The demo always goes well. Someone types a question about the product, the chatbot replies in two seconds with a fluent paragraph, and the room nods. Then a medical reviewer asks it for the hazard ratio in the elderly subgroup, and it gives a number that appears nowhere in the approved deck.
That moment is the whole problem with generic document chatbots in Life Sciences. A fluent answer is not the same as a defensible one. When a rep is about to quote an efficacy figure to a specialist, "roughly right" is wrong. This piece sets out the four ways generic chatbots fail on pharma and MedTech product knowledge, what a grounded alternative does differently, and the questions to ask any vendor before you put one in front of your field force.
Most chatbots are built to sound right, not to be right
Many document chatbots follow the same basic pattern, usually called retrieval-augmented generation, or RAG. The system splits your documents into chunks, finds the chunks that look most similar to the question, and asks a language model to write an answer from them.
That is a big step forward from a model answering from whatever it absorbed on the open internet. It genuinely improves access to documentation. But it has a weakness that matters enormously in a regulated setting: semantic similarity is not exact evidence matching.
Picture a rep asking, "What was the reduction in hospitalisation in patients with moderate renal impairment?" The chatbot finds chunks about hospitalisation, chunks about renal impairment, and perhaps a chunk about the overall population. It writes a clean sentence with a percentage. The questions nobody sees answered are the important ones:
- Did that exact number appear in the source, for that exact subgroup?
- Was the condition the same, or did the model borrow a figure from the overall population?
- Does another approved document say something different?
- Was the real figure sitting in a forest plot that the system never read?
These are architecture questions, not interface questions. A nicer chat window doesn't fix any of them.
We are not the only ones worried. In 2024, researchers at Stanford ran a preregistered evaluation of specific commercial AI legal research tools, all built on retrieval, and found that they still produced hallucinated or incorrect responses on between 17% and 33% of queries. Be careful with that figure: it measured those legal tools at that point in time. It says nothing directly about pharma assistants, about any particular vendor in our sector, or about us. What it does show is that retrieval alone did not remove errors in another field where vendors had claimed it would.
Failure one: numbers that look precise and aren't
The most dangerous answers are the most convincing ones.
A typical approved package for a product contains dozens of percentages: primary and secondary endpoints, different timepoints, the intention-to-treat population and several subgroups, safety rates by dose, comparisons with placebo and with standard of care. A generic chatbot that retrieves "related" content can easily stitch a figure from one context onto a question about another. The sentence reads well. The number is plausible. It is also not in the label.
A grounded system should follow a much stricter rule: every numeric value it reports must appear verbatim in the indexed evidence, with its condition. If the exact value isn't there, the system should say so and point to where the closest evidence is, rather than filling the gap.
Some people see that as the chatbot being less helpful. We see it as the chatbot being honest. A rep who hears "that subgroup figure is not in the approved material, ask your MSL" has been protected. A rep who hears an invented figure has been set up.
A scenario we see often: a new rep, three weeks into a launch, is preparing for a meeting with a key opinion leader who sat on the steering committee of the phase III trial. That physician knows every number in the publication by heart. If the rep quotes a subgroup figure that is off by two points, the meeting is effectively over, and the KOL will remember it at the next advisory board. The cost of one fabricated number is not one bad answer. It is the credibility of the whole team with the people whose opinion shapes the market.
Failure two: the charts the system can't see
A great deal of the most important evidence in a product dossier does not live in paragraphs. It lives in Kaplan-Meier curves, forest plots, dose-response charts, device specification tables and comparison figures in the visual aid.
Many document chatbots rely mostly on text extracted from PDFs. That means the chart caption gets indexed, but the values inside the chart are invisible. Ask for the event rate at 24 months and the system either says it doesn't know or, worse, answers from a nearby paragraph that describes a different timepoint.
A more rigorous approach reads figures and tables as evidence in their own right, using a model that can interpret images, and then cross-checks what it read against the document's own text. Where the two disagree, a human should see it before anyone in the field does.
A rep who hears "that figure is not in the approved material, ask your MSL" has been protected. A rep who hears an invented figure has been set up.
Failure three: contradictions that stay hidden
Product documentation changes all the time. A label update adds a new indication. A publication adds a longer follow-up. A local affiliate adapts a global deck for its formulary context. An older training document never gets retired.
