains reproducible, auditable, and interactively refinable while preserving human oversight.

As life sciences data ecosystems grow to include more sources, reporting tools and cross-functional dependencies, many pharma organizations reach a point where different teams report different versions of “the truth” - even when they believe they are working from the same data.

This interactive session examines why that happens, why the underlying issue may not be a lack of data or technology, and what it takes to resolve it.

Through an anonymized real-world case and realistic scenarios, participants will explore how inconsistent metric definitions, missing metadata, duplicated reporting and unclear governance and ownership can undermine trust and create decision paralysis.

Participants will examine:

• Where conflicting metrics and definitions originate across functions.
• How to distinguish intentional, role-specific views from accidental inconsistency.
• What governance, ownership and shared context must be in place before AI can reliably support decision-making.

Attendees will leave with a practical approach to clarifying KPI ownership and decision rights, strengthening shared metric logic and reporting foundations, and building a phased roadmap across governance, data, reporting and AI.

Forecasting plays a critical role in biopharma decision-making, influencing portfolio strategy, launch planning, commercial investments, and resource allocation. This workshop provides a practical introduction to forecasting fundamentals, helping participants understand how forecasts are built, how market insights and data are translated into assumptions and decisions, and how emerging technologies, including AI, are reshaping the future of forecasting. Designed for professionals new to forecasting or those who work closely with forecasting teams, the session combines industry frameworks, real-world examples, and practical applications to build foundational forecasting knowledge.

Organizations across life sciences are investing heavily in Generative AI and Agentic AI, yet many struggle to move beyond successful pilots to achieve sustained adoption and measurable business impact. This interactive panel discussion will explore the organizational, behavioral, and governance challenges that often limit AI adoption, including trust, change management, training, and value realization.

Drawing on real-world experiences, panelists will share practical strategies for embedding AI into everyday workflows, building user trust, creating champion networks, and sustaining momentum at scale. Attendees will leave with actionable insights for transforming AI pilots into lasting habits that drive meaningful business outcomes.

Choosing comparable product launches is a consequential step in pharmaceutical forecasting, yet analog selection often depends on familiar examples and subjective judgment. This case study examines how a global biopharmaceutical company and Axtria developed Analog AI to bring a consistent evidence base and structured review to that process.

The session explains how curated clinical, market, access, and company attributes support expert-led filtering, LLM-assisted recommendations, and machine-learning predictions of launch performance. A narrated walkthrough using a blinded asset will show how forecasters assess candidate analogs, interpret recommendations, and review competitive sets through a governed Market Basket workflow. We will discuss how to evaluate recommendation quality, distinguish a plausible explanation from supporting evidence, and decide where human review remains essential.

The presenters will share lessons on forecaster trust, proprietary-data integration, and phased adoption, alongside the role of agent-assisted curation in keeping the evidence base current. Attendees will leave with a practical blueprint for designing an analog intelligence capability and evaluating its use in their own forecasting processes.

More information coming soon.

Most of us measure analytics success by outputs: dashboards built, reports shipped. However, the metric commercial leaders care about is time-to-decision which can have a direct and often significant impact on business outcomes. For us, that process often took days. During a business review, a leader would ask a follow-up question, discover the answer wasn't in the deck, and hear the familiar response: "Let's take that offline." The question became a ticket, joined by dozens of others in the queue, and momentum stalled while teams waited for answers.

We wanted to close that gap and put a number on it. We first designed and built a governed semantic model that translated complex business logic into a trusted, reusable analytics layer. By publishing that model in ThoughtSpot, analysts gained a self-service foundation for exploration, and business users gained access to Spotter, ThoughtSpot’s agentic search experience. Users can ask questions in plain English and receive trusted answers without SQL, ticket queues, or report rebuilds, while the agent handles the follow-up questions that previously stalled or required analyst intervention.

Two outcomes emerged from this work:

• The first is we reduced decision latency. Questions that sat in the queue for a day or two are now answered in minutes, on the spot. More than thirty people across brand, field, and leadership now explore their data independently, eliminating roughly two-thirds of that fifty-plus monthly backlog. The result is over 100 analyst hours returned each month—more than 1,000 annually—while preparation time for major business reviews has fallen to approximately one-quarter of previous levels.
• The second finding was less expected. By reducing the friction to data driven insights, the questions changed. Leaders increasingly asked "why did it happen?" and "what should we do next?" rather than “what happened?”. Our usage data offers a window into how curiosity evolves when the cost of exploration approaches zero, creating a more analytical and decision-oriented culture.

None of this was really about the software. The harder shift was what it did to our analysts. Analysts moved from answering questions one at a time to owning the governed semantic model that answers them at scale. They became stewards of trusted definitions rather than operators of a request queue. That is the part that made leaders willing to act on an answer in the moment: they knew the business logic behind the agent’s answer was owned, governed, and accountable.

Sharing our journey, we will also be candid about what it took to get there. Adoption did not happen automatically. A handful of early answers exposed gaps in the model and threatened confidence before we strengthened governance and refined the underlying business logic. The longest pole in the tent was not the technology—it was establishing trust in the data.

