Picture the week the launch materials clear MLR. There's a core visual aid, a summary of the phase III trial, an objection handler that went through six rounds of comments, a competitor grid, a safety section nobody wants to paraphrase, and a folder of local adaptations still arriving from the affiliates. Somebody asks the obvious question: how do we turn all of this into something the field will actually learn before the kickoff?
Not long ago the honest answer was "slowly". Today a generative model can draft a full launch curriculum from those documents in days. That part is real, and we use it every week. But speed is the easy half. Our argument in this piece is simple: AI should build launch learning fast, stay locked to approved sources, and then hand over to measurement, because the only thing that matters is whether the knowledge reached the field and stayed there. Below we set out what AI does well, what it must never do in a regulated setting, and how to turn every generated question into evidence.
Launch teams are being asked to do more launches, with the same people
IQVIA's Early Bird review (March 2026) estimated that 73 novel active substances launched globally in 2025, above the average of the past decade. On the US side, RAPS, reporting FDA's own CDER figures in January 2026, said the agency approved 46 novel drugs in 2025, 23 of the 46 had orphan designation, and 85% were approved on the first review cycle.
Expedited pathways and first-cycle approvals compress the time between "likely approval" and "field ready". The label can land earlier than the plan assumed.
Then there's the portfolio math. IQVIA's Launch Excellence IX report (June 2025) put roughly $200 billion of revenue at risk from loss of exclusivity by 2030, and concluded that companies will need an unprecedented number of launches with consistently strong performance. The training team did not grow to match.
The question is where generative AI helps and where it quietly creates risk.
Where AI earns its place: drafting, languages and better wrong answers
Our case library covers 100 programs; launches are one use case among several, alongside kickoffs, certifications, congress follow-ups and always-on product knowledge. In all of them the bottleneck was rarely ideas. It was hours. Writing good multiple-choice questions is slow, careful work, and most medical and brand teams simply don't have the time during launch preparation. This is where a model, used properly, changes the economics.
- Drafting at speed. Questions, short micro-content, nano-content that fixes a single misconception, even short podcast scripts, all built from the approved documents in a fraction of the time a writer needs.
- Language scale. Content prepared in many languages (our platform runs in more than 50) so that global and local teams start from the same message on the same day, instead of affiliates waiting weeks for a translation cycle.
- Calibrated distractors. The wrong options in a question are where the measurement lives. A good distractor mirrors a real mistake: the older guideline threshold, the competitor's dosing, the secondary endpoint confused with the primary endpoint.
- Coverage. A model can map every claim in a visual aid to at least one question, which makes it much harder for a whole section of the story to go untested.
That third point deserves a moment. If every distractor is obviously silly, everyone scores well and you've learned nothing. If the distractors reflect real confusions, the answer pattern tells you which confusion is out there, and in whose territory.
Speed also changes design. When drafting is cheap, you can split the curriculum into waves (disease first, then product) or prepare a deeper version for the MSL team.
What AI must never do in Life Sciences
Here's the uncomfortable part. A general-purpose assistant that improvises is a liability in this industry. It will happily "improve" a claim, round an efficacy number, borrow a comparison from somewhere on the internet, or describe an indication more broadly than the label allows. Each of those is a compliance problem. If it goes into a question, it's worse: you are actively training the field to say something you can't say.
A generated question that teaches an off-label nuance is not a small error. It is the error, repeated by every rep who gets it right.
So the rules we apply are strict, and we'd suggest any team adopt them regardless of which tools they use.
Source-locked generation. Only approved documents go in. Nothing outside them is used to build a question, an explanation or an answer. If the source doesn't say it, the content doesn't say it.
Experts sign off. AI drafts, people approve. Medical, regulatory and brand reviewers see every item before it's published, exactly as they would for any other promotional or training material that goes through MLR.
Conflicts get flagged, not resolved by the machine. Launch documents often disagree with each other: an older slide says one thing, the updated summary says another, a local adaptation uses a different number. A disciplined system surfaces those conflicts for a human to decide. It never picks one quietly.
Your data stays yours. For us that means hosting in the EU and tenant isolation per customer, so one company's launch materials never touch another's.
A fair objection is that all this review slows things down again. Doesn't sign-off put us back where we started? In our experience, no. Reviewing a well-drafted question is far faster than writing one from scratch, and when the draft already cites the exact passage it came from, the reviewer checks it against the source in a minute. The drafting saving survives, but be precise about what it is: AI shortens drafting, not your MLR cycle. Total time to approval still depends on how many reviewers you have and how many rounds they need. What changes is that reviewers spend their time judging, not writing, and the risk of an unsourced claim slipping through goes down, though not to zero.
One question, from document to field
An illustration of the workflow, not a real item. The approved trial summary states the primary endpoint result, with its comparator and population. The model drafts a question with four options: the correct result, the secondary endpoint figure, the result in the wider population, and a competitor's number. It cites the page. The medical reviewer spots that the draft's question stem says "patients" where the label says "adults with moderate to severe disease", corrects the stem, and rejects the competitor distractor because that figure is not in any approved document. The corrected item goes back through the normal review step and only then goes live. The drafting took minutes; the human judgement took the time it always should.
