SPEAKERS:
Elizabeth Beringer, President, Real Chemistry Media
Leo Londono, Executive Director, Head of Omnichannel Customer Engagement, Ionis Pharmaceuticals
KEY TAKEAWAYS:
- Adoption of personalization engines requires insight-backed action items for the field
- Scalable content frameworks are necessary to develop enough of the right content rapidly to meet the needs of HCPs as they are on their customer journey
- Content infrastructure debt - not engine quality - is the primary bottleneck blocking personalization ROI
- Trust must be pre-engineered into system design before adoption data can prove the system works
- Treating personalization as a platform launch rather than a structured experiment makes ROI indefensible
"The problem isn't data," Elizabeth Beringer argued at Reuters Events: Pharma USA. "The problem is that there is too much noise."
That reframe matters more than it sounds. Commercial organizations that have spent the last three years building next-best-action engines, enriching data lakes, and signing enterprise agreements with personalization platform vendors have been solving for data volume and algorithmic precision. They have largely ignored whether field teams can hear the signal once it arrives.
Consider what happened when one organization handed reps a single statistic: 94% commercial insurance coverage for their product. Accurate. Well-sourced. Operationally useless. Reps received it and moved on, because a population-level coverage figure tells a physician nothing about the specific patients in their practice, the prior authorization requirements attached to their particular payer mix, or what the rep should say at the next call. The recommendation was right. Nothing changed.
That gap, between algorithmic accuracy and field-team action, is where most personalization investment currently disappears. Beringer's prescription for closing it is not a better model. It is a fundamentally different design philosophy.
Three Failure Modes, One That Matters Most
Diagnosing why personalization systems fail requires distinguishing between problems that are hard to solve and problems that organizations have simply chosen not to prioritize.
"Why most personalization systems fail, in my opinion: it’s because they're slow. Data comes in a month later … I don't care, literally does not matter anymore. They're opaque, they're black boxes. I have no idea how this signal came to me. Should I trust it? I don't know. And they're disconnected."
Three failure modes, but they are not equally tractable. Latency is increasingly a solved problem. As Londono put it, "Real-time targeting is nowadays like table stakes. That's something we want to do. That's something that we are really doing." Disconnection - the fragmentation between CRM systems, content libraries, and field force tools - is an integration problem that modern data architecture addresses, expensively but definitively.
Opacity is different. It persists not because organizations lack the technical capability to explain recommendations but because explanation was never part of the design specification. The system was built to be right. Whether the rep could understand why it was right was treated as a training problem, not an engineering problem.
Most organizations measure recommendation accuracy and adoption rate as separate KPIs, then wonder why the correlation is weak. The missing metric is explanation sufficiency - whether the system provides enough context for a rep to stake their professional credibility on the recommendation in a physician's office. A rep who cannot articulate why they are recommending a particular topic or resource will not recommend it, regardless of the algorithm's confidence interval. They revert to what they already know, what they can defend, what does not require trusting a black box in real time.
Beringer's standard for what transparency requires is precise: reps need to understand "what to do, why to do it, and why now." Current systems routinely answer the first question. The second and third are where adoption breaks down. The implication is that organizations diagnosing low NBA utilization should audit explanation quality before optimizing model performance. "If you send a suggestion and it doesn't say why, the rep is going to ignore it, guaranteed."
The Content Supply Chain That Doesn't Exist
Even where explanation gaps have been closed and field teams trust the recommendation, a second structural failure emerges downstream. "Even when the recommendation is right, the content often isn't there to support it."
This is the part of the personalization failure story that commercial organizations are less willing to tell. The engine identifies that a specific physician is ready for a dosing conversation. The rep understands why the recommendation was generated and believes it. Then they search for supporting content and find either nothing tailored to that message or an asset that went through regulatory review six months ago for a different campaign context. The recommendation dies at delivery, not at generation.
The investment mismatch is structural. Commercial organizations have funded engine sophistication - data infrastructure, ML development, platform licensing - while treating content as a downstream operational concern. Londono's diagnosis of the resulting problem is deliberately provocative: "You don't need 6 million pieces of content. You need the ones that are going to perform well." The scale-first content strategy produces libraries so large that retrieval and relevance become their own bottlenecks.
PRC review cycles in most pharma organizations run three to six months per asset iteration, which makes the A/B testing velocity that personalization requires effectively impossible under conventional workflows. The organizations closing this gap are not accelerating review timelines - they are changing what gets reviewed. Pre-approved modular content libraries, where medical-legal-regulatory teams review and approve at the component level rather than the assembled asset level, allow the same approved claims and visuals to be reconfigured for different physician contexts without triggering a full review cycle each time.
This is the practical middle path between one-off asset production, which has no scalability, and the theoretical ideal of fully dynamic content generation, which most regulatory environments cannot support. "Personalization needs content modules. That's what it needs, period. But simple. It needs content that could be adaptable easily. It's already approved and you can use it differently."