This is what "the convergence of data, talent, and AI" looked like in practice: not analysts replaced, but decision-makers equipped and empowered. We will share how we measured the impact, governed the semantic layer, analyzed usage patterns, and learned from missteps, providing a practical framework for organizations looking to reduce decision latency and scale trusted, AI-enabled analytics.

Oncology trigger programs combine claims-derived signals, biomarker-testing indicators, treatment-pattern information, healthcare professional (HCP) targeting, and field engagement data. Multiple alerts may be generated for the same HCP, requiring users to reconstruct the “why now” from fragmented records. We developed a context-engineered generative AI workflow to convert multi-source trigger data into concise, standardized summaries.

More information coming soon.

Every pharma commercial organization now wants an agent that can look at an account, decide it deserves more resourcing, and act with minimal human friction. Almost none are ready to build one safely, because the account the agent reasons about was assembled by a resolution method nobody ever interrogated. Two reasonable ways of mapping a provider (NPI) to a parent account can disagree on a meaningful share of a book of business, and that disagreement propagates silently into every tier, recommendation, and eventually every autonomous action built on top of it.

This session presents a full-stack account intelligence architecture in two parts. First, a proven, production-tested foundation: a dual-bridge resolution engine (calls-based and ZIP-based, unified behind a common interface) and a six-component, config-driven Center-of-Excellence scoring model, built to make resolution bias visible and auditable instead of hidden, validated through 28 passing unit tests and a synthetic tier-flip example showing how the same account lands Platinum under one method and Silver under another.

Second, built directly on that foundation, we present the design for an Account Strategy Agent: a bounded, explainable agent that monitors tier movement, drafts specific resourcing recommendations with a documented rationale tied to exactly which CoE components moved and how confident the underlying resolution is, and routes every recommendation to a human owner for approve, modify, or reject before any budget or field-force action is taken.

More information coming soon.

More information coming soon.

Commercial decisioning today is a tangle of segmentation, targeting, next-best-action, and omnichannel tools — stitched into rigid workflows that rarely flex across a brand's lifecycle. This session introduces the Adaptive Decision Intelligence Engine: a network of collaborating AI agents that reframes decisioning as continuous reasoning, not a fixed sequence of steps. Give it an objective — "accelerate adoption among high-potential cardiologists" — and it reasons through who to prioritize, what's holding them back, what to say, and how to reach them, coordinating field, medical, marketing, and access into one coherent strategy. The same architecture scales with your data: it reasons from playbooks and analogous-brand memory when signals are thin, and progressively activates propensity models, hyper-personalization, and reinforcement learning as evidence grows. Come see how one agentic framework delivers sophisticated, explainable, and compliant decisioning across the entire brand lifecycle — from launch to maturity.

Enterprise analytics runs on table sprawl. To answer each new question, teams build their own processed, use-case-specific tables, so the warehouse fills with overlapping datasets that compute the same metric under different names, definitions, and logic - and quietly disagree. Every one of these tables carries its own data-engineering pipeline to maintain. When agentic GenAI assistants are pointed at this landscape, they tap whichever use-case table a team happens to own, inheriting its fragmentation, its maintenance cost, and its contradictions. The missing piece is a shared, governed, AI-consumable data foundation the agents can reason over - not another table.

We present a foundation-first data architecture that classifies enterprise data into three explicit tiers. Raw datasets arrive from vendors untouched. Foundational datasets - claims, sales - are cleaned and conformed once into durable, enterprise-wide data products. Functional datasets hold use-case-level metrics filtered from the foundational layer. The governing principle inverts current practice: agentic systems tap the foundational layer wherever possible instead of proliferating functional tables. Foundational datasets are re-engineered for AI consumption - normalized, with every column unique and every column name and definition written to be unambiguous to both a human and a model.

Agentic AI is creating new opportunities for market intelligence teams to move beyond traditional search and standalone generative AI tools to-ward coordinated workflows that can plan, re-search, evaluate, and synthesize complex business questions. But realizing this potential re-quire more than simply adding AI to an existing process. Teams must determine how to structure tasks, assign specialized roles, validate information, and maintain human oversight while producing insights that decisionmakers can trust.

In this session, audience will be exposed to a practical framework for applying agentic AI to market intelligence. The session will walk through how a complex intelligence request can be decomposed into a structured research plan, delegated across specialized AI agents, and in-dependently evaluated before being translated into a business-ready deliverable. Particular attention will be given to source attribution, confidence assessment, cross-verification of critical claims, and safeguards designed to reduce un-supported or fabricated conclusions

As pharmaceutical organizations increasingly adopt Generative AI and Agentic AI, the challenge is no longer model capability but ensuring AI systems understand enterprise context, governance requirements, business rules, and commercial objectives. Without this context, AI-powered recommendations can be technically accurate yet commercially ineffective, reducing trust, explainability, and adoption.

