What AI Actually Does for Sites Right Now

Most of what gets pitched to sites as "AI for clinical trials" is built for sponsors and CROs. It predicts enrollment across a portfolio, flags protocol risk, optimizes country allocation. None of that touches what happens in your building on a Tuesday. But a handful of automation patterns have actually landed at the site level, and it's worth being clear about which ones earn their keep.

What is actually working at sites

Pre-screening triage against inclusion criteria. Sites with EHR access or a sizable registry can now run natural-language matching against protocol criteria instead of manual chart review. The honest framing: it produces a ranked list of maybes, not a list of eligible patients. A coordinator still confirms. But reordering 400 charts so the likely candidates surface first changes the economics of a chart pull considerably.

Transcription and visit documentation. Ambient recording with automated note drafting has moved fast in clinical care and is bleeding into research visits. Useful for informed consent discussions, symptom narratives, and adverse event detail. Still requires review before anything goes near source — and requires a documented process for what counts as source, which is where most sites get into trouble.

First-touch response to inbound inquiries. Automated replies, scheduling links, and structured pre-screen questionnaires do not require AI at all, strictly speaking, but they're often sold that way. They work because speed matters more than sophistication in inbound response. A same-hour reply from a template beats a thoughtful reply three days later.

Drafting routine written output. Recruitment ad variants, participant reminder language, lay summaries, follow-up emails. Low-risk, high-volume, and easy to review. Sites commonly find this is where the first real time savings show up.

What is still mostly marketing

Enrollment forecasting at the single-site level is weak. The sample sizes are too small and the variance too high. A model trained on your last twelve studies cannot tell you much about the thirteenth, especially when the thirteenth has a different therapeutic area and a different referral base.

Automated eligibility determination is also not there. Vendors will frame it as "AI-assisted eligibility" and the assistance is real, but anyone implying you can shorten your medical review is selling something you shouldn't buy.

How to evaluate a tool before you commit

1. Ask what happens when it's wrong. If a wrong answer creates a protocol deviation or a missed exclusion, the tool needs human review built into the workflow, not bolted on.
2. Ask where the data goes. Participant data in a third-party model raises BAA questions, IRB questions, and sponsor questions. Get all three answered in writing.
3. Ask whether your team will actually use it. A tool that requires coordinators to open a second system during a visit will not survive contact with a busy clinic day.
4. Pilot on one study, not all of them. Pick a study with steady volume and a coordinator who will tell you honestly whether it helped.

The practical stance

Treat automation as a way to get low-judgment work off coordinator plates so they can spend more time on the parts that require a human — consent conversations, retention calls, problem-solving with an anxious participant. The gains are real but incremental. Sites that expect a step change get disappointed; sites that chase two or three hours a week per coordinator, repeatedly, end up meaningfully better off.