State of AI 2026 in Market Access and Pricing

13 talks at EPA. One board: the complete picture of AI in Market Access & Pricing. The use cases are all over the place, the success factors are not. Zoom in to see what everyone’s doing; zoom out for the patterns:

The human-in-the-loop consensus. Every single speaker (pharma, biotech, vendor, academic) mentioned keeping humans at critical checkpoints. Not as a caveat or a compliance hedge, but as a design principle: to go beyond the jagged frontier of AI.
Speed yes, quality quietly. Most impact claims were about speed. Faster dossiers, faster SLRs, faster submissions. But two speakers mentioned that quality is also improving. Will this raise the HTA bar?
‘Buy’ gets you started, ‘build’ gets you ahead. Most companies started with off-the-shelf tools. Many are now building something proprietary: custom RAG pipelines, internal agents, domain-specific models trained on their own data and expertise. Tech-enabled services are coming, slowly but surely.
Efficiency now, strategy later? Almost every use case on the board is about doing existing work faster: localising dossiers, running SLRs, answering tender questionnaires. The truly strategic applications are mentioned, but mostly in the future tense (e.g. finding high-value patient segments, forecasting PICOs early).

The Miro board below is read-only and free to explore. If you were at EPA and something is missing or wrong, I would be glad to hear from you.

The fuller picture

Back from the World EPA Congress 2026 in Amsterdam, and the contrast with last year is striking. In 2025, I wrote that we were probably approaching “the first AI read-out” for market access and pricing, and predicted a mixed picture.

That prediction was roughly right, but the overall mood in Amsterdam was less hedged. The question has shifted. In 2025, the conversation focused on whether AI would prove transformative or merely tangential. In 2026, that debate has largely been settled. The room was not arguing about whether AI delivers; it was arguing about how.

How do you design a workflow that captures the efficiency gains? How do you move from a validated pilot to something embedded in daily operations?

The talks in Track 8 — AI & Digital Transformation — and several keynotes reflected this maturation: fewer concept slides, more architecture diagrams, more numbers, and a striking number of references to things already in production.

What follows are my notes. Let’s compare — and please do let me know if I got you wrong or missed a point.

Practical use cases of GenAI for HTA

Sven Klijn — Director, HEOR Evidence Acceleration & Innovation, Bristol Myers Squibb

The burden of evidence generation in HTA is the primary driver for AI adoption, not the technology itself. JCA is adding to an already substantial workload, and the question is whether AI can help absorb that pressure without sacrificing scientific rigour.
Interacting with non-text data (e.g. HEOR Excel models) requires deliberate architectural design: structured data must be distilled into formats the LLM can process, with guardrails defining how the AI reads and modifies it.
Generic AI applications (chatbots, Deep Research, Copilot) are inadequate for HTA work — lack of traceability and no access to proprietary unpublished data. Rather than relying on the built-in reasoning of models, the more controllable approach is to decompose the workflow explicitly: define step A, then B, then C.
A rigorous human-in-the-loop QC process can actually exceed the quality of a fully manual process. BMS reports catching more errors under their AI-assisted workflow than before, while compressing timelines from months to days with smaller teams.

From tools to orchestration: agentic systems in P&MA

Daniel Moreira — Co-CEO & Co-Founder, Cellbyte

Productivity from AI should be measured across three dimensions simultaneously: time to outcome, quality of outcome, and time to validate; rather than speed alone. Most current tools optimise for only one or two of these.
Generic AI tools and siloed point solutions represent opposite failure modes: Copilot-style approaches are fast but sacrifice quality and traceability; manual workflows with individual AI tools preserve quality but remain slow. Neither achieves a real efficiency breakthrough.
The solution lies in a structured data layer integrating all relevant sources (regulatory databases, HTA documents, guidelines) directly with the agentic system, enabling automated steps, fast iteration, and linked original sources for validation.
Context specificity matters considerably more than it might appear. For G-BA submissions, for instance, terminological precision (AND vs. OR connectors, specific assessment terminology) can determine black-and-white differences in outcomes; a generic LLM will miss this.
To build a business case for automation, prioritise tasks by three criteria: frequency (how often is it performed?), consistency (how structured is it each time?), and clarity (how well-defined are inputs and outputs?).

The future of AI in market access

Marco Rauland — Vice President, Market Access, Strategy, Pricing and Analytics, Merck KGaA, Darmstadt, Germany

Localisation of global master documents into local dossiers works well with AI — particularly for structured tasks and major languages (English, Spanish). The approach breaks down significantly for smaller markets: less training data means worse outputs, precisely where resource constraints make AI support most needed.
AI-driven tender price prediction works in markets with published, high-volume tender data (e.g. Italy), and Merck reports the model outperforms human prediction in two out of three cases. However, if the approach becomes industry-wide, competitive dynamics may erode the advantage.
Value-based prediction in multi-factorial HTA decisions (e.g. AMNOG) remains unsolved. The correlation between clinical evidence and reimbursed price after G-BA assessment has historically been low, and AI has not yet cracked the underlying decision logic.
AI-powered payer negotiation simulation, including avatar-based mock negotiations with adjustable payer personas, is already in use at Merck for training. The tool records sessions, provides structured feedback, and allows repeated practice without the cost of external facilitators.
The long-run scenario Rauland raises: both sides of a negotiation may eventually deploy AI agents, potentially leading to machine-to-machine negotiation. Whether this improves or merely replicates current outcomes is an open question.

