AI in clinical research operations means the software and AI models that a contract research organisation (CRO) builds to do its trial work with far fewer people. This work includes monitoring hospitals, cleaning the trial's data and writing reports. Until recently, almost all of it was done by large teams of people, many of them in India.
Here is why it matters to a CRO. A CRO earns money by running trials for drug companies. Its main cost is the salaries of its monitors, data managers, statisticians and medical writers. If software can do part of their work, the CRO can run the same trial at a lower cost, or run more trials with the same staff.
Consider a CRO's data team in Bengaluru that is cleaning the data of a trial at 120 hospitals. Every week, the hospitals enter about 40,000 new values into electronic data capture (EDC). In the past, data managers read lists of possible errors and wrote each query, a question to the hospital, by hand. Now an AI model reads the week's new values and compares them with the patient's earlier visits. It flags 600 values that look wrong. For each one, it drafts a query, such as "Week 8 weight is 12 kg lower than week 6. Please confirm." A data manager reviews the 600 drafts. She approves 450, edits 90 and rejects 60 that are not real problems. Only the approved and edited queries, 540 in all, go to the hospitals.
The model drafts and the person decides. This way of working is called human in the loop. It matters here because a wrong query wastes a hospital's time, and a missed error can affect the trial's result. An AI model can also state something false with confidence, as explained in AI hallucination.
Monitoring changes in a similar way. Instead of checking every value at every hospital, the software scores each hospital for risk. A hospital where every blood pressure reading ends in zero, or where patients report far fewer side effects than elsewhere, gets a high score. The monitors then visit the high-risk hospitals first.
Software that carries out a sequence of office tasks by itself is called agentic automation. In trials, every step the software takes must still be recorded under GxP rules, so that the regulator can see what was done by the model and what was done by a person. Regulators are still writing their rules for AI in trials.