AI Is Now Running Clinical Trials. The Question Is Whether Anyone Is Watching.

Miriam Gonzalez is 35 years old, has metastatic breast cancer, and built her own AI research team. She named it Polaris. It reads studies, searches global trial registries, converts her scans into 3D bone maps, and prepares clinical briefings for her physicians. In six months, it helped surface a biopsy opportunity in Zurich and brought Dana-Farber into her case. Her doctors make the decisions. Polaris does the work that no coordinator, no patient navigator, and no oncology team operating at human speed could do at scale. The bottleneck now, as she describes it with candid frustration, is compute capacity. Not biology. Not bureaucracy. Cloud compute limits.

Read that twice. A patient has outpaced the clinical trial system by building a better access layer herself, and her limiting reagent is API throughput. That tells you more about where the industry stands than any conference panel or vendor white paper ever could.

This convergence of signals arriving simultaneously on LinkedIn this week is not a trend story. It is a status report. AI has crossed the threshold from background analytics into active operations: finding patients, drafting protocols, compressing timelines, and now, in ARPA-H’s SURPASS program, threatening to dissolve the phase structure of clinical development entirely. The question sponsors, CROs, and regulators must now answer is not whether AI belongs in trials. It is whether the infrastructure, oversight, and validation frameworks have kept pace with the ambition.

The Operator Has Entered the Building

PhaseV’s AI Conductor is generating what Raviv Pryluk describes as pharma executives saying they actually want to write a protocol. That reaction deserves a moment of attention. Protocol development is, in most organizations, the most adversarial internal process in clinical development: scientific, regulatory, biostatistical, and operational teams negotiating over every endpoint, every visit, every inclusion criterion, producing documents that routinely exceed 150 pages before the first site is activated. If an AI platform is collapsing that process from clinical intent through SAP, schedule of assessments, and CSR in a way that generates genuine enthusiasm, something structurally significant is happening to the labor model of drug development.

Paradigm Health, backed by DFJ Growth, is pursuing the same terrain from a different angle: making trials faster, cheaper, and more accessible through AI, per Justin Kao‘s account of the company’s trajectory over the past 18 months. The pairing of Kent Thoelke, a clinical development veteran, with Jonathan Hirsch, a health data architect, is itself a signal: the organizations making credible progress here are combining operational domain knowledge with infrastructure-grade technical depth. Hype-cycle entrants rarely have both.

Then there is SURPASS. Richard Graham, writing from a biopharma executive perspective, frames ARPA-H’s program with careful optimism and useful skepticism. SURPASS proposes collapsing the multi-phase, multi-stop clinical development model into a seamless, phaseless framework, carrying a compound from first-in-human through confirmatory evaluation within a single broader trial structure. The operational demands that come with that vision are extraordinary. Graham highlights ARPA-H’s stated target of one week for data cleaning and construction of analysis-ready datasets. One week. Current industry norms for database lock in a late-phase trial run to months. The gap between those two numbers is not a process improvement problem. It is a civilizational lift.

The counterintuitive read on SURPASS is this: most observers will position it as a regulatory acceleration story. It is actually a data quality story. Graham makes this explicit, citing Andy Studna’s Applied Clinical Trials analysis, which connects faster IND pathways and AI-enabled decision support back to a foundational constraint: none of it works if trial data are still captured late, monitored retrospectively, and cleaned up after problems occur. The phaseless model Graham finds compelling depends entirely on quality moving upstream, with discrepancies identified at the site level in real time rather than discovered during SDV weeks later. SURPASS does not assume AI solves this. It demands that the execution layer solve it first.

Speed as a Metric, Oversight as an Afterthought

Joe Mullings is making a different, sharper argument. His post frames the Director and VP of Clinical seat as a sales role that has been misclassified. A complex late-stage Class III device trial can burn several million dollars a month. Every month of drag is another block of capital incinerated. His prescription: bring in commercial DNA, give clinical leaders enrollment quotas, and pay them to drive toward a finish line rather than manage a status update. Pull 30 to 50 percent out of a timeline, he argues, and you have saved tens of millions.

