SPEAKERS:
- Indraneel Mukherjee, Life Sciences Partner, Guidehouse
- Anchal Joshi, Product Development & Commercialization Lead, Nephrology, Biogen
- Michael Palumbo, VP, Immunology, Eli Lilly
- Jay Sabbah, US Head of Marketing, AstraZeneca
KEY TAKEAWAYS:
- AI recommendation override rates increase as decision stakes rise, inverting AI’s intended value proposition
- Pharma leaders systematically misattribute weak AI ROI to model quality rather than organizational behavior
- AI commercial value is epistemologically invisible - the counterfactual cannot be observed or measured
- Field-level decisions offer the fastest path to building institutional AI trust through rapid feedback loops
- Making override a documented, justifiable act rather than a silent default breaks the self-reinforcing trust deficit
The question pharma keeps asking about AI is the wrong one. At Reuters Events: Pharma USA 2026, Indraneel Mukherjee opened the session with the provocation the room already felt but hadn't named: "We keep asking how to get more value from AI. The real question is why we keep not getting it - despite the investment, the pilots, the platforms." That reframe matters, because the answer isn't technological.
Michael Palumbo provided the immediate illustration. A forecasting model generated a resource allocation recommendation that conflicted with what the leadership team expected to see. "The model was telling us something we didn't want to hear," Palumbo said. "And so we didn't listen to it." The launch proceeded on the leadership team's instinct.
That is the pattern this panel surfaced. Not an isolated failure, but a structural feature of how pharma organizations lose value from AI. The highest-value applications - launch forecasting, territory reallocation, in-market pivots - are also the applications most reliably neutralized by the leaders who commissioned them.
The override happens where it hurts most
The instinct is to treat AI override as an adoption problem, something time, training, and better change management will erode. The evidence from this panel suggests otherwise.
Palumbo described the organizational conditions where AI-informed decisions survive versus where they evaporate. "When the decision is reversible, when we can course-correct in thirty days, people follow the model," he said. "When it's a launch forecast, when careers are attached to the number, the model becomes a reference point, not a decision." The pattern isn't resistance to AI in principle. It's a rational risk calculus: high-stakes decisions attract scrutiny, and no senior leader wants to explain to the board that a launch missed because an algorithm said so.
Jay Sabbah named the industry's misdiagnosis directly. "When AI doesn't deliver, the first thing people do is question the model - the data quality, the integration, the vendor," he said. "Nobody asks how many times the output was overridden before it reached a decision." That misattribution is consequential. When weak AI ROI is blamed on model inadequacy, companies respond by investing in better models. The organizational behavior that neutralized the existing models goes unexamined and unchanged.
Mukherjee provided the consulting-side corroboration. "Across clients, we see the same thing in launch planning - AI-generated forecasts that go into the process, get refined, get validated, and then get set aside in the final meeting when leadership decides to go with a different number because that’s more defensible," he said. The final meeting is always the tell. That's where the model's recommendation has survived every technical review and dies at the decision threshold.
The override paradox is not an arc that matures into trust over time. It is a structural feature of organizations making high-stakes decisions under career risk. Without deliberate intervention, deploying more sophisticated AI produces the same override behavior at higher cost.
AI cannot build the case for its own trust
The override paradox persists partly because AI has no mechanism to prove itself wrong when it gets ignored. This is the measurement trap, and it is more fundamental than most commercial leaders recognize.
Anchal Joshi framed the challenge from his launch planning experience in nephrology. "When a launch performs well, how much of that is the AI-informed allocation and how much is the team, the market dynamics, the competitive situation?" he asked. "You can't isolate it. The counterfactual doesn't exist." This isn't a data science limitation that better attribution modeling will eventually solve. It's epistemological. AI's commercial contribution lives in what would have happened without it, and that alternative timeline is permanently unobservable.
Sabbah pushed this further with a concept the panel hadn't previously named: the override tax. "Every time a recommendation gets overridden, there's a cost - a misallocation, a forecast error, a field decision that went the wrong way," he said. "We don't measure that cost. We don't add it up. So we never see the bill." The override tax is real and accumulating in every commercial organization running AI tools. It just isn't on any dashboard.
