In April 2026, the FDA issued a Request for Information on a proposed pilot program for AI-enabled optimization of early-phase clinical trials. It’s the kind of signal the industry has been waiting for – a formal Agency invitation to define what trustworthy, high-value AI in trial conduct looks like.
CRIO responded. And because our perspective comes from an unusual vantage point – the site-technology layer where source data is created – we think it’s worth sharing what we said.
Every AI system operating on clinical trial data is downstream of source data capture. The AI conversation cannot be separated from the data-capture conversation.
Start with the Use Cases That Actually Work
The temptation in any AI pilot is to lead with the most ambitious applications – dose escalation modeling, biomarker-based stratification, population-level fairness assessment. We understand the appeal. But in our response to the FDA, we argued for a different sequencing.
The highest-value, lowest-risk early use cases are workflow-oriented: participant pre-screening and eligibility matching, source data quality monitoring, adverse event detection and triage, and protocol deviation surveillance. These use cases:
- Produce measurable value within months, not years
- Don’t require disease-specific model validation
- Lend themselves to clean before/after comparisons against historical baselines
- Exercise the trustworthy-AI principles the pilot is designed to evaluate: validity, transparency, and human oversight
Higher-stakes use cases like dose escalation have a place in the pilot – but only where AI is decision-support, with human adjudication preserved. That’s not a conservative position. It’s how you build the evidence base that earns expanded AI authority over time.
Data Integrity at the Source Is Non-Negotiable
This is the argument we made most forcefully, because it’s the one most often skipped in the AI-in-trials conversation.
CRIO operates the data-capture layer at more than 2,900 research sites in 33 countries. What we see repeatedly is that AI tools fail – not because the models are weak, but because the underlying data they’re running on is paper-based, late, or transcribed across multiple systems. No AI model can compensate for that.
A recent Tufts study found that a significant share of source data at investigative sites is still captured on paper. That’s the real constraint on AI readiness – not the models.
We told the FDA that the pilot should be designed with this interdependence in mind. An AI pilot layered on top of poor source data capture will produce findings that understate AI’s potential and overstate its failure rate. Getting the capture foundation right first isn’t a detour from AI adoption – it’s the precondition for it.
We have the data to back this up. In one documented case, Benchmark Research saw a 40% reduction in protocol deviations following deployment of CRIO’s eSource platform across nine sites – before any AI was introduced. That’s the floor AI gets to build on.
Don’t Gate Participation on AI Maturity
One of the more counterintuitive positions we took: the FDA should not require high AI maturity as a condition of pilot participation.
If the pilot only includes sponsors and sites that are already AI-sophisticated, the findings will be interesting but narrow. The industry needs evidence about how AI performs across the full spectrum of participants – including small biotech sponsors running first-in-human studies, community research sites, and academic medical centers.
The most realistic on-ramp for less mature participants is vendor-supplied AI delivered through existing site-technology platforms – eSource, CTMS, eConsent – rather than each participant building or procuring AI capability from scratch. This approach also has an analytic advantage: when the underlying capture foundation is constant across participants, observed differences are more cleanly attributable to the AI component itself.
We proposed a tiered participation structure: a foundational tier for participants using vendor-supplied AI through existing platforms, and an advanced tier for those developing custom AI components, with shared evaluation metrics across both.
Measure What People Actually Do, Not What They Say
The last theme running through our response is about evaluation design. Survey-reported trust, perceived value, and projected scalability are weak evidence. They’re subject to novelty effects, social desirability bias, and the simple fact that people’s stated beliefs about AI rarely match their revealed behavior.
Stronger evidence comes from the platform itself:
- Override rates and outcomes: When AI suggests, how often do users override it? And when they do, are they right?
- Feature utilization rates: Are users using the AI features, or clicking past them?
- Time-on-task: Does AI make coordinators faster, or just add a step to review?
- Post-pilot continuation: After the pilot ends and the novelty fades, do sites keep using the tools?
- Source-to-EDC concordance: Is the data that enters the EDC more accurate when AI-assisted capture is in place?
We also flagged something we observed during CRIO’s own AI Study Design Builder rollout: trust calibration takes weeks to months of regular use. Short-exposure trust measurements – a survey at week two – substantially understate the trust that emerges with sustained use. Measurement windows matter.
Why This Matters Beyond the Pilot
The FDA’s RFI is a starting point, not a finish line. The pilot it describes will shape how regulators, sponsors, CROs, and sites think about AI in trial conduct for the next decade. The frameworks that get established – for data integrity, governance, evaluation metrics, and participant selection – will outlast any specific AI model or technology.
CRIO remains what it is today – the site-technology foundation on which AI-driven trial conduct is built. We’re not a dose-selection algorithm or a biomarker discovery engine. We’re the system that captures the data those tools run on, at 2,900+ sites, in real time, under full regulatory compliance.
That’s why we responded to this RFI. And it’s why we’ll keep showing up in these conversations – at SCDM, at SCRS, at DIA, at SCOPE, and directly with the Agency – wherever the rules of the road for AI in clinical research are being written.
Contact us or subscribe to our AI newsletter to learn more about CRIO’s approach to AI.