AI automation for clinical trial sites: what it can and cannot do

Direct answer: what AI can take over at a trial site

A clinical trial site is the place where a study's subjects are enrolled, seen, tested and followed up. The site's office team, usually led by a site manager and carried out by clinical research coordinators (CRCs) and data coordinators, spends much of its time on administrative work around those visits. Most of that work is repetitive: working out which visits fall due, finding the paper or electronic document that proves a result, checking whether a data query has been answered, and sending the same reminder again and again.

An AI worker is a software agent that reads structured information such as the protocol visit schedule, the subject tracker, and the query log, and then performs defined actions on a schedule. For a trial site, the work it can take over is largely the following: scheduling subject visits within protocol windows, chasing missing source documents, tracking query follow-up, and sending visit reminders to subjects. It can also prepare the lists a monitor will ask for before a visit.

For the office team, this means fewer hours spent on reminders and status chasing, and more time on the work a person has to do: preparing subjects, checking data against source, and resolving the questions the system cannot answer. The team does not disappear. The number of people who need to read through a backlog of queries does fall, and that is the change an owner should plan for.

Some limits apply from the start. Informed consent, protocol deviations and decisions, and the assessment and reporting of safety events remain with the principal investigator and qualified staff. An AI worker can prompt a person that a safety-related visit is overdue, but it should not decide what a safety event means.

The whole workflow, end to end

The stages below follow the order in which a study moves through a site. For each one, the first line describes what staff usually do by hand and the second describes what an AI worker can do instead. Where a stage needs a person's judgment, the description says so.

1. Study start-up and feasibility

Staff complete feasibility questionnaires, gather the site's capacity and staffing details, and prepare submissions to the ethics committee or institutional review board and to the competent authority. Much of this is copying the same site facts into different forms. An AI worker can keep one master set of site facts, pre-fill feasibility forms from it, and track the date each submission was sent and when a response is due. A person still signs off the content and sends it.

2. Ethics and regulatory file maintenance

Approvals, renewals, amendments, training records and delegation logs all carry expiry dates. Staff often find out a renewal is due only when a file is checked. An AI worker can list every document with its expiry date, warn the regulatory coordinator 60 or 90 days ahead (the lead time is set by the site), and draft the renewal request for review. Checking that the right protocol version and consent form version are in use remains a person's task.

3. Subject screening and recruitment

Recruiters and CRCs screen referrals against inclusion and exclusion criteria, then book screening visits. An AI worker can sort incoming referrals against a checklist the investigator has approved, flag those that look eligible, and book a screening slot when the coordinator confirms. It should not make the eligibility call on its own. The investigator, or a delegated and trained member of staff, makes that decision and records it.

4. Informed consent

Consent is a conversation. Staff explain the study, answer questions, check understanding, and document the signed or electronically captured consent. An AI worker can prepare the consent visit: confirm the right consent form version is loaded, remind staff which version applies, and log the date the form was signed. It must not explain the study to the subject, take consent, or decide that a subject has understood. The trial site's consent process remains with qualified staff.

5. Enrolment, randomisation and visit scheduling

Once a subject is enrolled, the CRC works through the protocol schedule: day 1, week 4, month 3 and so on, each with a window in which the visit must happen. Staff build calendars by hand, check against lab and imaging slots, and rebook when a subject moves. An AI worker can generate the visit calendar from the protocol timeline, check each planned visit against the window, and flag clashes with lab or imaging capacity. When a visit moves, it recalculates the windows that depend on it and shows the coordinator what changed.

6. Visit reminders to subjects

Subjects are reminded by phone calls, texts or emails, and many sites still make these calls personally. An AI worker can send reminders through the channel the subject has agreed to, at the time the site chooses, and log whether the subject replied. Any message that touches health information needs to be approved by the site and sent only to subjects who have agreed to contact of that kind. Calls where a subject reports symptoms or asks a clinical question go straight to a person.

7. Source documentation

Source documents are the original records that support what goes into the study database: clinic notes, lab reports, questionnaires and dispensing logs. CRCs spend a surprising share of their time chasing these from departments that are busy with patient care. An AI worker can list which expected documents have not arrived for each visit, send a request to the named department, and escalate to the coordinator after a set number of days. It cannot make a missing document appear. It shows the gap and keeps the request moving.

8. Data entry and quality checks

CRCs and data coordinators enter data into the electronic data capture (EDC) system and check it against source. This is the most repetitive part of the day, and the most sensitive. An AI worker can prepare a draft of the data for a visit from the source records, compare it with what has been entered, and list the differences. A person confirms each entry. The AI worker should not enter clinical values into the EDC system without review, because the data must be attributable to a named person under the site's procedures.

9. Query management

Monitors and data managers raise queries when data is unclear or missing. The coordinator has to read each query, find the source, answer it, and track whether the answer was accepted. Queries often sit open for weeks because nobody owns the list. An AI worker can keep a single list of open queries with owner, age and due date, send the owner a reminder, and produce the list the monitor will ask for at the next visit. The answer itself is written by the person who knows the subject.