This session introduces a practical Context Engineering framework that enables AI agents to operate with enterprise awareness by combining governed data foundations, business ontologies, organizational knowledge, decision policies, and human oversight. Attendees will learn how pharmaceutical organizations can build trusted, auditable, and scalable AI solutions across commercial analytics use cases including HCP engagement, next-best-action recommendations, forecasting, market access analytics, and launch planning.

Using a real-world pharmaceutical workflow, the session will demonstrate how organizations can transform fragmented data, business knowledge, and operational processes into actionable intelligence that supports faster, more informed commercial decisions. Participants will gain a practical blueprint for operationalizing agentic AI while maintaining governance, transparency, compliance, and business trust.

Key Takeaways: By the end of this session, attendees will be able to:

• Understand why context engineering is emerging as a critical success factor for enterprise AI and agentic systems.
• Learn how to bridge the gap between enterprise data, business knowledge, governance, and AI decision-making.
• Explore a practical framework for building context-aware AI solutions within pharmaceutical organizations.
• Understand how enterprise ontologies, business rules, and decision intelligence can improve trust and explainability.
• Learn best practices for deploying scalable and governed Agentic AI systems in commercial pharma use cases.
• Walk away with actionable strategies that can be applied immediately to their own AI transformation initiatives.

PRISM presents a supervised multi-agent framework designed to make AI-driven commercial analytics in life sciences more trustworthy and defensible. The session will show how governed data access, SME rules, validation, provenance, and human review can be embedded directly into the analytical workflow to improve reliability and decision confidence.

More information coming soon.

More information coming soon.

A Governed Semantic Context Layer for Trustworthy Agentic AI in Pharma Commercial Analytics

The session will outline how commercial decision intelligence can be advanced through effective forecast informed IC planning. The session will take the audience through different forecasting techniques that can help creating an informed IC Plan to ensure fairer goal allocations.

Learn how a Medical Care Gap (MCG) framework can help Medical Affairs teams identify, prioritize, and measure clinically meaningful care gaps. By linking scientific engagement activities to measurable patient outcomes, this approach offers a practical pathway from activity metrics to impact-focused decision intelligence.

AI is exposing biopharma commercial data environments for what they really are: heterogeneous data sources that lack proper integration, entity relationships, and data modeling best practices.

Commercial and IT leaders are discovering this the hard way, the moment they put an AI model on top of their data. They can't get consistent numbers, responses are slow, and token costs skyrocket for seemingly simple questions.

Teams have relied on analysts to wrangle the data together to fit reporting needs, mostly out of necessity, because the data architecture was never defined to support cross-domain or self-service querying.

Fortunately, there is a path forward for using AI for self-service insights and analytics. It starts with designing the right data model and finishes with a thoughtfully governed semantic layer connected to whichever AI platform your organization runs on.

This poster walks through the specific modeling failures that cause this breakdown, the architecture that fixes it, and resulting high-value AI use-cases for commercial teams.

This session presents a case study in moving a global pharma marketer from siloed, channel-specific marketing mix studies to a unified, AI-driven investment system. Built on Google Meridian and tailored to the client's data, systems, and business context, the platform brings together a diverse marketing mix — spanning digital ads, TV, streaming audio, print, outdoor media, and in-person HCP engagement — into a single cross-channel measurement, scenario simulation, and budget allocation environment, replacing a fragmented, channel-by-channel view with one true omnichannel view of performance.

This poster describes an ontology-driven knowledge foundation supporting Medical Affairs and commercialization AI applications through a semantic layer.

The model defines core concepts — diseases, assets, indications, evidence, healthcare professionals, accounts, organizations, and their relationships — and maps heterogeneous structured and unstructured data into a consistent representation used for customer and account intelligence, patient journey analysis, launch and brand analytics, and competitive monitoring.

Topics include ontology design, data source mapping, governance, and trade-offs between semantic richness, implementation complexity, and the use of existing healthcare terminologies.

This poster addresses one of the oldest unsolved problems in pharma medical and commercial strategy: identifying key opinion leaders and understanding what they actually believe. An expert's footprint is fragmented across publications, trials, congress activity, payments, funding, and digital channels with no common identifier; standard signals mislead (citation lag, name-only co-authorship); and belief signal today comes only from slow, qualitative advisory-board sensing. The poster presents a three-layer architecture that addresses each failure: identity resolution linking unlinked public datasets to a single HCP with match-confidence review; LLM-based belief and sentiment extraction from unstructured sources — entity-anchored, theme-classified, auditable to the evidence span, tracked over time; and transparent, decomposable influence scoring with per-channel formulas that drill to their evidence. Visitors will take away a concrete, auditable blueprint for measuring what experts believe, not just how much they publish.

