Medical Affairs Insights Platforms: Turning MSL Field Notes Into Evidence

What an Insight Actually Is

Medical affairs has spent a decade arguing about its own strategic standing. McKinsey’s widely cited view is that the function has the opportunity to become a primary strategic pillar of the pharmaceutical organization, alongside research and development and alongside commercial and market access.¹ The trade press has been asking ever since whether that has actually happened.² The honest answer is that it happens in the companies where medical affairs produces something the rest of the organization cannot get anywhere else, and does not happen in the companies where it produces activity reports.

Field insight is the clearest candidate for that something. No other function has thousands of unpaid, unstructured, non-promotional conversations a year with the people who make the treatment decisions. Market research is bought and framed. Sales calls are transactional and, by design, promotional. Published literature is eighteen months behind. The MSL conversation is the only place where a clinician says, unprompted, that they have stopped using the therapy in a particular subgroup and here is the reasoning, and that reasoning is not in any dataset the company owns.

The Medical Affairs Professional Society frames insights as action-oriented conclusions drawn from analysis of multiple sources, not as observations that stop at stating a fact.³ That is the right instinct, but the definition is hard to apply at the moment of capture, when an MSL has ninety seconds and a phone. A more usable working test at the point of entry is this: could a named person, in a named function, start work because of this? If nobody can, it may still be worth recording, but it is not an insight and should not be counted as one.

It helps to be explicit about the three things that get recorded in the same box and should not be treated the same way.

The difference between the second and third rows is not effort in the software. It is whether the MSL was asked the right follow-on question and had somewhere useful to put the answer. That is a design problem, and it is where most insights programs are won or lost.

Why the pattern matters more than the individual note

A single clinician’s reluctance to treat older patients with reduced renal function is a data point. It might reflect the evidence, or it might reflect that clinician’s caution. Fifteen clinicians raising the same gap across four countries inside a quarter is a different object entirely. It is a defensible input to a study concept, a rapid literature review, a new standard response document, or a decision to prioritize a subgroup analysis of existing data.

This is the whole argument for treating insights as structured data rather than as reading material. Nobody can hold fifteen conversations from four countries in their head, and nobody will read three thousand call notes looking for them. The pattern is only visible if the underlying records share enough structure to be counted.

The Capture Problem: Free Text and Rigid Dropdowns Both Fail

Almost every insights program starts by choosing between two bad options, and most choose badly.

Free text preserves everything and enables nothing

An insight recorded as an open paragraph in a call note is faithful to the conversation and useless to the organization. It cannot be aggregated, because two MSLs describing the same clinical gap will use different words. It cannot be trended, because there is no consistent unit to count. It cannot be routed, because no rule can reliably tell a publication gap from a safety observation from a competitive comment. It cannot be searched with confidence, because the terms that matter (a subgroup, a comorbidity, a dosing interval) appear in a dozen forms.

Free text is a container, not a record. The information is genuinely in there. That is not the same as the organization having it.

Rigid dropdowns produce complete fields and empty content

The standard fix is to put structure in front of the free text: a category picker, a subcategory picker, a therapeutic area, a priority rating, a sentiment score, an impact estimate, a strategic pillar alignment. Twelve required fields, most of them offering values that were written by people who have never had the conversation.

What happens next is predictable. The MSL, writing at nine at night after a day of travel, picks the least-wrong option in each dropdown and writes two sentences of free text. The categories are all populated. The completion rate is excellent. And the actual content, the specific reasoning that made the exchange worth capturing, was never written down because the form consumed the attention that would have gone into writing it.

This is worth stating plainly because it is the central failure of the category: over-structuring does not produce structured insight, it produces compliance-shaped nothing. The system now contains thousands of well-formed records that no analyst can do anything with, and the organization concludes that field insight is low value, which is a conclusion about the form design rather than about the field.

The real constraint is the MSL’s attention, not the MSL’s willingness

Insight capture happens in cars, in airports, in hotel rooms, and in the ten minutes between meetings. Every field you add trades directly against the fidelity of what gets written. This is not a motivation problem to be solved with training or with performance metrics. It is a budget of roughly two to four minutes, and the design question is what to spend it on.

Spending it on classification is the wrong choice. Classification is the part a person in an analytics or medical information role can do later, at a desk, with the full text in front of them, and increasingly with model assistance. Spending it on specificity is the right choice, because specificity is the part that is only available in the moment and is gone forever once the MSL has moved on to the next day.