10. Monitoring visit preparation

Before a monitoring visit, staff pull together the files, check which source documents are still missing, and answer the questions the monitor asked last time. An AI worker can assemble the visit pack: the open query list, the missing document list, the visit calendar for the period, and the signature log. The coordinator still walks the monitor through the file.

11. Safety reporting

Adverse events must be recorded, graded, and reported within timelines set by the protocol, the sponsor and the regulator. An AI worker can flag that a visit recorded a symptom, remind the investigator that an event is open, and show the reporting deadline. It should not grade the event, decide whether it is serious, or submit a report. Those tasks belong to the investigator and qualified staff.

12. Subject reimbursement, sponsor invoicing and close-out

Subjects are often paid travel or time reimbursements, and sponsors are invoiced for completed visits against a payment schedule. Staff match visit completion to the payment table, prepare invoices, and chase payments. An AI worker can match completed visits to the payment schedule, draft the invoice for the finance assistant to check, and track unpaid items. At close-out, it can list which records still need archiving. The sign-off and the decision to pay remain with people.

Staff today and with AI workers

The table below covers the office roles a trial site typically uses. The job titles follow common usage in clinical research. The middle column describes the work the role does today, the third column describes what an AI worker takes over, and the last column describes what remains with a person. Headcount is not part of this table because it depends on the number and phase of studies the site runs.

The pattern is consistent across the rows. The AI worker takes the lists, the calendars and the chasing. People keep the judgment, the conversation with a subject, and the signature.

The 8 highest-value automations, ranked by effort against impact

Ranking the options is a judgment, not a measurement. The ranking below uses two questions: how hard is it to connect the data and set up the rule, and how often does the office team feel the gap when it goes wrong. Low effort and high impact come first. The ranking is a starting point for the owner's own list, and the free automation audit can test it against the site's own figures.

Item 8 has the highest value and the highest risk. Start with items 1 to 3. They change the daily routine quickly and carry the least risk to data integrity.

Compliance and regulatory points for this business type

Trial sites work under good clinical practice (GCP), and several rules shape how an AI worker can be used. The main one is the ICH E6(R3) guideline. The International Council for Harmonisation adopted its final text in January 2025, and the European Medicines Agency lists 23 July 2025 as the legal effective date in the EU. Australia's Therapeutic Goods Administration gives 13 January 2026 as its effective date, with a transition period of 12 months. The guideline's main theme is a risk-based and proportionate approach to trial conduct. Sites in different countries face different dates, so an owner should check the date that applies to the regulator in each country where the site operates.

The second rule concerns electronic records. In the United States, 21 CFR Part 11 sets the conditions under which electronic records and electronic signatures are accepted in place of paper. The FDA's final guidance on electronic systems in clinical investigations, issued in October 2024, states that electronic systems used in trials should be validated before use and that audit trails should record who changed what and why. An AI worker that writes into a trial's records is part of that system, so it needs validation and an audit trail. A reminder tool that only sends messages carries a lighter burden, but the site should still record what it sent.

The third rule concerns subject data. In the European Union, the General Data Protection Regulation treats data concerning health as a special category that is prohibited unless one of the listed exceptions applies. Article 9 sets out those exceptions, including explicit consent, and separate grounds for health care and for scientific research under safeguards. In the United Kingdom, the UK GDPR applies the same structure through Article 9. Sites need to confirm, with their data protection lead, which basis they use for reminder messages and for the records an AI worker reads.

Finally, the site's own contract with the sponsor may restrict which systems can hold study data, where data can be stored, and who can access it. Those terms are set in the contract and the site's procedures, and the AI worker has to follow them.

This explains the rule in general terms; it is not legal or tax advice for your case.

How the transition happens without disruption

The safest way to introduce an AI worker is to leave the current team in place and change one task at a time. In the first month, the AI worker runs in shadow mode: it produces the query list, the reminder list and the visit calendar, and the coordinator compares them with the manual version. Differences are logged and fixed before anything is sent to a subject or a department.

In the second stage, the AI worker takes over one routine, such as the open query list, and the coordinator handles only the exceptions the list raises. Exceptions are the cases the system cannot resolve: a query that needs an investigator, a document that a department refuses to send, a subject who asks a clinical question. Those go to a named person with a short note on what the system found.

Only after several monitoring visits, when the team trusts the lists, should roles change. At that stage, the work of a data entry or chasing role shrinks to review. Coordinators spend the time saved on subject care and on the work that needs training. If a role's hours or duties change as a result, the notice, consultation and severance rules of the country where the site operates apply. Any change should be agreed with the employee or checked with a lawyer before it is made.

Common mistakes

• Automating the query list before the query log has a named owner for each query. The list then reminds nobody.
• Sending subject reminders without checking that each subject has agreed to that channel, and that the wording has been approved by the site.
• Letting an AI worker write clinical values into the EDC system. Every entry needs a named person who confirmed it against source.
• Building the visit calendar from a generic template instead of the protocol version in force. Windows that differ by one visit cause protocol deviations.
• Skipping the validation step for any tool that writes into trial records, then finding the gap during an inspection.
• Treating the AI worker as a substitute for the principal investigator's oversight. It tracks; it does not decide.
• Cutting a role before the team has run the new routine through at least one monitoring visit.