An Agentic, Multimodal Analytics Engine Unifying Structured Data, Dashboards, and Documents for Commercial Excellence

Market research and competitive intelligence teams in pharma don't lack external signal: trial registries, congress abstracts, news, real-world evidence, sentiment, they lack a way to turn that signal into a specific next action a brand team will actually take. This poster presents a two-layer market intelligence architecture: an operational layer (personalized daily brief, automated news/alerts pipeline, congress/trial-timeline tracker) and an analytical layer (landscape exploration, KOL mapping, evidence-gap and sentiment analysis, semantic-layer querying), grounded by a citation-backed, hallucination-monitored AI Co-Pilot that answers only from a curated, brand-approved corpus. It then presents the honestly-scoped next step: evolving that Co-Pilot from a question-answering tool into a Recommendation Agent that proposes a specific next action such as update rep messaging, flag a publication opportunity, prioritize KOL outreach, tied to a named decision owner, governed by the same raw-vs-curated human validation discipline.

In relapsed/refractory multiple myeloma (RRMM), patients approaching relapse may be underrepresented in EHR-derived cohorts because progression is often not captured as a discrete structured event, care may be fragmented across settings, and patients may not appear in a contributing EHR system until referral to a specialist or initiation of a new treatment.

EHR-identified RRMM cohorts capture patients who have already progressed through multiple lines of therapy and entered a documented care pathway, but miss the broader population of myeloma patients approaching relapse who remain invisible until they arrive at a specialist. By the time they appear in structured data, the commercial and clinical window has narrowed considerably.

This session presents a three-stage real-world data framework developed and validated in RRMM that addresses this gap directly.

Stage 1: Patient Amplification Beyond EHR. Using linked claims, pharmacy, and clinical notes data, we first establish an EHR-confirmed RRMM reference cohort that serves as a clinically enriched “gold standard.” Longitudinal treatment patterns, drug sequencing signals, comorbidity profiles, and other observable features from these patients are then used to distinguish clinically similar patients in broader claims datasets from claims-based false positives. This digital twin matching approach surfaces patients who share the clinical fingerprint of RRMM despite lacking a confirmed RRMM designation in the contributing EHR data. In a tested cohort, the approach meaningfully expanded the identifiable population while maintaining clinical plausibility against the EHR-confirmed reference cohort and established epidemiologic estimates.

Stage 2: Transportability Across Claims Datasets. Once developed and validated using the EHR-confirmed reference cohort, the amplification algorithm relies on claims-observable features—such as longitudinal treatment patterns, drug sequencing, utilization, and comorbidity profiles—to identify likely RRMM patients. The algorithm can therefore be applied to other claims datasets without requiring patient-level linkage to EHR or clinical notes in each new source, provided the required variables are available. We evaluate consistency in population estimates across claims assets and define the validation and recalibration required when data coverage, structure, or completeness differs. This transportability supports scalable payer analyses, field force sizing, and patient-finding programs without dependence on EHR availability in every target dataset.

Stage 3: Earlier Prediction. The most forward-looking application of this framework is predictive: using treatment sequencing patterns and clinical signals present in the data 6 to 18 months before a confirmed RRMM designation, we show it is possible to identify patients who are likely approaching relapse before they are formally coded as refractory. This earlier identification window has direct implications for patient support programs, clinical trial recruitment, and pre-launch commercial strategy.

Together, these three stages constitute a replicable, defensible methodology for commercial teams, market access analysts, and real-world evidence researchers working in hematologic malignancies and beyond. The session will include explicit decision criteria for when amplified populations are appropriate for decision-grade versus regulatory-grade use cases.

Learning Objectives:

• Apply a linked multi-source data amplification framework to expand RRMM patient cohorts beyond EHR-identified populations using digital twin matching logic
• Evaluate the transportability of patient amplification methodologies across real-world data sources and assess population estimate consistency
• Use longitudinal treatment sequencing signals to identify patients approaching relapse/refractory status earlier in the disease course, before formal diagnosis code confirmation

Attendee Engagement Strategy:

Open with a visual: a Sankey diagram showing patient attrition from diagnosed myeloma to EHR-visible RRMM, making the "missing patient" problem visceral before introducing the solution. Mid-session, run a structured audience prompt asking attendees to map where in their own workflows they would apply each of the three stages. Close with a Q&A anchored to practical barriers: data access, validation thresholds, and internal stakeholder communication.

A language model is fluent by construction — it sounds confident whether or not the knowledge behind it is current. In pharma, that's a real risk: marketing mix models, omnichannel targeting rules, and safety monitoring all draw from a shared knowledge layer, but each changes at a completely different speed. When governance doesn't account for that, agents produce fluent, convincing, wrong answers that pass review and propagate downstream. This poster presents a practical governance framework — calibrated to both risk and rate of change — built from real deployment experience across commercial and medical agentic AI in pharma.

More information coming soon.

This session addresses the uncomfortable truth behind pharma's AI race: fluent AI outputs are often unreliable due to fragmented data hierarchies, conflicting definitions, and siloed identities. We argue that the missing ingredient in agentic AI is a "Company Brain"—a semantic context layer unifying data, ontologies, and rules. We will demonstrate how knowledge graphs and GraphRAG enable explainable, auditable reasoning, and outline a reference architecture for building this foundation. We will also cover governance, human-in-the-loop checkpoints, and a maturity framework to help data leaders determine if their foundation is truly ready for trustworthy agentic AI.