Industry investment is not currently pointed at this. ZS surveyed more than 200 medical affairs professionals from more than 35 companies and 150 key opinion leaders in early 2025, with 76 percent of internal respondents at director level or above. Nearly three-quarters said their organizations are prioritizing investment in insights collection and analytics systems, and roughly 80 percent said they are concentrating on platform integration, scalability, and advanced analytics in their CRM.⁴ That is a great deal of spend on the pipes. Very little of it addresses the two-minute problem at the point where the water enters.

A Light Structure That Preserves Substance

The design goal is a small number of fields that make a record countable without making the MSL do the analysis. In practice, four fields carry almost all of the value, and they should sit after the free text, not before it.

Two more fields belong on the record and should never be typed by a human: therapy area and product, and geography and specialty. Both should be derived from the interaction record and the clinician profile already in the system. Asking an MSL to re-enter data the system already holds is the fastest way to lose their trust in the tool.

Three rules that keep the structure honest

Free text first, structure second. The entry screen should open on an empty text box with one prompt: what did they want to know, and why does it matter to them? The structured fields appear underneath, after the substance is captured. This ordering matters more than it sounds. Structure presented first frames the MSL’s recall around the available categories and the specific reasoning never surfaces.

Never make a field required unless you can describe a wrong answer. If the governance group cannot articulate what an incorrect value for a field would look like, the field is not measuring anything and it is spending attention it has not earned. This single rule removes most of the fields on a typical insights form, including nearly every sentiment score and strategic-alignment picker.

Version the taxonomy and keep it small. Each field should hold roughly a dozen values, no more than twenty. When values change, and they will as a product moves through its lifecycle, the change is a versioned event with a documented mapping from old values to new. Retrospective analysis is impossible if the vocabulary silently shifted eighteen months ago and nobody wrote down how.

The Compliance Boundary Is the Architecture

Everything above is standard data design. What makes medical affairs insights different from any other qualitative data program is that a hard regulatory and ethical boundary runs through the middle of the dataset, and the boundary is not a policy you write after the system is built. It is the architecture.

Why the separation exists

Medical affairs interactions with healthcare professionals are non-promotional by definition. The OIG’s compliance program guidance for pharmaceutical manufacturers, published in 2003 and still the reference point for how US enforcement thinks about this, treats the independence of medical and scientific functions from sales and marketing influence as a core control, and identifies the blurring of that line as a recognized fraud and abuse risk.⁵ The PhRMA Code carries the same principle into industry self-regulation for US interactions.⁶ The IFPMA Code, updated and expanded for global application, requires that scientific exchange be non-promotional in intent, content, and nature, distinguishable from promotion, and overseen by the company medical function.⁷ The EFPIA Code embeds the same principles for Europe.⁸

FDA’s position is equally clear on the specific point where the two functions most often collide. Its guidance on responding to unsolicited requests for off-label information directs that such requests be routed to a medical or scientific function, with sales and marketing personnel not involved in preparing the response.⁹ Its final guidance on communications regarding scientific information on unapproved uses, issued in January 2025, sets out at length what firm-initiated scientific communication to healthcare providers must look like to stay outside the promotional frame.¹⁰

What an insights platform does to that boundary

Here is the failure that quietly undoes a well-built program. An insights platform that captures rich scientific detail about named clinicians, and then makes that detail available to commercial teams for targeting, has converted a non-promotional scientific exchange into commercial intelligence. Three things follow.

First, the regulatory exposure is real and it is documented. The system generates a durable, timestamped, searchable record of exactly who saw what and when. That is ordinarily a virtue. In an enforcement or litigation setting it is also the clearest available evidence of whether the line held.

Second, the underlying data source degrades. Clinicians participate in scientific exchange because they believe it is scientific exchange. The moment a key opinion leader concludes that their candid reservations about a therapy are being routed into a targeting model, the candor stops. The organization has not gained commercial intelligence. It has lost the only channel where clinicians said true things.

Third, and most practically, medical affairs loses the internal argument. Field medical leadership is the natural owner of an insights program, and the fastest way to lose that ownership is for compliance to discover that the outputs are being consumed commercially. Programs rarely die from a compliance finding. They die from the six months of restriction that follows one.

So the boundary has to be explicit, written down, and enforced by the system rather than by good intentions.