The word "simple" carries weight. The temptation in modular content design is to build for maximum configurability - dozens of interchangeable components, elaborate tagging taxonomies, complex assembly logic. The organizations that have deployed this at scale report that simpler module structures produce faster adoption and fewer compliance edge cases. Complexity in the engine is a feature. Complexity in the content module is a liability.
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.
From Population Data to Physician-Level Intelligence
The insurance coverage case study illustrates a principle that applies across every category of personalization signal: aggregate data does not change behavior. Physician-level specificity does.
The pivot Beringer describes is operationally demanding but conceptually straightforward. Rather than telling a rep what percentage of patients have commercial coverage - "because that doesn't actually mean anything" - a transparent system tells them what percentage of that specific doctor's patients have that coverage type, what prior authorization requirements apply to that payer mix, and delivers that information alongside the HCP record before the call. Same underlying data. Radically different utility.
The Dr. Smith recommendation model demonstrates what this looks like at the message level. The opaque version: "Recommended topics: efficacy and dosing." The transparent version, as Beringer frames it: "Dr. Smith has engaged heavily with our efficacy content in the last seven days. Recommend focusing on dosing information because 40% of HCPs who wrote their first script engaged with dosing information next." The 40% figure is drawn from organizational prescription data rather than published research, and its generalizability across therapeutic categories is untested. The structural principle, though, is independent of the specific number: a rep who understands the behavioral pattern driving the recommendation can evaluate it, contextualize it against what they know about this physician, and decide whether to act. That is the difference between a recommendation and a directive. Field teams act on the former and resist the latter.
Trust, once established through this kind of specificity, compounds. Londono describes the mechanics from the field side: "Having a clear framework with your sales reps starts building confidence and starts giving them more confidence so they can trust you. And when you activate and they see it and they value the data, they're going to trust even more." Each accurate, specific, explained recommendation that a rep validates in the field makes the next recommendation easier to adopt. The engine's authority is built incrementally, through demonstrated precision at the physician level - not asserted through platform launch communications.
The maturity ceiling for this model is real. Organizations with physician-level claims data, prescription attribution, and HCP digital engagement signals can operationalize physician-specific profiles relatively directly. Organizations relying primarily on third-party data aggregators face a specificity problem: the data may never reach individual physician granularity with enough reliability to pass field teams' trust threshold. For these organizations, the first investment is not in a better recommendation engine but in a better data foundation - or, paradoxically, in being transparent with field teams about the data's limitations. A system that accurately describes what it knows and what it doesn't generates more durable trust than one that projects false precision.
The Engine Is the Hypothesis, Not the Answer
What connects Beringer's explanation design framework, the modular content prescription, and the physician-level intelligence model is a deployment philosophy that most commercial organizations have not adopted. "Transparent personalization systems really have to answer these questions before they're asked to even gain the adoption to tell if it worked."
The sequencing is counterintuitive but logically necessary: trust must be pre-engineered into the system before adoption data can validate the system's accuracy. Without adoption, there is no performance signal. Without trust, there is no adoption. Organizations that treat trust as the outcome of demonstrated accuracy have the causal chain inverted.
Londono's organization has run this model through two consecutive product launches. "We have gone through two successful launches, back-to-back launches in less than two years" - evidence that the iterative, trust-first approach is not a theoretical preference but a tested operational methodology. Back-to-back launches under a single commercial infrastructure are exactly the stress test that reveals whether a personalization system has been built for scalability or for a single campaign.
The implication for organizations currently planning personalization programs is a direct challenge to how those programs are typically scoped. "A lot of people boil the ocean and then they don't have a way to measure it," Beringer observed. "My one pro tip is: identify, treat personalization systems or engines like exploration experiments. Develop a hypothesis, test it, have it incredibly measurable and defensible, and prove that value quickly and keep iterating from there."
The phrase "measurable and defensible" is doing specific work. Personalization programs that cannot produce defensible ROI evidence within a defined timeframe do not survive the next budget cycle - and they shouldn't. The organizations that have sustained investment in this capability share a common discipline: they scoped their first deployments narrowly enough to generate clean signal, proved value fast enough to protect funding, and built organizational capability through successive iterations rather than attempting comprehensive deployment at launch.
The combined evidence from Beringer and Londono supports a conclusion worth naming: the personalization programs that will prove most valuable are not the most sophisticated at launch - they are the most legible. To field teams, to executives, to the medical-legal-regulatory stakeholders who must approve the content those engines deploy. Sophistication that cannot be explained is indistinguishable from noise. The organizations that close the explanation gap are the ones that will generate the adoption data to prove what the engine actually knows. The engine is one thing. Making that rightness visible, specific, and trustworthy enough to act on. That is the work that will drive adoption and success across the business.
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.