Traditional launch signals such as activity volumes and prescribing trends, often provide visibility only after significant lag. However, the earliest indicators of launch success emerge across customer engagement quality, field execution, market access, patient support, and customer feedback, and are typically fragmented across multiple systems and functions.

This presentation introduces an AI-led Launch Excellence Engine that creates a unified insight-to-action capability by connecting and interpreting these early launch signals. Leveraging specialized AI agents focused on Performance Intelligence, Engagement Quality, and Customer Feedback, orchestrated through a cross-functional intelligence layer, the solution transforms disparate data into actionable insights and prioritized recommendations.

The approach enables launch teams to identify emerging opportunities and barriers earlier, optimize launch tactics, prioritize interventions, and make faster, data-driven decisions to improve execution precision and accelerate launch success.

A gross-to-net trend for the access team, a territory review for commercial operations, a contract term buried in a deck from two quarters ago - the answers already exist inside documents the organization owns. Yet the workflow is unchanged: someone files a request, it enters a data engineering queue, and the decision window closes before the answer arrives. This session asks whether that queue can be retired, and what engineering discipline is required before it safely can be.

Most teams attack this as either a retrieval problem or a generation problem. Neither framing survives contact with regulated commercial analytics. A deterministic pipeline preserves structure and provenance but cannot reason. An unconstrained language model reasons fluently but produces claims no compliance-conscious team can act on. Our contribution is not a new retrieval primitive; it is a deliberate architectural separation that assigns each concern to the layer that can actually guarantee it.

We present a ten-stage architecture in two halves. The first five stages are a deterministic data engineering substrate: batch ingestion with checksum-based deduplication, Markdown-first parsing across common office formats, near-duplicate and boilerplate removal, heading-aware chunking that preserves citation boundaries, and metadata tagging down to page, slide, and sheet. The remaining five stages are an agentic retrieval-and-generation layer: hybrid semantic and keyword retrieval over a single index, relevance reranking, citation-aware grounded prompting, per-claim source attribution with an explicit no-answer path, and end-to-end observability of every query. The design principle throughout: the language model is used only where reasoning is required, and never as the system of record.

We evaluated against a blind suite of 76 business questions curated by domain experts over proprietary commercial documents. The deterministic substrate alone, with no model reasoning, answered 12 of 76. The full grounded agentic pipeline answered 75 of 76, each answer citing an exact source location. The lift came from the separation itself - a provenance-preserving substrate is what makes the reasoning layer trustworthy enough to deploy. We will walk through the single unanswered question and what it reveals about the boundary of the approach.

We also treat the uncomfortable parts directly. An explicit no-answer path is a product decision, not a technical default, and teams must be willing to accept "not found" over a plausible guess. Chunking choices silently determine citation quality. Answer accuracy is bounded by document coverage, so gaps in what gets ingested become invisible gaps in what can be known. And self-serve access to document insight raises access-control questions that must be resolved before rollout, not after.

The approach runs on the unstructured commercial and operational documents analytics teams already hold, generalizes across formats without document-specific workflows, and remains model-agnostic as underlying models change. Attendees leave with a stage-by-stage architectural blueprint, an evaluation design they can replicate on their own document corpus, and a candid view of what shortens the analyst queue versus what merely moves it.

Rare diseases present unique challenges for patient identification and provider engagement. Low disease prevalence, delayed diagnosis, fragmented care pathways, limited data coverage, and historically inconsistent disease coding often result in significant portions of the patient population remaining unidentified within traditional commercial datasets. For rare renal diseases such as IgA Nephropathy (IgAN) and C3 Glomerulopathy (C3G), reliance on a single data source or diagnosis based targeting approach can substantially underestimate patient opportunity and limit effective healthcare provider engagement. To address these challenges, Novartis developed an integrated patient finding and HCP activation framework that combines connected data assets, AI based predictive analytics, and field execution strategies.

Pharmaceutical landscape assessments require analysts to reconcile evidence across claims, literature, trial registries, and syndicated sources. Differences in indication names, disease subtypes, and coding can change the meaning of an analysis before synthesis even begins.

This poster presents an architecture for landscape intelligence that combines a knowledge graph and semantic mapping layer with agent-assisted workflows for landscape synthesis, literature review, and patient analytics. It explains how harmonized context can connect evidence on disease burden, treatment patterns, pipeline activity, and pricing and access to downstream work such as target product profile development and forecasting.

An illustrative example will follow an indication through source mapping, evidence extraction, and synthesis. The poster will examine where deterministic rules are useful, what humans need to review, and how source references and validation checks can help identify errors before they affect subsequent outputs. Attendees will leave with a practical framework for evaluating the data foundation, review responsibilities, and evidence needed to assess an agent-assisted landscape intelligence capability.

This session explores how life sciences organizations can bridge the “context gap” that prevents AI systems from reliably interpreting enterprise data. It introduces a framework that combines semantic layers, knowledge graphs, and GraphRAG (Graph Retrieval-Augmented Generation) to transform traditional metadata and data catalogs into an executable business knowledge layer that provides AI agents with governed business definitions, entity relationships, metric logic, data lineage, and contextual business rules.