Faster content is only half the answer
This is where many AI-for-training conversations stop, and it's where the real problem begins. A launch team can now produce more learning material than ever. That doesn't mean the field knows more.
Murre and Dros's 2015 replication of Ebbinghaus, a single-participant relearning experiment with nonsense syllables, shows that unrehearsed material fades fast and that spacing helps. Nobody should map it onto a sales force's product knowledge. But it is a useful warning. A beautifully produced module consumed once, the week before kickoff, is fighting a steep curve.
Now multiply the content. If AI helps you produce three times as much and nobody checks what stuck, volume without verification is just a bigger pile.
That's why we treat generation and measurement as one system. Every AI-drafted question has two lives. First, it teaches. Then, once the field answers it, it becomes a data point.
Every generated question becomes a measurement
When a rep answers a question, you learn something specific: this person, on this topic, got it right or wrong at the first attempt. We call a wrong first answer a knowledge gap. It closes only when that same person answers the same question correctly later. Not a different person, not a similar question, not an average.
That definition sounds pedantic. It's the whole point. Across 100 programs in our case library, spanning 39 organisations from 2016 to 2026, we recorded about 9 million answers. Those answers identified 794,130 knowledge gaps at first attempt. Of those, 345,366 were verified closed, about 43%.
First, the spread. Closure across programs ranged from 18% to 80%, with 43% of identified gaps verified closed across the library. Similar content, similar audiences, very different outcomes. We cannot isolate a single cause, but in our experience the design around the content (timing, repetition, competition, follow-up) explains more of it than the content itself.
Second, timing. The 16 programs tied to a business moment (a launch, a kickoff, a certification, a congress, a cycle meeting) closed 55% of gaps, against 41% for the other 84. That is an observed association, not proof that the moment caused it, but it fits our experience that people work harder on knowledge they are about to use. AI makes it possible to have content ready for those moments, rather than a month after them.
For the most important items we set a stricter bar. A critical gap, on a claim, a safety message or a key differentiator, counts as closed only at 5 of 5 later check questions on the same concept. Repeated correct answers reduce, but do not eliminate, the possibility of guessing.
Some gaps won't close with more content, however fast you make it
Here's the finding that should temper anyone's enthusiasm for content volume. In one in five of the example questions we tracked, fewer than 20% of the people who were wrong corrected it; some of those were never retested, so not all are proven misconceptions. Overall, those tracked examples split roughly into thirds: closed, half closed, still open.
Where people did see the question again and kept the wrong answer, the problem is usually not a missing fact. It's a misconception. A rep who believes the mechanism works one way will keep choosing the answer that fits the belief, and another explanatory paragraph from the model won't shift it. In one global immunology kickoff in our library, a mechanism misconception held by 125 reps barely moved with repetition, even while the program as a whole closed 53% of its gaps.
A scenario we see often: week three after launch, a pulmonologist asks how the product differs mechanistically from the current standard of care. The rep answers confidently and incorrectly, because that's the version they believed before training and the training never really challenged it. More AI-generated micro-content is unlikely to fix that. A manager on a ride-along, or a short coaching conversation built around that specific misunderstanding, has a much better chance.
This is where the measurement pays for the generation. Because every question is tracked person by person, you can hand managers a short list: these are the three misconceptions repetition isn't fixing, and these are the people who hold them. Coaching time goes where it pays, instead of being spread across everything.
The loop that keeps running after launch
The kickoff isn't the end of the content problem. After launch the field starts asking questions nobody anticipated: a payer raises a comparison, a KOL mentions a subgroup, a nurse asks about administration in a specific setting. Usually those questions vanish into CRM call notes.
We think this is where source-locked AI is at its most useful. Atrivity Assistant helps field teams find answers based on approved content, with defined review and escalation procedures. And every question they ask becomes a signal. If dozens of reps in two countries ask the same thing in a fortnight, that's a gap in the curriculum, and the next wave of questions should cover it. The assistant and the question bank feed each other.
The result is a cycle rather than a campaign: generate from approved sources, measure what landed, fix each person's gaps, coach the stubborn misconceptions, listen to what the field asks, generate again.
What to take to your next launch meeting
The AI conversation usually starts with speed, and drafting today is very quick. We'd push for two more questions.
- How will we prove, person by person, that the critical messages reached the field and stayed there by the first cycle meeting?
- What will we do about the misconceptions that repetition doesn't fix?
If there's no clear answer to both, the AI project is a production tool, not a readiness program.
That's the approach we've built Atrivity around. AI drafts questions and content from your approved documents only, your experts sign off, and conflicts are flagged before anything is published. Then the field plays, every answer is recorded, Gap Fixing sends each person the fix for their own mistakes, and readiness indexes by team, country and topic show where knowledge is holding and where it isn't. AI to build fast. Measurement to prove it worked.