AI-driven commercial transformation

Fernando Ventureira — CEO, Stratence Partners

AI fails to deliver commercial value when treated as an IT project without cross-functional adoption, economic accountability, or a realistic integration path. The graveyard of unsuccessful pilots is largely populated by initiatives that lacked governance architecture rather than technical capability.
The data foundation must come first: a single point of truth integrating ERP, CRM, market data, tenders, access agreements, discounts, and rebates by SKU, customer, and region. Without this, AI generates unreliable or misleading commercial intelligence.
Sustained impact from AI in commercial settings requires embedding it into decision governance (ownership by decision type, escalation matrices, audit trails), operating cadence (weekly pricing forums, tender reviews), and economic accountability frameworks with tracked metrics.
AI adds measurable value in four specific decision domains: strategic (launch sequencing, portfolio simulation), access (reimbursement pathway, payer impact), pricing (corridor optimisation, net price forecasting), and execution (account prioritisation, promotion ROI). Outside these domains, its contribution is limited.

GenAI for HTA — panel discussion

Bill Malcolm (moderator) · Bristol Myers Squibb · Siguroli Teitsson, Bristol Myers Squibb · Dr. Rajdeep Kaur, Pharmacoevidence · Oliver van Zon, Crinetics Pharmaceuticals

The field has shifted in one year from abstract proof-of-concepts to board-level mandates for adoption, with validated use cases now reaching HTA review tables and scientific publications documenting results. The bottleneck is no longer the quality of the underlying LLM but the design of the solution around it.
RAG-based architectures address the core failure modes of generic LLMs in HTA contexts: hallucinated citations, lack of proprietary data access, output inconsistency, and data security concerns. The architecture matters as much as the model.
The competitive advantage in AI-enabled HEOR does not lie in access to better models — it lies in deep domain expertise. The most important function of subject-matter experts is workflow decomposition: breaking tasks into subtasks and assigning the right role.
Smaller biotechs face a structural disadvantage: less historical data, fewer prior submissions, and limited internal AI capacity, combined with vendor pricing that often remains calibrated for large pharma budgets. Avatar-based negotiation practice tools may partially compensate, allowing smaller country teams to conduct mock negotiations that were previously impractical.

Current use cases and prospects for GenAI in market access

Jagpreet (Jag) Chhatwal — Director, Institute for Technology Assessment, Massachusetts General Hospital, Harvard Medical School

Domain-specific AI platforms substantially outperform general-purpose LLMs on HTA-relevant tasks. In a direct comparison on rare-disease health-economic model conceptualisation, a HEOR-specific platform produced more HTA-aligned outputs, better citation quality, and adherence to CHEERS and modelling guidelines — tasks where ChatGPT and Gemini have gaps.
A full landscape assessment report (150 pages, 300–500 verified references, covering epidemiology through competitive pipeline) can be produced in 48 hours using agentic platforms, versus a conventional timeline of 3–5 months.
AI-based validation of Excel health-economic models against NICE’s internal checklist demonstrated that the system could replicate the full checklist; in one failure case, it correctly identified that the error was introduced by a human reviewer.
The right operating model for AI deployment (fully in-house vs. hybrid vs. fully outsourced) depends on internal capacity and domain expertise — and should be determined before deployment, not after.

AI-enabled automation of evidence synthesis

Shubham Pandey — Head of Modeling and Analytics, Pharmacoevidence

Pharmacoevidence is the first company to have an AI-assisted SLR accepted in a NICE UK HTA submission. The EAG reviewed the process and concluded the AI/ML tool was used appropriately — a meaningful regulatory precedent.
Siloed AI tools (one for SLR, another for ITC, another for medical writing) create fragmentation that loses context across the evidence pipeline. The value of integration is that outputs from each stage feed forward: SLR context informs ITC, which informs modelling, which informs the dossier.
Time savings from AI-assisted SLR are estimated at 50–80% across the lifecycle, compressing a process that runs 30–35 weeks manually to approximately 10–12 weeks.
Simulating HTA and FDA review questions before submission (using RAG-based multi-agent systems trained on past appraisals, guidelines, and sponsor submissions) predicted approximately 80–85% of actual regulatory questions in validation testing.

