Introduction

Key Takeaways

·      18 Clinical success rates stagnated at 10% for twenty years

·      Psychiatry TPP compressed from six weeks to three days

·      Five-point success improvement means 50% better program odds

·      Organizations need months updating post-readout competitive scenarios

·      Generic AI tools cannot address unstructured evidence environments

The pharma industry faces a stubborn paradox: despite decades of innovation in clinical operations, trial success rates remain trapped at historical lows. Karl Moritz Hermann, CEO and co-founder of Reliant AI and former senior staff research scientist at Google DeepMind, opened his Clinical Innovation 2025 presentation with a stark statistic and an unconventional solution.

"For the past 20 years, clinical success rates have hovered around the 10% mark and anything we can do to change things upstream, to have better decisions upstream, should impact that number, " Hermann explained.

His framework positions upstream evidence automation as the highest-leverage intervention point in pharma development, with implications extending far beyond incremental productivity gains.

The Compounding Cost of Slow Evidence

Hermann began by challenging the audience with a live poll: How long does it take your organization to update competitive scenarios after a readout? The uncomfortable silence spoke volumes—most admitted to two- or three-month timelines, with TPP development stretching three to six months or longer. These aren't merely productivity problems; they're strategic vulnerabilities that cascade through the entire development lifecycle.

The issue stems from what Hermann calls the compounding effect of upstream decisions. "Everything you do upstream compounds, right? And when we talk about pre-trial decision making, we talk about formulating TPPs. It's about evaluating evidence, it's about deciding which assets to promote, " Hermann said.

Every choice made during TPP formulation— which endpoints to pursue, which populations to target, which assets to advance—sets the trajectory for hundreds of millions in downstream investment. Yet these critical decisions often rest on fragmented evidence: PDFs from clinical data providers, spreadsheets from past consultants, conference abstracts in inconsistent formats, and whatever the current advisory team has compiled in PowerPoint.

This evidence fragmentation creates organizational friction that extends far beyond slow cycle times. Hermann noted a pattern across pharma companies where inconsistent data leads to misalignment across functions.

"The problem of course is when you build a TPP, when you do all of this on inconsistent data, you will have two teams looking at the same data and drawing three different conclusions at the end of that, " Hermann observed.

The result: protracted internal negotiations, duplicated analyses, and delayed trial initiations—all while competitors move faster with more unified decision frameworks. The stakes extend to the highest levels of capital allocation.

"Portfolio planning is ultimately what decides where to allocate billions, depending on the scale of your company, which trials get started, who gets treated and ultimately how competitive can you be at launch, " Hermann said. When these decisions rest on incomplete or inconsistent evidence, organizations aren't just losing time—they're systematically underperforming on the precision of their strategic bets.

For an industry where 90% of clinical investments fail to reach market, even marginal improvements in upstream decision quality represent transformational opportunities.

Why Generic AI Cannot Solve Unstructured Evidence

Hermann was direct in his assessment of the current AI enthusiasm in pharma R&D: productivity tools like email drafters and slide generators deliver marginal value, but they don't address the structural problem of unstructured evidence. The challenge isn't that companies lack access to data—clinical data providers offer extensive databases, and consultants can produce beautifully formatted analyses. The problem is neither delivers decision-ready evidence at the speed and completeness that competitive strategy demands.

Generic large language models, Hermann argued, cannot bridge this gap without purpose-built infrastructure. "If your data today is PDFs, spreadsheets, some data from the FDA portal or the EMA one as well, if you're so inclined, conference abstracts and different formatting and so on, if that's your decision basis, then no, something-something GPT is going to fix that for you, " Hermann said. The transformation requires three distinct capabilities, each representing a technical breakthrough unavailable until recently.

“This is about ingesting data, it's about structuring data, and it's about activating data. And this is the AI slide, right? Because this wasn't possible before," Hermann explained. designs optimized for differentiation signals.

This three-stage framework distinguishes transformational AI applications from incremental automation. Traditional approaches force organizations to choose between completeness and timeliness— consultants can structure evidence comprehensively but deliver static point-in-time analyses, while data vendors provide current information in formats requiring extensive manual interpretation. Hermann's argument is that AI enables a previously impossible combination: comprehensive, structured, and continuously updated evidence that responds automatically to competitive developments.

The implications for competitive intelligence are particularly acute. "There's no reason why it should take you more than—well, why it should take you months to update scenarios just because a competitor reads out. If you have a model set up, if you have this worked out, where your data comes into a structured place, this should be able to update automatically, " Hermann said. Organizations shouldn't require months to incorporate competitor readouts into strategic scenarios—with structured evidence platforms, updates should be automatic, eliminating the strategic blind spots that emerge during critical decision windows.

