The 90% Automation Promise Means Nothing to a Coordinator Running 12 Systems

Picture a screening visit at a busy research site. A coordinator sits across from a patient, laptop open, scanner nearby, eSource tool loading. She is maintaining eye contact, keeping the patient calm and engaged, while simultaneously logging into system four of twelve. John Oncea, Chief Editor at Clinical Tech Leader, reported this scene after conversations with Nick Spittal and Raghu Punnamraju of Velocity Clinical Research: 29 discrete steps to complete a single screening, 17 of them system-facing, only 12 involving the patient directly. The workaround had become so routine it barely registered as a problem. It should.

That same week, Ryan Flinn published an interview with Medable co-founder and CEO Michelle Longmire, who articulated her goal plainly: 90% of clinical trial operations automated by 2030. One benchmark she cited: site visit paperwork that currently consumes eight hours of staff time reduced to 30 minutes of AI-assisted review. The vision is serious. The timeline is aggressive. And it exists in an entirely different world than the one that coordinator is navigating.

This is the central dysfunction of AI in clinical trials right now. The boardroom and the site floor are not having the same conversation, and the gap between them is widening with every vendor pitch that leads with the word “automation.”

What Recruitment AI Actually Runs Into

The EHR-matching hypothesis was elegant: clinical data already exists, AI can read it, match patients to eligibility criteria, and compress enrollment timelines from years to months. For broad-population therapies, Irina Gontschar, MD, CEO of Potter Research Solutions, acknowledges the model can work well. But she draws a hard line at oncology, hematology, and rare diseases, where a diagnosis code in an EHR does not confer eligibility. A rare mutation has to be contextualized within a patient’s full clinical history. That requires a physician or a highly qualified research professional, not a matching algorithm.

The deeper problem Gontschar identifies is that early AI recruitment tools solved for the wrong bottleneck. Finding a potential participant was never the primary source of delay. The real friction accumulates after identification: EHR review, clinical history reconstruction, records collection, additional testing, eligibility confirmation, and all the communication in between. These are the repetitive, time-intensive tasks that qualified humans are currently absorbing, and they are exactly where AI with appropriate EHR access could provide genuine relief without displacing clinical judgment.

There is also a biological reality that no algorithm overrides. A therapy requiring six months of investigational drug administration followed by two years of overall survival follow-up cannot be compressed regardless of how quickly the right patient is found. Faster recruitment cannot eliminate protocol-defined time. The sponsors who understood this distinction early are the ones building AI tools calibrated to the actual bottleneck, not the visible one.

Which raises the question Gontschar leaves open: if the real opportunity is in post-identification workflow automation, why has so much capital and attention landed on the matching problem? Partly because matching is legible. It is demonstrable in a sales deck. A reduced time-to-enrollment metric looks clean on a slide. The messier work of EHR reconstruction and records collection is harder to quantify and less flattering to investors. The incentive structure, in other words, has been pointing AI development in the wrong direction.

The Integration Problem Nobody Wants to Invoice For

Jenn Pages, CCRP, CEO and Founder at her site consultancy, puts the operational reality without decoration: “I do not care how advanced your clinical trial technology is if it makes the site’s job harder.” Her inventory of systems a coordinator is already required to manage includes a CTMS platform, EDC, ePRO tools, a randomization system, lab portals, imaging vendors, a safety database, an eTMF system, and sponsor-specific platforms. At some point, she writes, a “technology-enabled trial” starts to sound like “please remember 11 passwords.”

The Velocity Research data Oncea surfaced makes this concrete. Seventeen of 29 screening steps were system-facing. A coordinator counting logins, uploads, transcriptions, and cross-checks before she can spend meaningful time with the patient. That is a technology stack functioning as a data-entry burden rather than a clinical support system. The coordinator becomes the integration layer, manually moving information between platforms that were never designed to talk to each other.

