An agent that reads EDC data and drafts monitoring findings. First-pass catch rate went from ~60% to ~92% over six iterations; weekly review time from 8–12 hours to 2–3; override rate settled at 11% — and stayed there.
001-001: “Focal hypersensitivity / skin tenderness, L hip” · 001-006: headache narrative with the ConMed written into the term.
May be appropriate (untreated AEs), but the EDC will auto-query the treatment field for each one unless the CRF says otherwise.
Identical verbatim on two record IDs two days apart. Likely duplicate entry — flagged for the site, not auto-closed.
“Took x1 tab neurofen on Wed afternoon…” entered as the drug name. Will fail WHODrug coding and EDC validation.
Only 2 records closed. Every other subject has 100% of AE records unresolved — a site workflow issue, not a data one.
Misses over noise. I tuned the agent to catch more even when that meant more false positives for a monitor to dismiss — a miss is a safety event, a dismissal is four seconds. Decision support, not decisions.
— The call I madeRemote data review is where a monitor earns their keep: reading what sites entered, spotting the adverse event that was not followed up, the deviation nobody logged, the lab value that should have triggered a query. It is also mostly scanning records that are fine, for 8–12 hours a week per monitor.
Findings lived in people's heads and spreadsheets. Coverage depended on who was reviewing and how tired they were, and nothing about the process got better over time.
The agent reads the EDC through the analytics warehouse and drafts findings with evidence — adverse events, deviations, query status, lab anomalies, site performance — into the review workflow monitors already use. It was iterated six times against a labelled set of missed and over-flagged findings, and tuned deliberately toward catching more. Override logging became a first-class metric: the number the team actually watches.
Every change to the prompt and rubric was judged on one number the monitors already trusted.
First-pass catch rate rose from ~60% to ~92%. Weekly review time fell from 8–12 hours to 2–3. The override rate settled at 11% and stayed there — monitors disagree with the agent often enough to prove they are reading, and rarely enough to prove it is useful.
The moment that set the pattern for every regulated AI feature since: an auto-coding step hit a regulatory edge case, and within 24 hours it was pulled from autonomous to human-in-the-loop. Autonomy is something a system graduates into with evidence.
The override dashboard was the artifact — leadership, clinical ops and engineering decided with it instead of with opinions. Six iterations in six weeks, each one driven by the monitors' own labelled misses.
— How it came together