The math is correct. The framing is incomplete. Clinical trials are not sales funnels. The behaviors that produce quota attainment in a commercial organization, aggressive targeting, pipeline acceleration, objection handling, do not translate cleanly into an environment where protocol deviations carry regulatory consequences and premature database lock can corrupt a primary endpoint. The talent cross-pollination Mullings describes could generate real velocity, but velocity without the GCP scaffolding to support it produces exactly the data integrity failures that Graham’s upstream quality argument is designed to prevent. These two posts are in conversation with each other, even if their authors do not know it.

Jarrod Reilly‘s post on digital twins sits at the furthest horizon of this week’s chatter, gesturing at virtual clinical trials that simulate treatment responses before clinical intervention, predict adverse reactions, and personalize drug selection at the individual patient level. Reilly himself includes the necessary caveat: “Clinical validation, data quality, and patient safety remain essential before these technologies can guide real-world decisions.” That sentence is doing significant load-bearing work. Digital twin models are only as good as the physiological and pharmacological data they are trained on, and that data comes from the same trial ecosystem that currently takes months to clean a database. The simulation layer cannot outrun the data substrate beneath it.

This is the structural tension the industry is not yet confronting directly. The AI capabilities being deployed at the operator level, patient-facing agents, protocol drafting platforms, phaseless trial frameworks, predictive digital twins, are all drawing on the same underlying data infrastructure that was built for a slower, more retrospective model of clinical development. Miriam Gonzalez’s Polaris system is remarkable precisely because she is a software engineer who could build a high-quality retrieval and synthesis layer on top of public data. The average patient cannot. And the average trial site is still faxing lab results.

What Validation Looks Like From Here

The FDA’s 2023 Framework for the Use of Digital Health Technologies in Clinical Investigations and the agency’s Real-World Evidence Program guidance both establish that AI-assisted tools used in trial conduct require prospective validation plans, clearly documented intended use, and performance monitoring over time. Neither framework was written with patient-constructed AI agents or AI protocol drafting platforms in mind. The regulatory posture is understandable, written for a world where sponsors deployed validated eClinical technology from known vendors through formal procurement processes. That world is not the world Miriam Gonzalez is operating in, and it is not the world PhaseV or Paradigm Health are building toward.

The validation gap here is real and specific. When an AI Conductor drafts a protocol section, who audits the clinical logic? When a digital twin predicts an adverse reaction profile, what is the reference standard against which that prediction was validated, in what population, and over what time horizon? When SURPASS compresses data cleaning to one week, what quality assurance steps are being parallelized rather than eliminated, and who bears regulatory accountability when parallelization introduces error? These are not rhetorical objections to progress. They are the operational questions that separate transformation from acceleration theater.

Graham’s skepticism that any single initiative can solve the broader challenges facing U.S. clinical development is the most valuable sentence in this week’s chatter. SURPASS, PhaseV, Paradigm Health, Polaris: each represents a genuine capability advance in a specific layer of the trial stack. None of them, individually or collectively, resolves the accountability question at the center of GCP-regulated research: when AI is the operator, who signs the form?

Miriam Gonzalez’s doctors make the decisions. Polaris does the work. That division of labor is the only version of this future that holds. The industry’s job now is to build the oversight architecture that makes it enforceable at scale, before the platforms that are already live make that question academic.

References

1. Miriam Gonzalez, LinkedIn post on Polaris AI agent and metastatic breast cancer trial search
2. Joe Mullings, LinkedIn post on commercial DNA entering clinical trial leadership
3. Richard Graham, LinkedIn post on ARPA-H SURPASS program and Operation TrialBlazer
4. Justin Kao, LinkedIn post on Paradigm Health and DFJ Growth Founder Story
5. Raviv Pryluk, LinkedIn post on PhaseV AI Conductor and protocol development
6. Jarrod Reilly, LinkedIn post on digital twins and virtual clinical trials