So at any moment, a library of approved content is likely to contain two documents that say slightly different things about the same data point. That isn't negligence; it's how living documentation works.
A basic retrieval set-up often returns whichever chunk scores highest for similarity, and the contradiction disappears from view. The rep never knows there was a choice.
A defensible system does the opposite. When two indexed documents disagree on the same value under the same condition, it flags the conflicting sources and dates for the content owner and avoids presenting either value as an approved answer until the conflict is resolved. The job of a knowledge assistant in a regulated company is not to hide ambiguity. It is to expose it early, when fixing it costs an email rather than a correction letter.
Failure four: treating facts as loose text
Underneath all three failures sits a design choice. Basic RAG systems treat your documents as free text to be searched. That works for "what is the mechanism of action?" It breaks down for "what is the value, in this population, measured this way, at this timepoint?"
"12%" on its own is not knowledge. It becomes knowledge only with its context: 12% of what, in whom, compared with what, measured how. A stronger pipeline extracts each fact during indexing as a structured record, with the subject, the value, the unit, the method and the condition held together. When a question comes in, the system matches on those fields as well as on meaning.
This sounds like plumbing, and it is. But it's the plumbing that decides whether an answer can be traced back to a source line, which in turn decides whether Medical and Legal will ever sign off on putting the tool in front of reps.
Speed matters, but only with fresh evidence
Field users won't wait. If an answer takes twenty seconds, they'll stop asking. So it makes sense for a system to remember answers to frequent questions and return them quickly instead of running the full retrieval every time.
The catch is freshness. A cached answer that survives a label update is worse than a slow one. Any cache in a regulated setting needs automatic invalidation: the moment a source document changes, every stored answer that depended on it should be discarded and rebuilt. Speed is a feature only when the evidence underneath it is current.
The failed questions are the most valuable data you'll get
Here's the part of this story most teams overlook.
When a grounded assistant says "that is not in the approved material", it is not only protecting the rep. It is generating a signal. Every unanswered question tells you something:
- a gap in the documentation that Marketing or Medical should fill
- an objection HCPs are raising that the visual aid doesn't address
- a competitor claim that is landing in the field
- a topic where the team is uncertain and would benefit from reinforcement
- a question that belongs with Medical Affairs rather than the promotional team
Across a field force, those questions add up to a live map of where the story is weak. In our experience that visibility can end up being worth more than the answers themselves, because it feeds straight into the next cycle meeting, the next round of practice questions and the next version of the materials. A chatbot becomes a sensor, not just a help desk.
A fair objection: "our reps just need something quick"
A reasonable pushback we hear from Commercial Excellence teams goes like this: all these safeguards make the tool slower and more cautious, and reps will just go back to searching their email for the deck.
There's something to that. A system that refuses everything is useless. The answer isn't to loosen the rules on numbers and claims, though. It's to make the honest answers useful: return the closest verified evidence with its source, suggest the right person to ask, and make sure the common questions are fast. In practice, reps don't mind hearing "not in the approved material" once they learn that when the tool does give a number, it comes with the source line they can check. A citation is a way to verify, not a substitute for verifying: reps should still use the approved visual aid and label for what they say to an HCP, and the tool should be reviewed regularly by Medical. Trust built that way is what brings them back.
The other objection is cost: grounded systems take more work at indexing time. True. But compare that with the cost of a single off-label or misquoted claim reaching an HCP, and the effort looks modest.
What to do before you choose
If you are evaluating a knowledge assistant for your field force this year, don't start with the interface, and don't rely on comparisons made even a year ago: these tools change fast, so test the current versions. Start with a test set. Take thirty real questions from your reps and your MSLs, including subgroup figures, chart-only values and at least two known contradictions in your document library. Run them through every tool on your shortlist and have someone from Medical score the answers against the source, using the rubric in the box.
You'll learn more in an afternoon than from any number of polished demos.
At Atrivity we build knowledge assistants that answer only from the documents a client has approved, with those rules on numbers, sources and contradictions, and we feed the unanswered questions back into the practice games the field already plays. But whichever tool you pick, the principle is the same. In documentation-heavy industries the challenge isn't generating answers. It's generating answers your teams can trace to an approved source in seconds, and that Medical is willing to stand behind.