Through real-world commercial analytics examples, the session demonstrates how context-aware AI can improve the accuracy, traceability, and trustworthiness of insights by leveraging certified metrics, domain ontologies, and graph-based relationships across healthcare professionals, territories, payers, and products. Attendees will learn a practical implementation roadmap for building AI-ready analytics platforms that support trusted decision-making in life sciences organizations.

This poster presents the architecture of a multi-agent authoring assistant, live in production for multiple countries' HTA submissions, that acts as an intelligent co-author for HEOR teams. The poster walks through the seven-component pipeline — template creation from a country's regulatory blueprint, evidence ingestion with page-level provenance, section-by-section planning, evidence-grounded drafting, claim-to-source traceability, automated QC (hallucination detection, numerical fact-checking, citation completeness), and a co-pilot editor for natural-language refinement. The design premise is human-in-the-loop throughout: the system prepares a structured, fully cited draft and evidence trail; the author supervises, refines, and applies scientific judgement. Visitors will see how first drafts compress from months to hours, how authors gain leverage across parallel submissions, and how compressed authoring timelines ultimately help medicines reach patients sooner.

From Periodic Readouts to Always-On Intelligence, an Agentic AI Approach to Market Mix Modeling.

Pharmaceutical commercial organizations run on fragmented systems: CRM knows interactions, IC platforms know payouts, forecasting tools know projections, and BI dashboards know past performance. But no system holds a persistent, shared understanding of who a territory, HCP, or rep is, what has happened to them over time, and where they are headed. The result is organizational amnesia: a rep who missed goal by 10% looks identical in a dashboard whether they are four months into a tough new territory or a tenured performer genuinely trending down. This session introduces a simple, three-part identity framework designed to close that gap.

From Data to Decisions - Agentic AI Framework for Grading Complicated Medical Coverage Policies and Accelerating Uptake of Buy & Bill brands

Commercial teams hold detailed intelligence on every prescriber and every representative, yet at the one moment it matters most—just before a live HCP call—that intelligence is rarely in the rep's hands. Coaching remains episodic, dependent on a manager finding time for a ride-along and is often delivered weeks after the interaction it was meant to improve. This presentation introduces an agentic field coaching capability designed to close that gap, shifting coaching from a periodic event into a continuous, data-grounded discipline embedded in the rep's daily rhythm.

Field coaching has long relied on role-play with a manager or peer standing in for the physician — useful, but limited by availability, and rarely tailored to the specific HCP the rep is about to face. At the center of this capability is a configurable HCP avatar: a voice-driven digital persona that lets a rep rehearse a real conversation before ever entering the room. Rather than reciting a generic script, the avatar is grounded in commercial data already resident in enterprise systems — prescribing trends and patient volume, product share and TRx growth, targeting tier, formulary position, and prior call notes — and reasons in character as the physician it represents, drawing on that context the way an experienced rep would going into the visit. The avatar can be configured to a named HCP for high-priority accounts, or to a composite profile representing a broader segment or archetype, making the tool useful for both individualized coaching and scaled onboarding. What distinguishes this from a static role-play script is that the avatar behaves dynamically: it escalates skepticism when the rep's messaging is weak or generic, and softens or opens up when the rep addresses its specific clinical concerns or objections — mirroring how a real HCP interaction would unfold rather than following a fixed conversational path. The result is a rehearsal environment that sharpens message delivery, objection handling, and call confidence ahead of the actual engagement, with the potential to shorten ramp time for new reps and reinforce consistency in how brand messaging is delivered across the field.

This capability extends the coaching workflows commercial teams already run in their CRM rather than replacing them, adding what a system of record cannot natively provide. CRM-based coaching remains a valuable field force effectiveness tool for documenting manager observations, but it is inherently retrospective, dependent on ride-along availability, and scored against generic selling competencies rather than the rep's own territory performance. The agentic coaching layer closes those gaps: rehearsing call readiness before the HCP interaction rather than critiquing it after, grounding feedback in the rep's actual TRx trends, share attainment, and targeting priorities, and enforcing MLR-aligned messaging live during practice.

