Medha Thapliyal

Sheet 03 Writing Revised October 2026

Sheet 03/11. 28 September 2026. 3 minute read.

The 0.9 percent

Rapid qualitative analysis appears in less than one percent of implementation science studies. AI may finally make it common. The question is what we trade for the speed.

03/11 · Essay

A scoping review published last year counted how often implementation science studies use rapid qualitative analysis. The answer was 0.9 percent. Fewer than one in a hundred qualitative studies in a field that exists, in large part, to shorten the distance between evidence and practice.

I have sat with that number for a while. It is not a criticism of the researchers. Rapid analysis is hard to do well, and the template-based methods that work in a hospital quality team do not always survive contact with forty interviews in three languages across two districts. But the number is also an admission. We say that implementation research should move at the speed of programs, and then we take eighteen months to code the transcripts.

What the new tools actually do

The most mature use of AI in implementation science right now is not prediction or decision support. It is coding. Several groups have shown that generative language models can match hand-coding on structured qualitative tasks nearly as well as the supervised machine learning that came before them. A 2026 paper from Binghamton and the University of Colorado went further in a direction I find more interesting: they built an AI-assisted rapid qualitative analysis workflow on small, locally hosted encoder models rather than a cloud API. Their reasons were data sovereignty and customisability. Interview data from frontline workers should not leave the institution that promised to protect it, and a model trained on an expert’s codebook behaves more like that expert than a general-purpose assistant does.

Their conclusion was careful. Small models can meaningfully assist coding if expert coding practices are carefully encoded in training. The emphasis belongs on the second half of that sentence. The model is only as good as the human exemplars it learns from, and most of us have never had to write our coding practice down precisely enough for a machine to copy.

What I worry about

Speed is not neutral. The slow parts of qualitative analysis are where reflexivity lives. Reading a transcript for the third time is when I notice that a supervisor’s complaint about “data quality” is actually a complaint about being blamed for a form she did not design. A model will code that passage as data quality every time, and it will be right in a way that misses the point.

There is also the question of whose constructs we are encoding. CFIR and RE-AIM are useful scaffolds, but there is no validated pipeline for coding either framework with a language model, and the AIM-IS consortium’s reporting standard is still being written. Until it exists, every AI-assisted analysis is a bespoke method, and bespoke methods are hard to review and harder to replicate across model versions.

How I would use it

I would use AI-assisted coding the way I use a second coder: as someone who reads everything, flags where we disagree, and forces me to say why. That means keeping the transcripts local, training on a codebook I can defend, treating disagreement rates as a finding rather than a nuisance, and reserving time for the slow read that the speed is supposed to buy.

If the 0.9 percent becomes nine percent because the tools made rapid analysis feasible, that is good for the field. If it becomes nine percent because we stopped reading, we will have made implementation research faster and less able to see what it is looking at.

Sources

03 · Adjoining

Keep reading.

  1. Nov 2025 Who participates when a model is in the room Participatory research assumes people can see and shape the method. When the method is a model, that assumption needs rebuilding from the ground up.
  2. Jul 2026 Novices with a probe One day of training produced sonographer-level gestational age estimates in Lusaka. The hard question is what happens to the referral chain after the machine says high risk.
  3. Jun 2026 Silicon mothers Language models can produce plausible answers about pregnancy in Assam. Plausibility is not validity, and formative research cannot be outsourced to a model that has never missed an antenatal visit.