Industrializing Intelligence: What 450 Pharma Leaders Heard at Axtria Ignite 2026

More than 450 executives from dozens of leading life sciences, MedTech, and biotech firms gathered in Princeton, New Jersey, on June 10 and 11 for Axtria Ignite 2026. They came to hear about the hits, the misses, and what it takes to lead the agentic enterprise in their own companies.

Axtria founder and CEO Jassi Chadha opened with an argument, one that the rest of the day worked through: the foundation, not the agent, determines whether AI in pharma scales.

The Question That Anchored the Day

By 2026, the questions about AI had stopped being whether it works. The harder questions had arrived, and they did not have easy answers.

In his keynote, Chadha pointed out that 60% of generative AI algorithms in medical applications still produce significant hallucinations. The pharma accuracy bar sits at 99.5%. 89% of AI pilots never reach production. And the industry is racing to deploy a projected 100,000 agents on top of data that was never designed for them.

Axtria Principal and Ignite’s emcee, Priya Bhargava, shared another question, one that framed everything that followed. A VP of commercial excellence had compressed a launch from 18 months to four and still admitted, “We have a Ferrari, we just don’t know how to drive it.” A VP, weeks from a rare disease launch, could not yet identify which physicians had patients who needed the treatment. Four leaders, four organizations, four different problems, and one underlying question: Before you build the agent, what must be true?

AI-Ready Data and a Reframe of the Race

Chadha answered with three things the foundation must carry: AI-ready data, a working semantic layer, and governance designed for digital workers.

On data: 73% of biopharma still reports significant data issues, and analysts lose roughly 100 days a year cleaning data for analyses as basic as HCP-level reporting. On the semantic layer: agents need business rules, glossaries, and domain context to deliver reliably, because the absence of that context is what turns a request for incentive compensation into an answer that is “just about right” rather than correct. On governance: trust in AI systems has fallen from 61% in 2019 to 53% now, and agents have to be treated like new hires who earn trust through validation and audit rather than receive it on day one.

Chadha’s closing reframe was the line of the morning: “The key in this race in AI is not being the fastest, not being the first,” he said. “If we have that as an objective, we’re going to have a lot more failures.”

Bhargava restated it with even more force: “The agents simply amplify what lies underneath them. Put them on a broken foundation, and they’ll go faster and faster toward the wrong outcome.”

For the attendees in the audience? The rest of the day was aimed at working through that argument.

Data Designed for Humans Cannot Run Agents at Machine Speed

The data problem in pharma has been around for thirty years. What changed in 2026 is the cost of ignoring it.

A panelist in the trusted-data session framed the gap directly: “Clean data, as we thought about it maybe five years ago, just isn’t enough to really power an autonomous agentic future.” Humans, the argument goes, are smart enough to fill gaps. They make predictions, work around quality issues, and reconcile inconsistencies on the fly. Agents cannot. They scale whatever is underneath them, including the gaps.

Another panelist in the same session offered the day’s most memorable framing of the work ahead.“AI is kind of a toddler,” the panelist said.“We used to ask him, ‘What is this color?’ It said ‘Red,’ and we were clapping. Now it’s time to send that kid to school. First, you create a syllabus. You say, ‘These are the books, you have to read them.’ That’s basically your context.”

The toddler-to-college reframe showed what made the data conversation harder this year than last. Panelists across forecasting, field intelligence, and rare disease patient journeys returned to the same point: the data that needs fixing is not just the quality of inputs but the institutional knowledge that lives in spreadsheets, dashboards, and the heads of senior analysts. A panelist in the same session called this the “tribal brain” problem: how do you convert organizational know-how that everyone agrees exists into something a machine can actually read?

A panelist in the forecasting session put the consequence in operating terms. This is where AI-driven analytics for pharmaceutical commercial strategy meets its hardest test. Forecasting is a culmination of cross-functional inputs (market access, finance, brand teams). If the data and the semantic foundation are not ready, AI cannot generate those inputs out of thin air. It hallucinates them. The forecaster ends up defending an AI-augmented number against a room of stakeholders who already had pre-conceived notions of what the number should be.

The pattern repeated across tracks. The data foundation used to be a prerequisite before starting AI work. Now it is the AI work itself.

The New Battleground is the Layer Above the Data

The session most singularly focused on this pillar drew the largest cluster of definitional questions of the day.

