66.9%, 85%, 18 Months: What Clinical AI Actually Delivers When the Data Is Right

Key takeaways:

• The results organizations are achieving when AI is grounded in high-quality RWD aren’t incremental improvements on the status quo. They represent a fundamentally different level of performance. 
• The numbers below span data recovery, patient identification, protocol design, recruitment, and disease detection. They point to the same underlying variable: data quality is what separates these outcomes from the average. 
• These figures aren’t benchmarks to aspire to eventually. Organizations are achieving them now, with infrastructure that exists today. 

The first three posts in this blog series made the argument in words. The framework post, Why Two Clinical Teams Can Run the Same AI and Get Completely Different Results, explained why data quality, not algorithm sophistication, determines AI performance. The smoking data post, Three Out of Four Patients Are Missing Smoking Data. That’s Not the Real Problem, showed what a data quality failure looks like at the level of a single clinical variable. The evaluative post, A Diagnostic for Clinical AI Teams Whose Results Aren’t Matching the Promise, turned the framework into a checklist for evaluating where a team’s current foundation might be creating drag. 

This post does something different. It steps back from the argument and lets a set of numbers carry the weight. 

Each one comes from an organization that has paired artificial intelligence (AI) with comprehensive, current, clinically validated real-world data (RWD). Each is a real outcome, not a projection. And each reflects the same underlying pattern: when the data foundation is right, clinical AI performance isn’t marginally better than the average. It’s in a different category. 

From 22.3% to 66.9%

Increase in smoking status data coverage when TriNetX’s AI-driven extraction from clinical notes is applied alongside structured coded data. In most structured datasets, smoking status is missing for more than three out of four patients, even though it affects drug metabolism, disease progression, treatment response, and trial eligibility across nearly every therapeutic area. Recovering it at scale, with the validation and traceability required for regulatory use, is what it looks like to systematically address a data quality problem that most AI systems inherit without realizing it. (The full story is in the second post in this series, Three Out of Four Patients Are Missing Smoking Data. That’s Not the Real Problem) 

85%

Accuracy rate for identifying inflammatory bowel disease (IBD) patients at high risk of a disease flare, using a machine learning model built on real-time clinical data. The traditional approach, clinical staff identifying at-risk patients during appointments, had a success rate of 33% for the same task. The difference isn’t effort. It’s data currency. 

18 months

How far ahead of diagnosis an AI model can now predict pancreatic cancer risk, using RWD from 35,000 patients and 1.5 million controls. The model identified 87 predictive features. A prediction window that long doesn’t exist without the longitudinal data depth to support it, and it has direct implications for how and when trials can recruit patients with early disease. 

189%

Increase in the eligible patient pool for a UK–Germany COPD trial when TriNetX applied AI-optimized eligibility criteria to richer, better-structured RWD. Traditional coded criteria identified 230,750 eligible patients. The AI-optimized approach identified 666,200. The science didn’t change. The inclusion criteria didn’t loosen. The data were comprehensive and rigorous enough to find patients that coded criteria alone had missed. 

65% and 30–50%

Improvement in enrollment rates from AI-powered patient recruitment tools, and reduction in trial timelines when AI is integrated into clinical operations with the right data foundation. Both figures come from research published in the International Journal of Medical Informatics (February 2026). Combined with the 40% reduction in costs reported in the same paper, they represent a shift in the economics of clinical development, not a marginal efficiency gain. 

The Pattern Behind the Numbers

A single pattern runs through every one of these results. None of them were achieved by finding a better algorithm. Each was achieved by building AI on data that were comprehensive enough to see the relevant patient population clearly, current enough to reflect where patients actually are in their disease trajectory, and rigorous enough to produce outputs that hold up under scrutiny. 

That’s not a technology story. It’s a data foundation story. And the organizations getting these results have treated it that way from the start. 

There are more numbers where these came from. An AI model that predicts hepatocellular carcinoma risk 6 to 36 months before diagnosis, with strong performance maintained across U.S., Latin American, and Asia-Pacific cohorts. An early-warning system that flagged 103 potential lupus trial candidates before they met formal eligibility. A pharma organization that embedded these capabilities directly into its protocol development workflow and moved from insights that took days or weeks to answers available during the authoring session. Each is detailed, with the underlying data architecture that made it possible, in the full guide. 

About Joshua Hartman 

As Senior Director, Clinical Study Feasibility & Analytics, Josh leads a team of clinical experts in utilizing real-world data to enhance clinical trial design, feasibility, and site identification. His work reflects a deep commitment to advancing public health through data-driven insights.