Because the rehearsal is grounded in the representative's own commercial standing — their actual accounts, quota position, and targeting priorities — feedback moves beyond delivery mechanics to address whether the right HCP was called, with the right message, against the right quota. Most role-play coaching today can only assess how something was said; this shifts the evaluation to whether it should have been said at all, to that physician, at that point in the rep's territory plan. A compliance layer operates beneath every exchange in real time, pausing the rehearsal the instant a rep drifts into an off-label claim, competitor disparagement, or minimized safety language. Rather than flagging violations after the fact in a post-call audit, the system intervenes in the moment of practice — so repetition reinforces approved messaging instead of quietly eroding it through habit. The same architecture is designed to scale beyond the individual rep. Coaching output can roll up to the district manager as a reviewable, ready-to-send note rather than one the manager has to compile from scratch, extending a manager's coaching capacity without adding headcount or time. Over time, the capability is built to benchmark each rep's patterns against top-quartile performers, surfacing where a rep's approach diverges from what the best performers in the same territory or segment consistently do. In doing so, the tacit know-how that typically lives only in a top rep's head — how they open a call, when they push and when they hold back, how they handle a specific objection — becomes a shared, teachable asset across the territory rather than an individual advantage that leaves when the rep does. Success is measured against tighter new-rep ramp time and a decisive shift toward more frequent, signal-based coaching. Attendees will leave with a clear view of the design choices, data dependencies, and open questions involved in operationalizing an agentic coaching capability for a live commercial team—including where the approach delivers value today and where its boundaries still lie.

This session presents a governed context-engineering framework for reliable natural-language analytics in pharmaceutical commercial organizations. It explores how a semantic layer combining pharmaceutical ontology, business-rule reasoning, deterministic validation, and auditable decision lineage can extend existing text-to-SQL capabilities without replacing current data investments. The session also examines how non-technical domain experts can own and continuously steward semantic definitions, supported by a working reference implementation, enterprise vendor-evaluation requirements, production experience, and a defined approach for evaluating answer quality and stewardship efficiency.

Why can the same business question produce different answers across the same enterprise?

This poster will uncover how fragmented metric definitions, unresolved entities, inconsistent ontology mappings, and entitlement rules can quietly undermine the consistency, trust, and auditability of enterprise AI. It explores how a shared context layer, built on semantic governance and a pharma-native knowledge graph, can give conversational and agentic AI a common understanding of business definitions, entities, relationships, and access controls.

The poster will also show how context engineering can create a reusable foundation for more consistent, explainable, and governed AI responses across commercial use cases, while reducing reconciliation effort and incremental agent build effort.

Field teams are where life sciences strategy meets reality, yet the information they need is often distributed across CRM platforms, operational systems, analytics tools, and dashboards they rarely have time to access. AI assistants promise to close this gap by delivering insights in natural language at the moment decisions are made.

This session draws on real-world experience designing, evaluating, and operationalizing an AI-powered assistant for field-based commercial and access teams. The evidence base includes analysis of user interactions, evaluation results across multiple testing cycles, and the design decisions informed by those findings.

Analysis of user queries revealed that a significant portion focused on understanding the business environment: key accounts, stakeholder dynamics, and market developments. Another major category centered on planning and execution, including account prioritization, territory management, and progress against objectives. Remaining interactions involved performance monitoring and managing activities through to resolution. Collectively, these findings suggest that users engage AI assistants throughout their workflow, from situational awareness and planning to execution and follow-through.

One of the most important engineering lessons was that prompt optimization alone delivered limited gains. Sustained improvements came from strengthening the underlying context layer: establishing robust data models, improving metadata management, and creating clear mappings between user terminology and enterprise data assets. We also found that users routinely rely on implicit context, expecting assistants to understand concepts such as ownership, organizational scope, and reporting periods without explicitly stating them. Effective solutions therefore require operational context to be incorporated by design.

Achieving production readiness required more than generating accurate responses. Domain adaptation techniques were necessary to address specialized business terminology and use cases. Governance controls were implemented to ensure responses aligned with appropriate business roles and responsibilities. Observability capabilities enabled end-to-end traceability, while structured user testing surfaced failure modes that were not apparent in internal evaluations.

This presentation introduces an AI-powered Rebate Integrity Solution that leverages an Agentic AI architecture to proactively identify, investigate, and prioritize rebate integrity risks throughout the rebate lifecycle.

To illustrates how enterprise LLM reasoning architectures and semantic data modeling convert complex, multi-source pharmaceutical datasets into an intuitive AI conversational interface. By replacing static dashboards with an intelligent, self-learning assistant operating inside a secure cloud environment (Snowflake), field forces gain cognitive relief, reduce administrative preparation time, and make faster, data-driven decisions directly at the point of customer engagement. Tuning critical guard-rails for proper directions and flow followed by the LLM to provide accurate and complaint responses.

A governed Analytics Ready Data foundation allows Medical Affairs to move from fragmented reporting toward continuous, evidence-driven scientific insight generation, improving the quality, speed, and consistency of decision-making. This is what data-to-decisions demands in a scientific setting: agentic AI is only as trustworthy as the data layer beneath it, and the traceability that makes an insight auditable is what makes a medical decision defensible.

What does it take to evolve a traditional omnichannel analytics ecosystem into an agentic omnichannel decisioning engine?

This poster will uncover how a harmonized KPI taxonomy, unified measurement framework, causal experimentation, and agentic workflows can work together to move commercial teams from descriptive reporting to continuous, prescriptive decisioning. It will also explore how monitoring, diagnostic, recommendation, and orchestration agents can shorten the path from campaign signal to action, while human-in-the-loop governance helps maintain trust, control, and scalability across brands and business units.

