Which Site Tasks Are Actually Ready for Automation

Most conversations about AI at research sites stall in one of two places: vague excitement about "AI matching," or blanket anxiety about handing patient decisions to a model. Neither helps you decide what to do on Monday. Here's a more useful frame — sort automation candidates by how much judgment the task requires and how expensive a mistake is.

The four tiers of site automation

Tier 1: Mechanical, zero judgment. Appointment reminders, no-show follow-ups, visit-window calculations, source document assembly from templates, transferring demographics between systems. These should already be automated. If a coordinator is retyping the same date into three systems, that's the first thing to fix — not because AI is involved, but because it isn't needed.

Tier 2: Pattern recognition with human sign-off. Pre-screening chart review against inclusion criteria, flagging protocol deviations before they're formalized, drafting first-pass responses to routine sponsor queries, summarizing a long medical history into the fields your screening log needs. Language models are genuinely useful here. The output is a draft; a person approves it.

Tier 3: Prioritization and triage. Which of 400 database contacts should get a call today? Which enrolled participant looks likely to drop? Models can rank, but rankings can encode whatever bias lives in your historical data. Use them to order a work queue, not to exclude anyone.

Tier 4: Anything touching eligibility determination, consent, or safety reporting. Keep a qualified human in the decision seat. Full stop.

What's actually shifting right now

The biggest practical change isn't matching algorithms — it's that unstructured text has become workable. Referral emails, call notes, scanned records, and free-text EHR fields used to be dead weight requiring manual reading. Now they can be parsed into something searchable. Sites sitting on years of unstructured prior-contact notes often find their most underused asset is text they already own.

The second shift is voice. Automated call handling for inbound recruitment lines has gotten good enough that after-hours calls no longer go to voicemail purgatory. Sites commonly find a meaningful share of inbound interest arrives outside business hours, and that interest decays fast.

The third is documentation load. Drafting monitoring visit responses, deviation narratives, and recruitment status summaries is high-volume, low-creativity writing. Assisted drafting cuts hours without touching clinical judgment.

Evaluating a tool without getting sold

Ask these before you sign anything:

1. What happens when it's wrong? If the failure mode is an awkward text message, tolerance is high. If it's a missed exclusion criterion, tolerance is zero.
2. Is the output auditable? You need to show a monitor who did what and when. "The system decided" is not an answer.
3. Where does PHI go, and does the vendor train on it? Get this in writing, in the agreement, not in a sales call.
4. Can a coordinator override it in under ten seconds? Tools that make the human path harder than the automated path produce workarounds, not efficiency.
5. What's the setup cost in staff hours? Many tools save five hours a week after forty hours of configuration nobody scheduled.

The realistic near-term picture

AI is not going to enroll patients for you. It will reduce the administrative tax on the people who do. Sites that benefit most tend to start narrow — one workflow, measured before and after — rather than deploying a platform and hoping.

Pick your worst repetitive task. Automate that. Then pick the next one.