Jessica Weddle (Partner) & Anna Léandri (Manager) — Elevio Group

Four adoption archetypes are emerging in pharma market access AI transformation: the Big-Bang Transformer (comprehensive programme now), the ROI-Oriented (OPEX reduction focus), the Careful Experimenter (one validated use case at a time), and the Pragmatic Optimizer (wait and adopt once others have validated). None is inherently superior, but each carries distinct risks and timelines.
Three recurring blockers cut across all archetypes: incomplete or absent strategic vision (failing to define the end state before starting), delivery friction between market access teams and developers (different vocabularies, working rhythms, and levels of granularity), and change management failure (teams fearing replacement rather than augmentation).
The friction between market access professionals and developers is routinely underestimated as a time cost. A bridging role — someone who can translate between domain requirements and technical implementation — is not a nice-to-have but a prerequisite for delivery.
The AI journey in market access follows a recognisable trajectory: yesterday’s chatbots (“summarise this HTA guidance”) → today’s specialised agents (“support my GVD development end-to-end”) → tomorrow’s orchestrators (“run the full process of market access deliverables and identify decision-ready options”).

Beyond the buzzwords — how MA professionals are actually using AI

Rafaat Rahmani — CEO, Lifescience Dynamics Limited

A survey of 84 market access and analytics professionals (5–15 years’ experience, Europe and US) found that AI use is now mainstream in day-to-day workflows, with the majority using AI tools multiple times per week. The top applications are data processing, secondary research, and competitive intelligence. Forecasting and scenario planning remain underused, likely reflecting confidence rather than capability gaps.
Speed and synthesis are the dominant sources of perceived value: faster turnaround, clearer storylines, and the ability to operate as an editor rather than a generator. Accuracy and hallucination remain the dominant concerns, followed by compliance, cost, and trust.
Despite strong current adoption, 35% of respondents believe AI will not replace their jobs — though the majority expect it to become a standard component of analytics projects within three years, embedded into normal workflows rather than operating as a separate initiative.
The most revealing finding: professionals articulate what they actually want from AI as a thinking partner — something that sparks ideas, connects insights across domains, and provides strategic recommendations, rather than a task executor. This suggests the ceiling for value is considerably higher than current use reflects.

Integrating AI into an evolving structure

Aodan Tynan — Head of Access Implementation, Pricing & Value Demonstration, Astellas Pharma

Building a new health-economics team from scratch (mid-2025) offered a rare opportunity to hire for AI receptivity from day one — making openness, curiosity, and pragmatic scepticism explicit hiring criteria rather than retrospectively trying to instil them in an established team.
A team of six that has existed for only six months has already built internal tools: an agent that crawls HTA assessment databases and extracts economic model parameters, survival curve and utility analysis apps, and model QC workflows tested by deliberately inserting errors to verify detection capability. These were built without external vendors.
The most common failure mode in vendor engagement is discovering mid-pilot that a “solution” is still in development and the pharma team is effectively serving as an unpaid co-developer. Explicit clarification upfront (“is this ready for mainstream or are you still building it?”) is necessary but often skipped.
A useful internal boundary: use external AI tools for publicly available data and specialised vendor capabilities; use internal GenAI (Astellas has an internal model called Stella) for proprietary data — and keep that boundary clear for compliance and data security reasons.

From a tool to a teammate — AI agents in tendering

Nico Bacharidis — Chief Commercial Officer, Cube RM

A survey of 30 pharma companies found that most are not yet systematically applying AI to tendering, despite the tender process being, as Bacharidis puts it, built for applied technology — high volume, repetitive documentation, structured inputs, and published outcome data in many markets.
Cube RM’s Clara agent automates 60–70% of tender questionnaire responses by reading submission documents, identifying questions, matching to a document library, and providing confidence scores with source references. Human review is retained for low-confidence responses. The learning loop allows the system to incorporate feedback and improve over time.

Why should you negotiate with AI?

Chirag Maheshwari — CTO, PharmSight Research and Analytics

AI-based avatar negotiation practice tools address a concrete structural limitation of conventional mock-negotiation workshops: finite rounds, high logistical cost, and inability to personalise by country, payer type, or weakness profile. The avatar can run unlimited rounds, adjust aggression level, provide structured feedback, and replay specific sequences for improvement.
The emotional realism of avatar interactions (the tool is described as triggering genuine stress responses during sessions) is cited as a feature rather than an incidental quality: emotional engagement improves retention of feedback and makes the practice transfer more reliably to live negotiations.
The technical architecture layers a foundational LLM enriched with pharma-specific knowledge (first layer), internal data including global value dossiers and objection handlers (second layer), and country/indication-specific fine-tuning via RAG (third layer) — with guardrails preventing responses outside the defined scope.
No validated outcome data yet exists: the tool launched approximately six months ago and case studies are in progress. The honesty about this limitation, alongside the clear articulation of the value proposition, is noted: the proof of concept is plausible but unverified at scale.