Trusted data alone does not make a trustworthy agent. AgentOps manages context as a product — owned, versioned, tested, gated — so pharma agents apply policy at decision time and stay auditable.

In the pharmaceutical industry, commercial decisions rest on data that arrives continuously from external sources and degrades quietly between deliveries. For years, the most reliable safeguard against that degradation was informal: an experienced analyst looked at an output and recognized that something appeared wrong. As organizations move toward agentic AI, that safeguard disappears. An agent does not hesitate, has no sense of what a period should look like, and will act on a flawed figure and explain it fluently. Presently, most quality assurance still relies on rule-based validation, including null counts, range checks, and referential integrity. These checks only catch the failures someone anticipated well enough to write a rule for. The defects that damage commercial decisions are usually the ones no rule existed for.

This approach treats data quality management as an engineered, AI-enabled capability, addressing three failure modes that rule-based checks routinely miss:

• Detecting Incomplete Deliveries through Forecast-Based Expectation: A period lands partially delivered, is trended anyway, and reads as a business decline, sending teams to investigate a drop that never happened. Time-series forecasting of expected volume converts under-delivery and reporting lag into an explicit signal, available before any analysis runs on partial data rather than discovered weeks later.
• Catching Silent Drift through Baseline-Relative Quality Scoring: Quality erodes gradually. No single delivery breaches a fixed threshold, yet the asset is materially worse than it was a year ago. Scoring every governed dimension (completeness, uniqueness, validity, cross-field consistency, temporal ordering) against a stable baseline, as well as in absolute terms, allows degradation to be detected as a delta rather than a violation.
• Isolating Buried Defects through Multi-Level Anomaly Detection: Aggregates look normal while a small set of records carrying implausible values distorts a segment, a forecast, or a downstream payout. Unsupervised outlier detection and distributional comparison form a cascade that move from total volume, to segments such as geography, specialty, product, and payer, down to the individual records driving the deviation.

Equally important is how findings are delivered. The same quality core must serve an executive who needs a defensible ship-or-hold decision, an analyst who needs diagnostics granular enough for root cause, and a data operations team that needs evidence specific enough to take to a vendor.

This presentation will cover the design, development, and implementation of the entire capability, including real-world examples, the maturity path from reactive checks to monitored drift to predictive quality. It will also offer a candid view of what proves difficult in practice: an anomaly is not an error, and threshold design matters more than algorithm selection. In addition, it will examine the organizational constraint by clarifying who holds the authority to stop a delivery.

Most enterprise analytics assistants start the same way: document tables, columns, metrics, and use cases, embed that documentation, and retrieve the top-k most similar chunks. This works at small scale and fails quietly at real scale. Once a chatbot spans 30 or more tables and dozens of use cases, similarity-based retrieval starts omitting a required join path, a governing business rule, or a related metric - and the model answers confidently on incomplete context. This is not hallucination in the usual sense; it is silent incompleteness, harder to detect and more damaging in a regulated commercial setting.

This session shows how we replaced top-k similarity with ontology-driven graph traversal in a production analytics context layer. Retrieval becomes a deterministic connected walk: identify the relevant entry nodes and filters from the question, then traverse defined relationships to pull every connected node - each carrying its full stored context - rather than hoping the right chunks rank highly.

Node types include datasets, columns, metrics, use cases, concepts, business rules, personas, and ontology entities. Relationship types include dataset-to-column, dataset-to-dataset joins, column-to-metric derivation, metric-to-use-case, use-case-to-concept, and concept-to-business-rule and concept-to-persona links, plus entity links into unstructured content. We organize these into three ontology tiers that we recommend as a reusable pattern: a structured ontology covering datasets, columns, metrics, joins, and business rules; a business ontology standardizing use cases, use-case categories, therapeutic areas, disease areas, and brands into shared terminology; and a document ontology that decomposes unstructured sources into documents, sections, paragraphs, statements, and textual entities, linking them back to business concepts for GraphRAG.

Construction is more tractable than teams expect. We reuse the same context documentation already written for embeddings, add an explicit declaration of desired node and relationship types, and pass both through an agentic pipeline that generates a first-pass graph in ingestion-ready form. An SME reviews and corrects before ingestion - this human gate is non-negotiable. Ongoing expansion follows the same loop, passing the prior graph definition alongside new dataset and use-case documentation so growth is incremental rather than a rebuild.

We are explicit about fit. For a single use case, one functional domain, or a handful of simple tables, a knowledge graph is an overfit - traversal returns more context than the question needs, adding cost and latency where embeddings would suffice. Graphs earn their keep when table counts grow and join logic, metric derivation, and cross-domain concepts become genuinely complex. The main cost is upfront: modeling and curation take real time, and that investment compounds only when the structure is defined well.

In production, this approach enabled analysis across many datasets and deeper, more varied concepts, with observed retrieval accuracy above 95% and a scalable path for onboarding additional datasets. Attendees leave with a node-and-relationship starter schema, the three-tier ontology pattern, a build-and-review workflow, and a clear decision rule for when to choose graph over vector retrieval.