From Six Weeks to Three Days

Hermann grounded his framework in a recent customer engagement: shaping an early-stage TPP for a new psychiatry program. The traditional approach would require six to eight weeks of manual work—identifying analogs in adjacent indications, cataloging endpoints from comparable trials, building evidence tables, and iterating across therapeutic area teams before reaching internal alignment. That timeline assumes no major competitive developments interrupt the process.

The structured evidence approach delivered dramatically different outcomes. "We were able to within three days, pull out 150 analog situations. From each of these situations, pull out endpoints to come out of this with a ranking of 10 to 15 endpoints that made sense for this particular situation, " Hermann said. The speed improvement alone—from weeks to days—would justify adoption for organizations managing multiple simultaneous TPP cycles.

But Hermann emphasized that precision gains matter as much as velocity. "Importantly, in this case, we got rid of two non-viable endpoints that we initially had on a shortlist, " Hermann noted. These represent avoided costs: resources that would have been invested in protocol development, regulatory strategy, and potentially trial execution for endpoints unlikely to deliver commercial value.

The case study also demonstrated how structured evidence accelerates organizational alignment. When all teams—therapeutic area leadership, portfolio planning, clinical operations—work from the same unified evidence base, iteration becomes dramatically more efficient. The friction of competing interpretations dissipates when everyone accesses consistent, complete evidence rather than fragmented sources requiring manual reconciliation.

Hermann concluded with a framework for action, urging organizations to conduct evidence automation audits using six diagnostic questions. The most critical: How long to update scenarios post-competitor readout? How long to reach decision-ready TPPs? Organizations exceeding one week on either metric should treat evidence structuring as a strategic priority rather than an incremental improvement opportunity.

The economic case rests on two levers. Speed improvements—compressing TPP cycles, accelerating trial initiations—preserve NPV and reduce the capital cost of prolonged development timelines. But the larger opportunity lies in precision: better-informed decisions about which assets to advance, which endpoints to pursue, which populations to target.

"If we can just lift this by 5 percentage points, that means your odds of coming through will actually go up by 50%. This is meaningful, "

Hermann said. Applied across a portfolio, this represents hundreds of millions in avoided costs from failed trials and accelerated paths to market for viable assets.

The organizational implications extend beyond technology adoption. Hermann noted that companies where central business functions— portfolio planning, R&D strategy—own evidence automation initiatives see faster adoption than IT-led efforts. The signal matters: treating structured evidence as decision infrastructure rather than data management positions it as accountable to commercial outcomes.

As pharma R&D faces mounting pressure to demonstrate capital efficiency, Hermann's message is clear: clinical innovation isn't only about executing trials faster. "Today, clinical innovation is not just about working faster in the clinic, it's also about thinking faster before you even go there, " Hermann said. It's about thinking faster before trials begin— structuring the evidence that determines which programs deserve those billion-dollar investments in the first place.

In order to get you the highlights of Pharma Clinical Innovation USA 2025 faster, we are using generative AI technology to summarise the transcripts of the sessions. The conference organiser is checking the summary for accuracy. If you have any feedback about the summary and the post-event report, please contact lucy.fisher@thomsonreuters.com

"70% of the material created by organization never reaches HCPs," Calliano noted. "And I think that as pharma industry we have an obligation wherever we invest money that this money is invested in a productive way."

AI's promise lies not in creating more content faster but in ensuring the right messages reach the right patients through optimized resource allocation.

The solution requires a fundamental mindset shift. Rather than innovation-focused experimentation, scaling demands enterprise-grade thinking around infrastructure investment, compliance frameworks, cost management, and workflow embedding. "AI for me compared to traditional ERP project is essentially a co-development exercise where you're working with your customers to develop something that adds value for them," Calliano explained. Organizations that develop repeatable scale-up methodologies will compound advantages over time, as each successful deployment builds platform capabilities reducing cost and risk of subsequent initiatives.

The discipline extends to use case selection. High-impact applications combine clear business problems, measurable outcomes, contained scope, and alignment with existing processes. Organizations spreading resources across dozens of disconnected pilots sacrifice the focused effort required to achieve production quality. The companies successfully scaling AI share common characteristics: executive sponsorship ensuring cross-functional coordination, dedicated technical teams building reusable components, and ruthless prioritization concentrating resources on initiatives with enterprise-wide applicability rather than localized optimization.

