SPEAKERS:
Ashutosh Katiyar, Executive Director, Commercial Strategy, Insights & Analytics, Regeneron
Chris Walker, Head of Product Marketing, Tellius
KEY TAKEAWAYS:
· Analytics latency is a decision-quality failure, answers delivered after decisions are made have no value
· The semantic layer is simultaneously a technical architecture and a strategic repositioning of the analytics function
· Trust erosion from premature AI deployment is irreversible, not recoverable through iteration
· Regeneron's virtual account platform targets HCPs receiving fewer than one field visit per quarter
· Change management deserves co-equal investment with technical architecture, not treatment as a follow-on phase
Consumer AI has permanently reset the tolerance threshold for analytics latency. When every business user carries a tool in their pocket that returns sophisticated answers in seconds, waiting three days for a commercial insight is no longer an inconvenience. It is an organizational credibility problem. "There's a lot of expectation in the ChatGPT or Claude era that I'm going to get my answer immediately," said Chris Walker, Head of Product Marketing at Tellius, speaking at Reuters Events: Pharma USA. "And how do you work in that world?"
Ashutosh Katiyar, Executive Director of Commercial Strategy, Insights & Analytics at Regeneron, offered a precise answer to what that failure actually costs. "Sometimes when you answer the question, it is already too late," Katiyar argued. "People have moved on and they've already made a decision. So how do you really solve that decisioning cycle and sort of get from days to hours and hours to minutes?" When latency crosses that threshold, the analytics function gets bypassed entirely, with business leaders filling the vacuum using whatever information is already in front of them.
AI agents represent the obvious structural solution: democratize data access, put the analytics capability directly in front of business stakeholders, and eliminate the bottleneck. Regeneron's deployment tells a more complicated story. Organizations that move to AI agents before building the architecture of semantic coherence and user trust don't get slow adoption. They trigger the kind of permanent rejection that no subsequent technical improvement can reverse. The urgency that justifies the investment is precisely the pressure that makes it dangerous.
The Invisible Architecture That Determines Everything
The commercial analytics ecosystem at most large pharma organizations is not a shortage of data. It is a surplus of disconnected data with no common interpretive framework. Katiyar described the operational reality his team was working against: a fragmented environment of reports and analytics sources spread across the commercial ecosystem, each telling a different version of the same story, with no unified space where a business user could pose a question and receive a coherent answer. Every ad hoc request required manual assembly, cross-referencing sources that often contradicted each other, producing answers that arrived late and sometimes conflicted with the next report the stakeholder happened to open.
The solution Katiyar built around was explicit: "a unified framework, creating the semantic layer that understands business questions and connects it back to the data." The second half of that statement carries as much strategic weight as the first. "If you democratize the data and put it in front of your business leaders and business stakeholders, you're not going to get bothered by all these small requests that are coming in so that you can spend your quality time thinking about growth opportunities."
Most organizations frame the business case for a semantic layer as an efficiency argument: faster answers, fewer tickets in the queue. Katiyar is describing something structurally different, a repositioning of the analytics function within the organization. The ad hoc request backlog is not just an operational drain on analyst hours. It is a structural mechanism that keeps analytics leaders perpetually reactive, consuming the bandwidth that would otherwise allow them to participate in strategic conversations upstream of decisions rather than downstream of them. The semantic layer's primary value is not the queries it absorbs. It is the strategic altitude it creates for the team behind it. This is the category of foundation Tellius builds for pharma commercial teams: a governed semantic layer with pre-built definitions for TRx, NBRx, payer hierarchies, and territory structures, sitting beneath a conversational interface so the same question returns the same answer regardless of how it is phrased.
Walker's framing placed this architecture decision in the context of what AI is actually capable of when the foundation is sound. "AI is more than assistive. It can be transformational. It can change the way that your organization completely operates." That transformation requires the semantic layer as its precondition. Without a unified interpretive framework beneath the AI agent, what users experience is an interface that occasionally produces brilliant answers and occasionally produces confidently wrong ones, with no visible way to tell the difference. That inconsistency is not a nuisance. It is the mechanism through which user trust collapses.
The organizations most likely to underinvest in semantic architecture are the ones under the most pressure to show AI progress quickly. The semantic layer is invisible to end users, produces no immediately visible output, and is difficult to demo to executive sponsors. Its absence only registers when the AI agent starts returning answers that don't cohere, and by then, the damage to user trust may already be done.
Why the Trust Cliff Has No Guardrail
Katiyar was unambiguous about the hierarchy of priorities in AI analytics deployment. "Trust is number one." His specific concern was behavioral: GenAI systems can return different answers to different phrasings of the same underlying question. In consumer applications, that variability is mildly frustrating. In commercial pharma analytics, where field teams are making call-planning decisions and targeting choices based on AI-surfaced insights, inconsistency is operationally dangerous. A representative who receives one strategic recommendation on Monday and a contradictory one on Thursday doesn't slow down. They stop trusting the system and start relying on their own judgment, which was the baseline the AI was supposed to improve upon.
The consequence Katiyar described is not a gradual fade in engagement. "If you throw too much at it and you're not able to solve for that trust issue, people will lose trust. People will just stop using the system altogether." Katiyar is not describing a recoverable dip in adoption metrics. He is describing a cliff. The system becomes an expensive platform that nobody logs into, and no subsequent technical improvement changes that outcome, because organizational memory of unreliable tools persists long after the reliability is restored.
