All articles
Industry TrendsNovember 17, 2026·3 min read

Which AI Tools Are Actually Earning Their Keep at Research Sites

A practical breakdown of which AI and automation applications are genuinely working at research sites, which are still oversold, and how to evaluate a tool without burning a quarter on it.

By Trialflow Team

Every vendor at every conference now has an AI story. Most of it is noise. But underneath the marketing, a few things really are changing in how sites run trials, and it's worth separating what's landed from what's still a demo.

What's actually working at sites today

Transcription and note drafting. Speech-to-text has crossed the quality threshold for clinical use. Coordinators dictating visit observations and getting usable draft notes back is probably the single most immediately useful application. It doesn't replace source documentation standards — someone still reviews and signs — but it cuts the after-hours charting that drives turnover.

Protocol summarization. Feeding a 120-page protocol into a model and getting a plain-language eligibility summary, visit schedule table, or coordinator cheat sheet works well. It's a reading-comprehension task, which is what these tools do best. Always verify against the source before anyone acts on it.

Pre-screening triage. Sorting inbound referrals by likely eligibility, flagging obvious exclusions from a questionnaire, and routing high-probability candidates to the top of the call list. Sites with heavy inbound volume commonly see this free up a meaningful share of screening time.

Communication drafting. Reminder texts, follow-up emails, recruitment ad variants. Humans review, but the blank page problem disappears.

What's still oversold

Eligibility determination. Any tool that claims it can read a chart and tell you a patient qualifies is making a regulatory claim it can't support. Models hallucinate lab values and misread negations constantly. Use these outputs as a worklist, never as a decision.

Feasibility prediction. Vendors promising to forecast your enrollment rate from historical data are mostly doing regression with a new name. The models aren't wrong, exactly, but they're trained on data that doesn't include your referral network, your PI's clinic volume, or whether your competing study starts next month.

Autonomous outreach. Fully automated conversational agents calling patients remain risky. Participants notice, and a bot that mishandles a medical question creates problems that outweigh the hours saved.

How to evaluate a tool without wasting a quarter

  1. Name the task, not the category. "Reduce time spent on pre-screen calls" is evaluable. "AI-powered recruitment" is not.
  2. Ask what happens when it's wrong. If a bad output reaches a participant or a sponsor without human review, you need a different workflow.
  3. Run it on last quarter's data. You already know the answers. See whether the tool gets them.
  4. Check where the data goes. PHI leaving your environment to a third-party model is a BAA question and an IRB question before it's a technology question.
  5. Measure coordinator hours, not accuracy scores. A 95%-accurate tool that requires full manual verification saves nothing.

The quieter trend worth watching

Sponsors are starting to expect faster site responsiveness — same-day referral contact, real-time enrollment reporting — partly because automation has made it possible for some sites. The competitive pressure is less about having AI and more about the response times it enables. Sites that automate the mundane parts of intake and follow-up will look more capable in feasibility conversations, regardless of what technology they used to get there.

Start with the task that your coordinators complain about most. That's usually a better guide than any vendor roadmap.

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