The specialty context makes the measurement gap more acute, not less. Joshi described the nephrology launch environment where patient populations are variable by indication, territory dynamics are idiosyncratic, and allocation errors are expensive to correct. "The AI's value proposition is strongest in exactly these situations - the data complexity is too high for human processing alone," he said. The tool is most needed where the stakes are highest, which is also where override is most likely and measurement is most difficult.
The cycle completes itself. AI can't demonstrate value because it gets overridden. It gets overridden because it hasn't demonstrated value.
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.
Field operations as the trust proving ground
Two structural moves emerged from the panel. Neither of them is algorithmic. The first is about where AI earns its track record. The second is about how insights survive organizational translation.
Palumbo pointed to real-time field-level decisions as the context where the trust-building calculus changes. "In the field, decisions are happening constantly, feedback comes back fast, and the individual consequence of any single call is bounded," he said. "That's where AI can prove itself - not in the annual forecast, but in the day-to-day." Field operations offer what launch forecasting cannot: high decision frequency, short feedback loops, and bounded downside. AI recommendations that prove accurate in this context accumulate the institutional track record that eventually creates permission for higher-stakes applications.
Sabbah reinforced the logic. "Real-time reallocation is where AI's speed advantage over human approval chains is most dramatic," he said. "A territory manager waiting three days for a reallocation decision that AI could inform in three minutes - that gap is where you find the value." The resistance is also highest here, precisely because reallocation feels like a challenge to field leadership judgment. Organizations that establish AI-first protocols for bounded, fast-cycle field decisions build the muscle memory that makes AI credible when the stakes eventually rise.
The cross-functional handoff is where this breaks down. Joshi identified the organizational translation problem: AI-informed insights that survive the analytics function rarely survive intact when they move to marketing and then to the field. "The insight gets summarized, then summarized again, and by the time it reaches the person who needs to act on it, what's left isn't really an AI recommendation anymore," he said. The structural fix isn't trust in AI. It's workflow design that preserves AI-informed decisions through organizational handoffs rather than allowing each function to reinterpret and dilute them.
Making override visible is the only lever that works
The session converged on a diagnostic that Mukherjee articulated as the distinguishing variable across organizations. "The companies getting value from AI aren't necessarily running better models," he said. "They've changed the conditions under which decisions get made - what's required to go with the AI recommendation and what's required to go against it." That asymmetry is currently inverted in most commercial organizations: following AI requires nothing, and overriding it requires nothing either.
The original insight this panel points toward: the companies that capture disproportionate AI commercial value in the next launch cycle are the ones that have made override a formal, documented act, subject to the same justification requirements as other high-consequence decisions. Not a cultural nudge, not a training program. A governance requirement. If the AI recommendation is overridden, the override is logged, the rationale is recorded, and the outcome is tracked. This single structural change generates the data that currently doesn't exist - override frequency, override cost, AI accuracy over time - and creates the institutional record that either validates or challenges the override instinct. It breaks the cycle not by demanding trust but by making distrust visible and accountable.
Palumbo described, "The goal isn't to have AI make the decision," he said. "It's to make sure that when a human overrides it, that override is a real decision - not just a reflex." That distinction between a decision and a reflex is where most AI value is currently being lost.
Sabbah captured the trust gradient that underlies all of this. "We trust AI completely for low-stakes operational tasks - no one argues with the CRM data," he said. "But as the stakes go up, the trust goes down, in almost perfect inverse proportion to where the value is." The inverse correlation is precise and predictable. It will not self-correct.
Petrocelli closed with the frame that holds the session together: "The human-AI partnership only works if both sides of the partnership have clearly defined roles - and right now, only one side does."
For any commercial leader running AI tools in launch or in-market operations: identify the three highest-stakes decisions in your next launch cycle. Map where AI output enters the decision process and where it exits. If the AI recommendation is present early and absent at the moment of decision, you have located where your investment is being converted into expensive wallpaper. The fix is not in the model. It starts with making the override visible - and in pharma's next launch wave, the organizations that govern override will outperform the ones that simply deploy better algorithms.
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.
