What AI Can Actually Do at a Research Site Right Now

Every vendor at every conference now has an AI story. Most of it is noise. But a few things have genuinely changed in the last couple of years, and sites that understand which ones are real can make better decisions about where to spend money and attention.

What has actually become reliable

Three capabilities crossed the line from demo to dependable:

Transcription and summarization. Speech-to-text is good enough now that recording a phone pre-screen and generating a structured summary is practical. Coordinators still review and correct, but starting from a draft rather than a blank note saves real time per call.

Unstructured text search. Pulling candidates out of free-text chart notes, referral letters, and old screening logs used to require someone reading everything. Language models handle the first pass well — flagging "possible match, mentions metformin and A1c above range" from a note that no structured query would catch.

Drafting repetitive correspondence. Visit reminders, referral follow-ups, non-qualifier letters. These are templated anyway; generation just makes the templating faster and more personalized.

Notice what these have in common. Each one produces a draft that a human reviews, and each one fails safely — a bad summary gets corrected, a bad candidate flag gets rejected at pre-screen.

Where it still breaks

Eligibility determination is the big one. Models will confidently declare a patient eligible based on an inclusion criterion they half-read. In site datasets where teams have tested this, the error pattern is consistent: the model handles clear-cut cases fine and gets confused precisely where a coordinator would also need to think — overlapping exclusions, washout timing, lab values that are borderline against a protocol-specific threshold. That is the opposite of helpful, because it automates the easy work and gives you false confidence on the hard work.

Similarly, anything that touches a participant without review is a liability. An automated text that tells someone they qualify, or that misstates a visit requirement, is a protocol deviation waiting to happen and potentially an IRB issue depending on what was approved.

Questions to ask a vendor

1. Where does a human sit in this workflow, and what exactly do they approve? If the answer is vague, the tool is either doing less than claimed or more than you want.
2. What happens to our data? Specifically: is it used for training, where is it stored, and will you sign a BAA. Sponsors increasingly ask sites this during qualification visits.
3. Can we see the source for every output? A candidate flag should link back to the note that triggered it. Without traceability you cannot verify anything.
4. What does it do when it does not know? Good tools abstain. Bad tools guess.

A practical sequencing

Start where errors are cheap and volume is high. Pre-screen note drafting, reminder sequencing, and first-pass chart review all qualify. Measure something concrete before and after — minutes per pre-screen, percentage of referrals contacted within 24 hours.

Resist starting with eligibility, randomization support, or anything in the regulatory binder. Those are the areas where a mistake costs you a finding, and they are also the areas where current tools are weakest.

The sites getting value from automation right now are not the ones with the most tools. They are the ones who picked two or three narrow tasks, verified the output for a few weeks, and then stopped checking so closely once they trusted it.