A panelist in the trusted-data session offered the clearest distinction between the two terms that everyone uses interchangeably. “Semantic is basically your metadata, business glossaries, business rules,” the panelist explained. “Context layer is when your structured data, unstructured data, all of that data come together using your semantic assets.”

The technical case for why this matters now is straightforward. LLMs are language models. They are not strong at processing integers and numeric information natively. Pharma’s analytic-ready datasets are mostly integers. The layer that translates between the two, the layer that tells the model what a column means and why it matters, is what allows an agent to give an answer the business can act on rather than one that needs to be checked against the underlying tables.

Another panelist in the same session put the strategic stakes in plainer language. The team that owns the context layer becomes, in effect, the team that owns the institutional logic. Some panelists framed that ownership as belonging to the brand team, which lives with that logic every day. Others framed it as a stewardship role within data and analytics. Neither framing settled it, but the question of who is the custodian of the context layer is becoming an organizational design question, not a tooling question.

A panelist in the high-performance customer engagement track tied the context layer back to field adoption. The unlock for reps is that they can ask the agent why it made a recommendation and get an answer that holds up. The recommendation itself is no longer the differentiator.

Governance is Becoming the Accelerator, Not the Brake

If data and agentic AI in life sciences were the longest-running themes, governance was the most-discussed across all of Ignite’s tracks.

A panelist in the agentic-AI mainstage session laid out one of the most concrete frameworks of the day: an AI augmentation scale from level zero (human does all the work) to level four (fully autonomous agents), applied role by role across the organization. The panelist’s company had run the exercise across six critical roles in clinical trial services, predicted where each role should sit on the scale in two and a half years, and then worked backward to redesign job descriptions, recruiting profiles, and required skills. “A general exercise won’t cut it,” the panelist said. “You have to do a role-by-role augmentation.”

A panelist in the same session drew an analogy that several attendees referenced for the rest of the day. The technology to put a self-driving car on the road existed in 2014. The reason self-driving cars are still not prevalent is not technical. It is governance, specifically systems-level governance, that lagged behind the capability for years. “Organizations that figure out this governance at scale,” the panelist said, “are the ones that are truly going to accelerate.”

A panelist in the dabbling-vs-demonstrating session called out the failure mode in current pharma practice. Most responsible AI boards are still operating as conventional toll gates. A use case shows up, the board says yes or no, and the work continues. What the industry needs, the panelist argued, is investment in agent observability, agent monitoring, and kill switches. Most governance organizations have not even upskilled on ISO or NIST frameworks. They are applying old technology governance to a new class of systems, which is what slows things down.

A panelist in the field-intelligence trust-stack session laid out a four-layer governance framework that an Axtria moderator returned to throughout the day: explainability that tells the basis for the recommendation, deterministic rather than probabilistic guardrails, decision lineage that creates an audit trail, and closed-loop feedback designed in from the start. Most pilots, the panelist argued, are killed because these layers are reactive afterthoughts.

The most quoted reframe of the headcount question came from a panelist in the pilot-purgatory session. “When the CFO office starts to do headcount tracking and says, ‘Let me have a headcount tracking between the carbon employee and the silicon employee,’ I think that is when we see the transformation.” Governance, in that frame, is what makes “silicon employees” auditable.

The Hard Part Nobody Talks About: People, Culture, and the 70 Percent

The most repeated message of the day cut across all three of Jassi’s pillars.

A panelist in the agentic-AI mainstage session put the number on it. “Change management and adoption cannot be the 45th slide in a 30-page deck at the business case approval. We have learned that 70% of the success depends on how you get that right.”

The afternoon fireside chat made the same case in operating terms. A panelist there described her company’s transformation in three phases: first build, then adopt, then impact. The first year and a half had been a build phase. The next phase was business accountability for adoption. Without business accountability, the products would sit unused regardless of how well they performed. The panelist’s company had recently rolled out licenses to roughly 30,000 employees and trained more than 1,000 marketers for generative AI in pharma for brand work, with reported adoption above 90% among that cohort. The cultural unlock, the panelist explained, was the framing: AI is not coming to take your parking spot. It is here to elevate what you do.