The design implications are concrete. Insight records live in a store with its own access model, not in a shared commercial data warehouse with a permission flag. Any release outside medical is an explicit, logged event with a named approver in medical governance, not a report subscription. Aggregation thresholds are enforced in the query layer, so a theme cannot be released until it reflects a minimum number of distinct clinicians. And the insight store shares no join key with commercial targeting systems, which is the control that prevents a well-meaning analyst from reconstructing individual attribution through a linkage nobody anticipated.

The reporting obligation that attaches the moment an MSL hears certain things

The second compliance dimension is not about the commercial line at all. It is about safety, and it is where insights platforms create genuine regulatory risk if the design is careless.

Under 21 CFR 314.80, an applicant must promptly review adverse drug experience information obtained or received from any source, foreign or domestic, and must report each experience that is both serious and unexpected no later than 15 calendar days from initial receipt of the information by the applicant.¹¹ The phrase that matters is “from any source.” An MSL is the applicant. The clock starts when the MSL hears it, not when a report is filed, not when a coordinator reads the note, and not when a text-mining job runs overnight. In the EU, GVP Module VI sets out the parallel obligation to collect and collate reports of suspected adverse reactions from both unsolicited and solicited sources.¹² ICH adopted E2D(R1) in September 2025, updating the international standard for post-approval safety data management specifically to address newer and increasingly used sources of safety information, including digital platforms and organized data collection programs.¹³¹⁴ Product quality complaints carry a parallel intake and investigation obligation on the quality side.

Four design decisions follow, and none of them are optional.

• A single blocking question at the top of every record. Did this conversation include a possible adverse event, a special situation, or a product quality complaint? It is the first thing on the screen and it cannot be skipped.
• A hard routing stop. A yes answer routes to pharmacovigilance or quality intake through the established path before the insight record can be saved. The insight record then carries a reference to the safety case, not a copy of it.
• Scheduled reconciliation. The insight store and the safety database are reconciled on a defined cycle to find cases that reached one and not the other. This is where the design either works or is discovered not to work.
• Text mining as a second net, never the first. Automated detection over free text is a valuable backstop and a poor primary control. It is not deterministic, its recall varies by phrasing and by language, and no inspector will accept it as the sole mechanism for meeting a 15-day obligation.

One further point deserves stating plainly to leadership before a program starts. A structured insights system is a discoverable record. It captures what field personnel knew about a therapy’s real-world behavior and when they knew it, in a form that is far easier to search than a call note archive. That is an argument for building it well, with a clean safety hand-off and a defensible retention policy. It is not an argument for not building it, and the alternative (the same knowledge sitting unstructured and unreviewed in thousands of notes) is not a better position to be in.

Aggregation: Seeing the Pattern Across Many Conversations

The value of an insights program is the pattern across many conversations. Almost every vendor demonstration shows this working. Almost every internal implementation finds that it does not, and the reasons are consistent enough to list.

What real aggregation requires

The denominator problem deserves the extra sentence. A rising count of a theme can mean the clinical reality changed, or that the field team grew, or that a new therapy area launched, or that a manager reminded everyone to log insights. Only the first of those is interesting, and the count alone cannot distinguish them. Rate per hundred interactions, by region and specialty, is the minimum honest presentation.

The reporting bias nobody controls for

Insight volume measures MSL diligence at least as much as it measures clinical reality. A region that looks quiet may have a diligent team facing an uneventful quarter, or it may have a team that stopped writing because nothing ever came back. A theme that appears to spike in one country may reflect one enthusiastic MSL who writes long records.

There is no statistical fix for this. There is an operational one: track contributor concentration. If more than a third of a country’s insight records come from fewer than a fifth of its field team, the trend line is telling you about the team, not about the medicine. Report that ratio alongside the themes and let the reader adjust.

Three views that earn their place

Most insights dashboards contain a dozen charts and get looked at twice. Three views do the work.

Theme rate over time, with denominator. Rate per hundred interactions for each theme, by quarter, filtered to a therapy area. This is the view that tells medical strategy whether something is emerging or whether the team simply logged more.

Concentration by specialty and region. The same themes, shown as where they concentrate rather than how they move. A question that appears evenly everywhere is usually a communication gap. A question that concentrates in one specialty in two countries is usually a clinical practice difference worth understanding.

Novelty. Themes appearing for the first time, and verbatims that classification could not confidently place. The unclassifiable records are frequently the most valuable ones in the system, because a genuinely new observation has no existing category by definition. Route them to a human reviewer weekly rather than letting them sit in an “other” bucket.