This irreversibility is what makes premature deployment a genuinely asymmetric risk. If trust erosion were recoverable, if organizations could deploy broadly, iterate on failures, and rebuild confidence through demonstrated improvement, then a speed-first strategy would be defensible. The evidence from Regeneron's experience suggests the recovery path is far harder than the deployment path. Organizations that spend two years rebuilding credibility after a failed rollout have effectively lost those two years twice: once to the failed deployment and once to the recovery effort, all while the analytics backlog they were trying to solve continues to accumulate.
The organizational requirement Katiyar identified as equally critical was change management, and he positioned it as a structural co-investment, not a Phase 2 consideration. "Invest in change management. If you're not changing the way that you're operating and you're not training your stakeholders to ask the questions to that platform, then you are doing a disservice to yourself because they'll keep coming back to you." If stakeholders are not trained to interact with the system effectively, to understand what kinds of questions it handles reliably and what its outputs mean, they will route requests back to the analytics team regardless of the AI's capability, recreating exactly the bottleneck the deployment was designed to eliminate.
Walker's distinction between assistive and transformational AI maps precisely onto this dynamic. An AI tool that business users interact with the same way they interact with a search engine is assistive. Transformation requires behavioral change at the organizational level: new workflows, new expectations about who owns which analytical decisions, and new discipline about what the platform is asked to do. That change does not happen through a training session on the platform's features. It happens through sustained engagement with how the organization makes decisions, which is a change management problem as much as a technology problem.
The organizations with the worst analytics backlogs are the ones under the most pressure to deploy immediately. They are also the ones with the least bandwidth for the semantic architecture work and change management investment that make deployment survivable. The constraint and the remedy compete for the same resource.
Discover more on this topic at Pharma Customer Engagement USA 2026 (October 27-28, Philadelphia) - where commercial, marketing, medical, data and AI pioneers converge. Explore the agenda here.
Low-Volume HCPs and the Commercial Proof Point
Regeneron's virtual account platform translates the architectural principles into a specific commercial problem: how to engage meaningfully with physicians who don't warrant the economics of regular field visits. "Some of these customers are not even getting a field rep once in three months," Katiyar observed. These are not unimportant prescribers. They are low-volume HCPs who collectively represent meaningful growth potential but fall below the threshold that justifies sustained field deployment under traditional resourcing models. The platform was designed to close that gap, using AI-driven targeting to identify the right accounts and enable outreach at a frequency and specificity that field resources alone cannot maintain.
The platform's core differentiator is pre-call predictive intelligence. "Before you go into that account or a given physician, you need to know everything there is to know about that physician or account," Katiyar explained. The Account 360 view surfaces not just prescribing history but physician mindset, adoption-journey stage, and anticipated objections. This is the contextual preparation that previously required hours of manual research before a meaningful call, now extended to the long tail of commercially relevant but resource-constrained relationships rather than reserved for high-priority accounts alone.
Walker’s read on why this approach lands is worth sitting with. “This is a white space problem,” he said. “There was no previous engagement model to replace—the platform created entirely new commercial value.” Most AI analytics business cases get framed around efficiency gains on existing workflows. The commercially interesting cases are the ones reaching accounts the old model couldn’t economically serve at all, and that is where the architecture investment starts producing results that wouldn’t have existed under any previous operating model.
The results Katiyar reported were directional: "a stark improvement in the customer engagement" with low-volume targets, with engagement improvement tracking alongside script lift, volume lift, and brand lift. No specific figures were disclosed, but the causal logic is clear. AI-enabled targeting produces engagement frequency, which produces commercial volume. The architecture built to make the platform trustworthy is the same architecture that makes those results traceable back to the investment.
The Paradox Demands a Sequencing Decision
The argument Katiyar and Walker assembled at Reuters Events: Pharma USA points toward a conclusion that most AI analytics business cases are structured to avoid: the foundational investments that make deployment survivable are most valuable precisely when they are hardest to justify. Executive sponsors under pressure to demonstrate AI progress want visible outputs, agents, interfaces, dashboards. They are rarely enthusiastic about funding a semantic layer that produces no user-facing artifact.
The Regeneron experience suggests the sequence is non-negotiable. Deploy the agent without the semantic foundation and you get inconsistent answers. Get inconsistent answers and you breach user trust. Breach user trust in an enterprise AI context and you face an abandonment problem that persists through every subsequent technical improvement. The organizations that skip the foundation to show speed don't arrive at transformation faster. They arrive at a trust deficit that forecloses transformation entirely.
Change management and semantic architecture are not risk-mitigation expenses. They are the mechanism through which AI investment converts to organizational capability. The analytics leaders who will extract durable value from AI agents are the ones who treat the invisible foundation as the primary deliverable and the agent itself as the downstream payoff. That requires a different conversation with executive sponsors, one that asks them to fund the architecture before they can see the output. It is a harder sell. It is also the only sell that ends with the platform still in use two years later, and with the analytics function finally operating at the strategic altitude the investment was supposed to create.
To get you highlights of Pharma USA 2026 faster, we are using generative AI technology to summarise the transcripts of the sessions. If you have any feedback about the summary, please contact lucy.fisher@thomsonreuters.com.
Discover more on this topic at Pharma Customer Engagement USA 2026 (October 27-28, Philadelphia) - where commercial, marketing, medical, data and AI pioneers converge. Explore the agenda here.