Proven Applications Delivering Measurable Impact

Beyond the pilots, some organizations have achieved tangible results by focusing on contained use cases with clear ROI metrics. ViiV Healthcare deployed AI-powered coaching assistants that analyze recorded physician calls, providing field teams immediate feedback on question quality, personalization, and conversation effectiveness. "We use recording during calls and we use that as a coaching assistant," Kate Francis described. "And it's real time because the individual doesn't have to have someone with them assessing how they do." The system eliminates weeks-long reporting delays and removes the need for physical accompaniment during calls, though obtaining physician consent for recording requires different approaches than traditional field rides.

The impact extends beyond efficiency. AI's capability to process vast information volumes is accelerating medical education and enabling more sophisticated shared decision-making between physicians and patients. "You can go through 200,000 scientific publications in a matter of hours, things that used to take months," Francis noted. This acceleration proves particularly valuable in rapidly evolving therapeutic areas where evidence bases expand faster than human review capacity.

GSK embedded machine learning tools into commercial investment workflows across five markets, enabling finance, commercial, and insights teams to make real-time tactical adjustments. The system functions as decision support rather than autonomous decision-making, maintaining human oversight while dramatically improving analysis speed. This human-in-the-loop design addresses both compliance requirements and user acceptance challenges, allowing teams to build confidence in AI recommendations through transparent logic and override capabilities when business judgment suggests alternative approaches.

The transformational potential is particularly evident in HIV and infectious disease, where patient advocates play active roles in treatment decisions and healthcare equity gaps persist.

"If what we do is improve the optimization of the message, improve the efficiency of how that communication happens, that it is tailored and personalized to the customer, not just the HCP, but the customer we're dealing with, then we're going to improve overall healthcare,"

Francis emphasized. AI applications extending beyond physician engagement to patient education, adherence support, and community connection could address systematic barriers preventing marginalized populations from accessing care.

Human Barriers Constrain Technical Possibilities

Despite AI's technical sophistication, the binding constraint on pharma ROI is not algorithms, data, or compliance but change management. Organizations underestimate the human barriers including works council concerns, job displacement anxiety, discomfort with changing physician relationships, preference for perfection over iterative solutions, and resistance to new working methods. "I think the biggest bottleneck obviously is compliance tools, data mindset," Francis acknowledged. "I think people's mindset, if they're close to it, they're worried about it getting through works councils, just getting people to embrace something that's completely different and new."

Even technically successful deployments fail without addressing training needs, demonstrating short-term wins, and phasing implementation to avoid overwhelming teams with simultaneous initiatives. The anxiety around AI displacement, while understandable, may be premature. "This idea that AI is going to take away a job, we're going to disappear, then a bot or an agent, which is for me still a glorified bot, can take over our jobs is not a scenario that's going to happen in the next two, three, four years," Calliano assessed.

The hard work required to make AI functional and accurate ensures human expertise remains essential. Organizations transparently communicating AI's role as augmentation rather than replacement, involving users in co-development, and demonstrating productivity gains rather than headcount reduction will accelerate adoption and capture value faster than those treating AI as cost reduction mechanism.

Sandoz identified medical inquiry management as a high-impact AI entry point. With nearly a billion patients treated globally, the company faces massive inquiry volumes from patients and healthcare professionals across markets like the US, Germany, and Japan. "Today there is a big budget," Pagano noted, describing the current manual process requiring medical affairs personnel to answer within strict timeframes. AI triage could perform initial question categorization and solution matching, routing inquiries to appropriate medical personnel rather than forcing 48-hour waits through multiple channels.

"There is still a human element because the final decision is taken by a human," Pagano emphasized. "But AI can help a lot."

The session concluded with a provocative challenge cutting to the heart of AI's paradox in pharma engagement. "If everyone is creating more content faster, but the HCP's time is capped, are we all at risk of spamming doctors?" Walton questioned. The answer may lie in using AI for restraint as much as generation, determining when not to engage rather than maximizing touchpoint frequency. Organizations that combine AI personalization capabilities with content discipline and attention economics will likely achieve better outcomes than those simply accelerating production.

The path forward requires balancing competing imperatives. Organizations must experiment to discover high-value applications while maintaining discipline to scale proven use cases. They must leverage AI's speed while respecting regulatory processes ensuring patient safety. They must pursue efficiency gains while investing in change management ensuring adoption. The companies successfully navigating these tensions will separate themselves not through superior algorithms but through superior organizational capabilities translating technical possibilities into business results.

To get you highlights of Pharma Customer Engagement EU 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 2026 (22-24 April, Barcelona) - the global collaborative network for leading pharma innovators. Visit the website here.