Routing and the Closed Loop

An insight that reaches nobody who can commission a study, amend a publication plan, or fill a medical information gap is a filing exercise with a dashboard attached. Routing is not a workflow detail. It is the reason the system exists.

Who receives what

Two features of this table matter more than its contents. Every row has a named function, and every row has a specific action that function can actually take. If a category cannot be given both, it should not exist as a routing destination, because records assigned to it will accumulate without a disposition and field teams will notice.

The feedback loop is the single biggest driver of capture quality

This is the strongest claim in this article and the one most often treated as a nice-to-have. MSLs write good insights when they have seen a good insight change something. They write short, generic, category-satisfying records when they have not. No amount of training, template refinement, or performance measurement substitutes for the experience of watching a field observation turn into a study, a publication, or an answer they can now give.

A working loop has three levels, and organizations routinely build the first and stop.

AI Assistance: Real Help and a Specific Risk

Nearly 90 percent of medical affairs respondents in the ZS survey said their organizations plan to invest in generative AI for data analysis and insights generation, alongside literature review and summarization and content creation.⁴ The direction is set. The question worth asking is which parts of this work models do well and which parts they quietly damage.

Where the help is real

Language models are genuinely good at several tasks in this pipeline, and dismissing them would be a mistake.

• Classification against an existing taxonomy. Assigning a written insight to a therapeutic question type and an evidence gap type is exactly the kind of task where a model performs consistently, and it removes the classification burden from the two-minute capture window.
• Clustering and theme discovery across large volumes. Published work in the Journal of Pharmaceutical Policy and Practice demonstrated a sentence-embedding and density-based clustering approach applied to medical information inquiry data to surface themes at a scale no manual review could reach.¹⁵ The same techniques apply directly to field insight text.
• Deduplication. Detecting that three records describe the same congress conversation is a semantic similarity problem and a good fit.
• Cross-language normalization. Global field teams write in many languages. Consistent classification across them is otherwise very hard.
• Search over years of accumulated text. Retrieval over an insight archive is a straightforward and high-value application, particularly when a new question arrives and someone needs to know whether it has been raised before.

The specific risk: omission, not fabrication

The failure mode that matters here is not the one most governance discussions focus on. Everyone worries about the model inventing something. The more damaging failure in this particular application is that the model summarizes the exchange and drops the qualifier that made it an insight.

Consider what actually carries the information in a scientific conversation. The clinician did not say they will not use the therapy. They said they will not use it in patients over 75 with a creatinine clearance below a threshold, and only when the patient is already on a second agent, and they added that they are not certain, and that a colleague disagrees. Every one of those qualifiers is the insight. A three-line summary that reads “concerns about use in elderly renal patients” is not wrong. It is simply no longer actionable, and nobody reading it downstream can tell that something was removed.

Summarization systems are trained to produce the typical. An insight is by construction atypical, which is precisely why it was worth recording. The two objectives are in tension, and the tension does not show up in the accuracy metrics most teams look at.

The published evidence on medical text summarization is worth reading honestly, because it points in both directions. A 2025 study in PLOS Digital Health evaluating large language models drafting emergency department encounter summaries found that 42 percent of GPT-4 summaries exhibited hallucinations and 47 percent omitted clinically relevant information, with omissions concentrated in the history and examination content.¹⁶ A 2025 framework paper in npj Digital Medicine, applying a structured error taxonomy and iterative workflow refinement across 12,999 clinician-annotated sentences, reported a 1.47 percent hallucination rate and a 3.45 percent omission rate.¹⁷ Separate work on evaluating omissions in medical summarization has argued that omission is systematically harder to detect and measure than fabrication, and is under-represented in standard evaluation approaches.¹⁸

The gap between those two error rates is the finding. It is not evidence that one model is better than another. It is evidence that error rates in this task are governed by workflow design, prompt construction, and human review placement far more than by model selection. A team that buys a platform and accepts its defaults will land near the high end. A team that designs the task, builds a labeled evaluation set from its own therapy area, and places human confirmation at the right point will land near the low end.

