All articles
Industry TrendsNovember 17, 2026·3 min read

What AI Is Actually Doing at Research Sites Right Now

Separating the AI capabilities that are genuinely working at research sites from the ones still being oversold, plus the five questions to ask any vendor before a pilot.

By Trialflow Team

Every vendor at every conference now has an AI story. Most of it is noise. But underneath the marketing, a few shifts in clinical trial operations are real, and they're worth understanding so you can tell which pitches deserve a demo and which deserve a polite no.

What is actually landing at sites

The AI capabilities that have found real traction at research sites share one trait: they handle language-heavy work that was previously manual, and a human reviews the output before it matters.

Chart review and pre-screening. Tools that read structured EHR data against inclusion and exclusion criteria have gotten genuinely useful, especially for studies with lab-value or diagnosis-code gates. The practical limit is unstructured notes. A model can flag "possible candidate" from a progress note, but a coordinator still opens the chart. Treat these tools as a filter that shrinks your review pile, not as a screening decision.

Document drafting. Recruitment ad copy, participant-facing visit reminders, lay summaries of protocol changes, first drafts of deviation narratives. Sites commonly find this is where time savings show up fastest, because the alternative was a coordinator staring at a blank page at 6pm.

Transcription and note capture. Phone pre-screens and consent discussions transcribed automatically, then summarized into your tracking system. Quality here has improved sharply. Verify how the vendor handles PHI in transit and whether audio is retained.

Scheduling and reminder logic. Less glamorous, more reliable. Rules-based automation that confirms visits, reschedules no-shows, and escalates to a human after two failed contacts does not need a language model at all, and it outperforms most AI pilots on measurable outcomes.

What is still oversold

Be skeptical of anything promising to predict enrollment, identify "high-propensity" participants from thin data, or auto-generate monitoring responses. Prediction requires history, and most site datasets are too small and too study-specific to train anything meaningful. A tool trained on sponsor-level aggregate data usually cannot tell you much about your clinic's population.

Also be careful with tools that write anything a monitor will read. Generated text that sounds plausible but misstates a date or a sequence of events creates more work than it saves.

Questions to ask any vendor

  1. Where does our data go, and is it used for training? Get the answer in the contract, not the sales call.
  2. What happens when the model is wrong? If there's no human checkpoint before the output affects a participant or a record, walk away.
  3. Is this auditable? You need to show a monitor who decided what and when. "The system flagged it" is not a sufficient answer.
  4. Does it write into our existing systems, or create a new one? A tool that requires double entry will be abandoned within a quarter.
  5. What does implementation actually require from us? Mapping, validation, and staff training usually cost more hours than the pilot saves in month one.

A reasonable posture

Pick one high-volume, low-risk task — pre-screen call summaries, or reminder sequences — and automate that well before touching anything regulated. Measure one thing: hours returned to coordinators, or days off your enrollment timeline. If you can't name the metric before you start, you're buying a story rather than a capability.

Enrolling studies shouldn't be this hard

Trialflow gives research sites one platform for lead management, AI trial matching, and recruitment analytics.

Request a Demo