Testing an eCOA system is not like testing most software. The logic you need to verify (compliance windows that open on day 14, adaptive triggers that fire at week 12, missed-visit alerts at day 47) plays out across months of a live study. The UAT window available to test all of it is typically a few days. That gap has a name: the temporal conundrum. And it sits at the center of nearly every eCOA validation headache.
On May 21, 2026, Lisa Charlton, Chief Product Officer at Castor, moderated a session with Connor Ladly Fredeen, Director of Delivery Engineering at Castor, and Gauri Nagrani, Co-Founder at Safira Clinical Research, covering why this problem is harder than it looks and what AI-assisted testing can realistically do about it.
What the session covered
Gauri opened with the regulatory frame. The shift from Computer System Validation (CSV) to Computer Software Assurance (CSA) (now aligned with ICH E6(R3)) means directing validation effort toward scenarios with the greatest patient safety and data integrity risk. For eCOA solutions with complex time-dependent logic, those high-risk scenarios are almost always the ones farthest along in the study timeline. And reaching them manually requires weeks of staged data entry before testing can even begin.
Connor described three AI-assisted approaches Castor uses to close that gap. The first is AI backdated data entry: the system programmatically stages chronologically consistent patient histories, allowing testers to jump directly to critical decision points such as day 90 eligibility reviews or missed-visit alert triggers. Every action remains traceable within the validated CDMS.
“If an AI system is performing an action, you have a human-auditable trace of what was done, as well as screenshots and other evidence produced by the system each step along the way.”
— Connor Ladly Fredeen, Director of Delivery Engineering, Castor
The second approach is exploratory AI testing: autonomous agents run browser-based scenarios that human testers would not think to design, surfacing edge cases through deliberate creative variation. The nondeterministic nature of AI (often cited as a concern) is in this context a feature. The third is AI-driven timeline compression: a separate scheduling layer outside the validated core system handles complex scheduling logic, compressing months of simulated study time to days without touching the CDMS validation state.
A firm thread throughout the session was regulatory independence. Gauri was direct: sponsors and CROs must own their UAT test scripts. Vendors cannot write UAT plans for their own systems. AI-assisted execution is acceptable. AI-owned strategy is not.
The discussion of mid-study change controls produced one of the session’s most memorable analogies. Gauri described what it feels like to insert a change into an existing multi-system tech stack — not the standard Jenga move of pulling from below and stacking on top, but forcing a new block into the middle of a structure that’s already standing:
“It’s like this Jenga where you’re trying to put a block in between somewhere and the whole thing can either stand up for a while or just totally is destroyed.”
— Gauri Nagrani, Co-Founder, Safira Clinical Research, on the instability introduced by mid-study amendments across a connected tech stack
Any change to an eCOA system requires assessing impact across the full tech stack, not just the modified module. Regression testing across all connected systems is the safeguard. And AI-assisted regression is where the next major time savings will come, as Gauri noted, because it is the area where manual testing most often misses coverage.
What’s in the full recording
The on-demand recording goes deeper on several fronts. Connor walks through how keeping AI-driven timeline manipulation in a separate scheduling layer outside the validated electronic data capture system, with enough technical detail to inform your own validation architecture decisions (29:00). The Q&A covers a question every eCOA team using AI-assisted testing will face: what safeguards ensure sponsor protocol data stays protected when it enters an AI environment (52:40). Connor addresses whether AI test output can be considered repeatable given its nondeterministic nature, and his answer reframes the question entirely (54:25). And for teams managing mid-study amendments across a multi-system tech stack, Gauri walks through the full risk-scoping framework she applies in practice (49:50).
Frequently asked questions
What is the temporal conundrum in eCOA UAT?
eCOA systems contain time-dependent logic including compliance windows, adaptive triggers, eligibility rules, and missed-visit alerts. This logic unfolds across weeks or months of a live study. UAT windows are typically a few days. Reaching high-risk scenarios such as day 90 eligibility checks or missed-visit alert triggers can require weeks of manual data staging before testing even begins. The gap between the study’s required time horizon and the available testing window is the temporal conundrum.
Can AI be used to write eCOA UAT test scripts?
AI can assist in generating test case suggestions and exploring edge cases, but sponsors and CROs must own the final test scripts. Gauri addressed this directly in the session: regulatory independence requirements mean vendors cannot write UAT plans for their own systems. The role of AI is to support execution and expand coverage. The strategy, the test case ownership, and the sign-off remain with the sponsor’s team.
How does AI-assisted UAT data maintain ALCOA+ compliance in eCOA?
Connor explained that AI actions in Castor’s system produce a complete human-auditable trace. Every step is logged with screenshots and documented evidence within the validated CDMS. The resulting study state is equivalent to what a human tester would produce, with full attributability and traceability. ALCOA+ compliance is preserved because the audit trail is generated by the validated system rather than the AI layer operating alongside it.