AstraZeneca appears to be reading the same signals at the institutional level. Danielle Bitterman, VP of AI for Clinical Development, posted an open role this week for a senior leader to build and lead a Clinical AI Research team covering Phase I through III, with an explicit focus on foundation models, trial simulation, causal modeling, and agentic systems. The framing is notable: AstraZeneca is not looking for someone to deploy existing AI tools. It is investing in the science and methods of AI for clinical development as a distinct research discipline. That signals the company recognizes current tools are insufficient for the problem at hand.

Benjamin Vandendriessche, CEO of the Digital Medicine Society, returned from DPHARM and CTTI meetings with a diagnosis that cuts through the enthusiasm: regulatory momentum is real, including FDA’s Digitally-derived Measures white paper and Operation Trialblazer, but adoption of the underlying tech stack remains slow. His framing deserves attention: “Access to technology is not the main blocker anymore.” The blocker is deployment intelligence, which means understanding incentive structures clearly enough to know where a tool will actually reduce friction versus where it will simply relocate it. Per Vandendriessche’s post, more than 80% of health systems have no clear AI governance system in place even as AI-based systems are already rolling out. Clinical trials are not exempt from that gap.

Pages and Oncea are both pointing at the same structural failure from different angles. Sponsors and vendors have optimized for feature sets and demo impressions rather than workflow integration. The coordinator who spent 17 of 29 screening steps inside disconnected systems is not a technology problem. She is an incentive problem. Nobody in the contracting chain is accountable for her total system load. The sponsor pays for the EDC. A separate vendor provides the ePRO. The imaging vendor runs its own portal. Nobody owns the seam between them, so the coordinator fills it with manual effort, every study, every site, every day.

Where the Pressure Has to Land

Medable’s 90%-by-2030 target is a useful north star precisely because it forces a binary question: if nearly all trial operations are to be automated within five years, which operations are on the list, and who is accountable for the integration that makes automation possible? An eight-hour site visit documentation process reduced to 30 minutes of AI review is achievable if the underlying systems share data reliably. It is not achievable if the coordinator is still manually transcribing between platforms to create the input the AI is supposed to review.

Gontschar’s argument points toward the model that actually has leverage: AI paired with EHR access, handling the repetitive post-identification workflow, while experienced clinical professionals retain judgment over eligibility determination. That is not a vision of replacement. It is a vision of appropriate task allocation, which is a harder organizational problem than building the algorithm. It requires sponsors, CROs, and technology vendors to agree on data standards, access permissions, and accountability structures before the AI layer can deliver anything meaningful at scale.

Pages has already drawn the evaluation framework sponsors should be using: does it reduce duplicate entry, save coordinator time, make patient visits easier to manage, and communicate with systems already in use? If the answer to any of those is no, the technology is a cost center wearing an innovation label. The standard, as she puts it, is not how impressive the demo looks. The standard is whether it makes the study easier to run.

AstraZeneca building a dedicated clinical AI research function suggests at least one large sponsor is taking the methods problem seriously enough to fund it internally. The question is whether the rest of the industry waits for that work to surface in publications and partnerships, or starts auditing their own coordinator workflows now, before the next platform lands on an already overloaded site stack.

The coordinator maintaining eye contact with a patient while logging into system eight of twelve is not waiting for a better demo. She is waiting for someone with budget authority to count her steps and decide that 29 is unacceptable.

References

1. Gontschar, I. (2026). When AI Meets the Real Complexity of Patient Recruitment. LinkedIn.
2. Pages, J. (2026). I do not care how advanced your clinical trial technology is. LinkedIn.
3. Oncea, J. (2026). Clinical Trial Sites Don’t Need More Tech. LinkedIn.
4. Flinn, R. (2026). Everyone’s talking about AI in drug discovery. LinkedIn.
5. Bitterman, D. (2026). The frontier of AI for drug development. LinkedIn.
6. Vandendriessche, B. (2026). Back from DPHARM: Disruptive Innovations to Modernize Clinical Trials. LinkedIn.