Why medical insights die in free-text fields

An insight typed into a free-text field is not lost. It is stored, searchable in principle, and almost never read again. Free-text capture fails for three structural reasons, and not one of them is that field teams should write better notes.

The usual diagnosis is effort. People are busy, the form gets filled in at the end of a long day, the writing is rushed. If everyone were a little more disciplined the data would be usable. That diagnosis is wrong, and it is popular because it puts the fix on the people with the least power to make it.

What qualifies as an insight in the first place is a separate question, answered in what counts as a medical insight. This piece assumes a good one has been captured and asks what happens to it next — which is where most of the measurement problem this sits inside actually bites.

What happens to an insight written in free text?

Follow one through. An MSL comes out of a good meeting and writes three sentences into a field.

It saves correctly. It attaches to that meeting, in that account, in that territory. Nothing has gone wrong.

It then appears in a report as a row. Somebody preparing a quarterly summary opens four hundred rows and reads perhaps forty of them properly — not through negligence, but because four hundred free-text entries is not a readable document and never was.

Finding it again requires already knowing it exists. Search only works if you can guess the words someone else chose, and nobody looking for evidence gaps in older patients will surface a sentence that reads "still not convinced about the over-80s stuff." The insight is present and unfindable at the same time.

The functions who could act on it never see it directly. Publications, evidence generation and medical content do not read the insight log. They receive a summary written by the person who read forty rows out of four hundred, filtered through what that person recognised as important.

So the insight is not lost. It is archived, which is worse, because an archive creates the appearance of a working process. Reports get produced. The field is being filled in. The organisation has a complete record of its insights and no knowledge of them, and nothing in the system will ever flag the difference.

Three reasons free text fails, none of them about effort

It is not comparable. Two MSLs describing the same underlying finding will write two different sentences, and nothing joins them together. Which means the thing actually worth knowing — that eleven clinicians across four territories raised the same concern this quarter — is invisible. A single insight is rarely valuable on its own. The repetition is the value, and free text can store insights while being structurally incapable of storing repetition.

It has no addressee. The field records what was learned. It does not record who needs to know. So routing falls to whoever reads the report, once per cycle, using their own judgement about relevance. Anything they do not personally recognise as important stops there, permanently, and no one is ever aware it stopped.

It cannot express absence. A blank field means either that nothing new came out of the conversation or that there was no time to type. Those are opposite facts and the system cannot tell them apart. That ambiguity undermines any measure built on top of it, and it biases every aggregate quietly towards whoever is most diligent with forms rather than whoever is having the best conversations.

None of the three improves if people write more carefully. A better-written sentence is still not comparable to another one, still has no addressee, and a blank is still a blank.

What structured capture actually requires

Four things. The fourth is the one people assume is excluded, and it is not.

A small set of types that reflect how insights are used, not how they arise. Something like an evidence gap, a practice pattern, an access or pathway barrier, a misreading of the data, an unmet need — illustrative rather than definitive, and the specific list matters less than the count. Small enough that a person holds the whole set in their head without looking.

A link to the thing it is about. Which product, which scientific topic, which planned talking point it came out of. Without that anchor an insight cannot be aggregated with anything, and comparability was the entire point.

A route. Each type maps to a function that receives it. This is the requirement most often skipped, and the test is simple: if a type has no owner, it should not be a type. A category that routes nowhere is a folder.

Somewhere for the words to still go. Structure wraps around free text; it does not replace it. The sentence the clinician actually said is the single most valuable element in the record and it has to survive intact. Anyone selling structured capture as the removal of prose has misunderstood what makes an insight worth anything.

Read together: structured capture is not a form instead of a paragraph. It is a handful of fields around a paragraph, which make the paragraph findable, comparable and routable.

Why typing an insight is not the same as categorising it

Here is the honest objection. More structure means more fields, more fields mean more time, and more time at the end of a long day means less gets captured at all. That objection is largely correct, and it is correct for a reason worth naming precisely.

Typing is recall. Categorising is judgement — and it is judgement demanded at the worst available moment, after the meeting, in a car park or a corridor, with the next appointment starting.

Worse, categorising asks the MSL to predict how somebody in another function will want to consume the information. That is not their job, they have no visibility of it, and they are badly placed to guess.

Two things follow, and neither is anybody's fault. The taxonomy gets used approximately, because when nothing fits exactly people select the nearest-looking option and move on. And what gets recorded is the MSL's interpretation at the moment of filing, not what the clinician said.

The fix is not a better-designed form. It is to derive the structure from the conversation and ask the person to confirm or correct it. Confirming a proposed classification takes a few seconds; producing one from scratch takes thought. It also changes what the record rests on, because the evidence for the classification is the thing that was actually said, which means it can be checked later by someone who was not there.

Which makes it worth asking a vendor directly who performs the structuring. If the answer is that the MSL selects from a drop-down, the workload objection stands in full and nothing has been solved.

What you lose when structure is too rigid

Structure has its own failure mode, and it is worse than the problem it replaces rather than better. This is the part most vendors leave out.

A large taxonomy produces noise, not precision. Forty categories do not yield forty kinds of insight. They yield the same handful, distributed close to randomly, because nobody holds forty options in working memory and everybody picks whichever looks nearest. The output looks highly specific and carries less information than the free text did, while being much harder to argue with.

Rigid structure cannot hold the insight that does not fit. This matters more than it sounds, because the finding nobody anticipated is by definition the one the taxonomy was not built for — and the unanticipated finding is frequently the valuable one. If there is no route for it, it does not get recorded, and its absence is invisible.

A fixed taxonomy freezes the questions you were asking when you wrote it. Two years on, the organisation is still asking them. The categories start shaping what the field notices, which is the same failure as a bad measure, arriving through a different door.

So the answer is a middle position, and it has four parts. Few types rather than many. An explicit path for the thing that fits nothing, going to a person rather than into a bucket labelled Other. The verbatim always preserved alongside whatever structure was applied. And the taxonomy reviewed on a schedule, because it will eventually be wrong and nobody notices that on their own.

There is one question that separates a structured system worth having from an expensive version of the field it replaced: can it record something it was not designed to expect? If the honest answer is no, the structure is not capturing insight. It is filtering it.

RocketMSL derives structure from the conversation and asks the MSL to confirm it, keeps what the clinician actually said alongside it, and routes each insight to the function that can act on it. See how insights are structured as they are captured →