Design rules for model assistance

• The verbatim is the record. The model output is a proposal. No model-generated summary ever overwrites or replaces the original text, and every downstream view offers one click back to the verbatim.
• Human confirmation on classification, at a defined threshold. Auto-accept above a confidence level for low-consequence fields, human review below it, and human review always for anything touching safety, access, or the compliance boundary.
• Never auto-route on model output alone across the boundary. No model decision should be the sole basis for releasing content outside medical affairs or for classifying something as not a safety event.
• Validate against your own labeled set. Vendor benchmarks are built on general clinical text. Build a few hundred labeled examples from your own therapy area and measure against those, and re-measure when the model version changes.
• Measure omission explicitly. Include an omission rate in the evaluation, not only accuracy or agreement. It will not appear on its own.
• Version and log the model as a controlled component. Which model version, which prompt version, which taxonomy version produced each classification. This is ordinary practice for any system whose outputs inform regulated decisions, and it is the difference between being able to explain a historical trend and not.

The broader question of where a person must sit in an automated pipeline in a regulated life sciences setting is one we have written about separately, and the principles transfer directly to this application.

Building It: Sequencing, Governance, and Metrics

Do not start with platform selection

The most common sequencing error is to run a vendor evaluation first. The market is mature and the products are capable, and that is precisely the problem: a capable product will configure itself around whatever taxonomy you give it, including a bad one, and you will not discover the taxonomy was bad for eighteen months.

Start with three questions instead, answered in writing by the people who would receive the output. What decisions do we want to be able to make that we cannot make today? Who owns each of those decisions? What would we have to see to make them differently? A team that cannot answer the third question is not ready to specify a taxonomy, and a taxonomy specified without those answers will be a list of topics rather than a list of decision inputs.

A workable sequence

• Define the insight in one paragraph, with three worked examples and three counter-examples drawn from your own therapy area. Circulate it until field medical, medical strategy, and compliance all agree it is right.
• Build the taxonomy backward from the receivers. Sit with evidence generation, medical information, and publication planning and ask what they would need to see. Their answers become the values in the evidence gap field.
• Wire the safety and quality hand-off before anything else. This is the one part that must be correct on day one. Everything else can be improved iteratively.
• Run one quarter manually. A structured document, a weekly review, and a monthly summary to the receiving functions. This is not a pilot in the technology sense. It is a test of whether the receiving functions do anything with the output, which is the assumption most likely to be wrong.
• Then select a platform, with a specification written from a quarter of real evidence rather than from a vendor’s feature list.

Governance

An insights program needs a standing group with authority, not a project team. Membership: medical strategy, field medical leadership, medical information, pharmacovigilance, publication planning, and compliance. Its charter should cover four things and stay short.

The definition of an insight and the taxonomy, with a version history. The boundary rules governing what may be released outside medical affairs, who approves each release, and how releases are logged. The disposition process, including service-level expectations for acknowledgment and disposition. And the review cadence for the taxonomy, which should be at least annual and always after a significant product lifecycle event.

Metrics, and the ones that corrupt the system

Validation posture

In most companies a medical affairs insights platform is a business system rather than a GxP system, and treating it as though it needed full computer system validation will slow it down for no regulatory benefit. But two parts of it are different and should be handled accordingly.

The safety and quality hand-off path supports a regulatory reporting obligation. It should be specified, tested, and change-controlled with the same discipline as any other route into the safety database, and the reconciliation between the insight store and the safety system should be a documented, scheduled, and evidenced activity. The boundary release process produces records that may be examined in an enforcement or litigation setting, so its approvals and its audit trail need to be reliable. The rest of the platform can be run as a normal business system with documented change control and periodic review. Getting this proportionality right is what keeps the program moving.

Conclusion

The reason field medical insight so rarely becomes evidence is not that MSLs fail to notice things or fail to write them down. It is that most organizations have treated this as a CRM configuration exercise, when it is a structured data problem with a compliance boundary running through it and a human feedback loop holding it together. Get the capture structure light enough that specificity survives the two minutes available. Make the medical-commercial boundary an architectural fact rather than a policy statement, and wire the safety hand-off so that a reportability clock never starts inside an analytics queue. Give aggregation a denominator. Route every theme to a named person who can start work, and tell the field what happened. Use models for the classification and clustering they are good at, and design the workflow around the fact that the failure you will not see is omission, not invention.

None of that requires a large program. It requires deciding, before any platform is selected, what decisions the organization wants to be able to make and who owns them. Sakara Digital works with pharma and biotech organizations building exactly this kind of structured data capability, where the data design, the regulatory boundary, and the operating process have to be solved together rather than in sequence. If you are standing up a medical affairs insights program, or you have one that is collecting records nobody acts on, and you want an independent perspective on where the design is going wrong, we are happy to have that conversation.

For Further Reading