Another panelist in the same session offered the counterpoint that landed hardest. “It takes people to build, but it takes equal people to drive adoption,” she said. Her two stories made the point. A rare-disease patient-finding program worked, but only after a year of sharpening the leads with constant field input. A next best action program at the same company had to be killed because the field never trusted it. The difference was not technology. The difference was who was on the phone every week with reps, learning what counted as a junk lead and what counted as a real one.

A panelist during the dabbling-vs-demonstrating session reframed the entire change-management challenge in a single sentence. “This is not an AI revolution,” he said. “It’s a people revolution. We have to set our mindset that way.” His company had restructured at the executive level around it, with one leader responsible for AI across the business reporting to the CEO, supported by a 900-plus-person unit.

A panelist in the pilot-to-enterprise commercial transformation session described the change-management unlock as inclusivity in the design phase. Top-down decisions, in his experience, required a great deal of selling after the fact, because no one in the affected functions had been part of building the change. Bottom-up co-design front-loaded the work but eliminated the selling. People who had been part of the design did the marketing for him.

Another panelist in the same session offered the most honest admission of the day. After a successful CRM transformation, they had been “a little drunk with the success of the massive stand-up” and neglected the unglamorous work of documenting processes and standing up the QC mechanisms that would carry the new system. The lesson, they said, was not to celebrate too early on the front-facing wins while the fundamentals quietly slipped.

Dabble Faster: Moving Agentic AI from Pilots to Production

The closing mainstage session was an explicit call to action for organizations still tiptoeing.

Asked to characterize the risk of staying on the sidelines in five words or fewer, the panelists did not soften the answer. “Talent loss,” a panelist said.“Who wants to work for Blockbuster when you can go work for Amazon? People will walk with their feet.”

Another panelist offered the line that got the most attention from the audience. “Dabbling is a necessary but insufficient condition to succeed. So continue dabbling, but dabble faster.”

A third panelist made the case for shipping value rather than chasing perfection.“Don’t wait for things to be perfect. Just start. Because if you’re trying to work for something now, it will change by the time you get it perfect. It doesn’t matter if you don’t have the foundations, just start.”

That last line stood in apparent tension with the keynote’s three pillars. It was not. The point both speakers were making was that the act of building on a real foundation is itself how the foundation gets fortified. Organizations that build while reinforcing the foundation end up six months ahead. Organizations that wait for the foundation to be finished end up six years behind.

The pace of the AI vendor landscape, with new pharma-specific tooling arriving from multiple providers, has accelerated the urgency. So has the steady drumbeat of board-level AI fluency programs. The boards have moved on from asking whether to invest in biopharma digital transformation. . They are now asking, in the words of a panelist in the agentic-AI mainstage session, three questions: are we investing at the right level, how is the risk managed, and is execution tracking? Those are the questions that move organizations from dabbling to demonstrating.

Recognizing the Leaders Doing the Work

The day included the Ignite Leadership Award recognitions for three life sciences leaders: Anvita Karara of Bristol Myers Squibb, recognized for the enterprise-scale AI transformation rewiring how BMS engages healthcare professionals; Jeremy Pincus of GSK, recognized for an AI-powered overhaul of medical affairs that has cut MLR review cycle time by 30%; and Bharti Rai of Biogen, recognized for a career spent building data, operations, and customer engagement as one integrated function instead of three bolted-together ones. The Axtria Bedrock Honor was given to Quest Diagnostics in recognition of a fifteen-year partnership and the recent implementation of a sales force alignment solution across a 34-role-type, thirteen-hundred-person field organization.

What 450 Leaders Carry Home

Bhargava closed the day with the question one moderator had posed mid-session: What is the decision you make differently tomorrow based on what you heard today?

The day’s answer was less a list than a single observation: The foundation is the strategy. The data, the context, and the governance are the AI work. Treating them as afterthoughts is what produces the pilots that never reach production.

Senior leaders from Bristol Myers Squibb, Gilead, Bausch+Lomb, Biogen, GSK, Thermo Fisher Scientific, Quest Diagnostics, Pfizer, AstraZeneca, Merck, Novartis, Takeda, Sanofi, BD, Regeneron, Novo Nordisk, Amneal, Astellas, Boehringer Ingelheim, Madrigal, Sobi, Immunocore, Incyte, Ferring, Grifols, Cloverpop, and MSD Europe joined panels across the day. Axtria principals moderated each session