Twelve Months, Three Patients: The Enrollment Crisis No Algorithm Can Fix

A founder called Mark Hensley last week with a question that had no clean answer. Twelve months into his clinical trial, he had enrolled three patients, and he wanted to know whether anything could still be salvaged. Hensley, who runs the consultancy Hensley Biostats, walked through the autopsy publicly on LinkedIn: you sit with the trial documents, the site reports, and the screen-fail logs, and you trace backward from the recruitment number to the moment the trial first started telling you it would fail.

What he finds, he wrote, is almost never what sponsors expect. “The marketing is fine. The sites are doing their job.” The damage was done earlier — “somewhere in the inclusion criteria, a line that sounded conservative on paper has filtered out the patients who actually exist.” A visit schedule that asks a working mother to take a half-day off three times in a row. A screening visit that runs two hours for candidates who had forty-five minutes to give.

The reflex, when a trial stalls like that, is to spend money on the symptom — buy a recruitment platform, license a patient-matching model, add another vendor to the pile. For most of the last decade, that reflex has had a favorite product to sell.

The $3 Billion Alibi

In the same few weeks Hensley posted his autopsy, two other clinical-operations veterans arrived independently at the same uncomfortable verdict: the most expensive tools the industry has ever purchased have not moved the one number that matters.

Elias Tharakan, a clinical-technology strategist, compressed the indictment into a line that traveled: “We bought AI to fix clinical trial enrollment. Instead, we automated the excuses.” His figures are the ones the field has recited for years without flinching — more than $3 billion poured into AI recruitment platforms, and roughly 80% of trials still missing their enrollment deadlines. Three years of “AI-powered” everything, he argued, took fragmented, two-week-old EHR data and asked a model to predict the future, then treated the hallucinations as a forecast. The same models trained on historical patient populations and then disappointed everyone by reproducing the very demographics the field was trying to escape. The dashboards built to lighten the load were dropped onto already-overwhelmed site coordinators, who quietly shelved them.

His conclusion is the part worth sitting with: the bottleneck was data quality and change management the entire time, and those were the line items nobody wanted to fund.

Andrea Hartman-Gonzalez, a clinical-operations specialist, named a quieter version of the same self-deception — the industry keeps grading itself on the wrong exam. Success gets measured in Cost Per Lead, when the only figure that decides a trial’s fate is Cost Per Enrollment. A low cost-per-lead looks impressive in a status deck right up until none of those leads qualify, none randomize, and the true cost per enrolled patient climbs past the point where the study’s budget still holds together.

A pile of leads has never once enrolled a patient.

She sketched the standoff in miniature. The sponsor needs fifty patients in eight weeks; the protocol carries thirty-seven exclusion criteria; the recruitment vendor reports four thousand clicks; the coordinator reports that none of them qualify. Every party is doing exactly what it was hired to do, and the trial still slides toward the screen-fail cliff — because the eight-week target and the thirty-seven criteria were never reconciled with each other before the study opened. No campaign repairs that contradiction. The remedy is a protocol amendment, the most expensive way the industry has to admit a design mistake; research from the Tufts Center for the Study of Drug Development has for years found a large share of amendments avoidable, each costing months of delay and six figures to execute.

Failure by Design

Step back far enough and the pattern should unsettle anyone who has ever signed a recruitment-technology contract. For ten years the field has aimed its most advanced instruments at the visible end of the problem — the leads, the clicks, the matching scores. But the decisions that actually determine whether a trial enrolls were made upstream, in a protocol design meeting, before a single patient was screened.

We got very good at treating the symptom and left the disease alone.

The reason the reflex persists is that the symptom is easier to sell to. A vendor can demo a dashboard in thirty minutes; nobody can demo the discipline of cutting a protocol’s exclusion criteria from thirty-seven down to nineteen. A recruitment campaign produces a number for next week’s status deck; a hard feasibility review produces a delay the sponsor has to defend to a board that wants the trial open now. The incentives reward motion over correction at every level — and software is the most frictionless motion money can buy.

The regulators have been pointing at the upstream cause in increasingly plain language. The FDA’s 2023 draft guidance on Decentralized Clinical Trials exists precisely because the visit burden Hensley describes — the travel, the repeated half-days, the two-hour screening — is a design choice rather than a fact of nature. The Diversity Action Plan requirement that arrived with FDORA in 2022, and the FDA’s 2024 draft guidance fleshing it out, exist because no sponsor can model its way to a representative population out of historical data that was never representative to begin with. Both are official acknowledgments that enrollability is engineered into a protocol, for better or worse, long before recruitment starts.

The Trust You Can’t Buy

There is already a proof of concept for the alternative, and it did not come from a software company. Tharakan, in a second post, traced what happened after Walgreens walked into the clinical-trial business in 2022 to industry shrugs. Three years later he tallied 17 million patients educated, more than 35 biopharma partnerships, 20 compliant trial sites built inside pharmacies, and a $25 million BARDA award — built on the one asset a pharmacy already holds and a platform cannot buy: a patient’s standing trust in the person who fills their prescription. It is also the most direct answer to Hensley’s working mother, who can join a study at the pharmacy she already visits without taking three half-days off to reach it. “Who else already has the trust we’ve been trying to buy?” Tharakan asked. It is the right question, and it has nothing to do with model architecture.

So here is what has to change before the next trial posts three patients in twelve months. Feasibility has to move to the front of the process, where it belongs — every exclusion criterion interrogated for the patients it silently deletes, every visit built around the lives of the people expected to attend, every recruitment dollar judged by who randomizes rather than who clicks. The technology was never the villain in this story; pointed at clean data and a survivable protocol, the same models earn their keep, and Tharakan concedes as much when he notes the teams now seeing 7x faster patient identification got there by standardizing their data first.

None of this can be bought late.

By the time a founder is calling a biostatistician to ask what can be salvaged, the answer was written into the protocol a year earlier. The trials that fill are the ones designed to be fillable — at the inclusion criteria, at the visit calendar, at the metric chosen for the dashboard. Everything after that is just paying for the excuse.

References

1. Mark Hensley, LinkedIn — “Twelve months in, three patients enrolled”: an autopsy of a stalled trial
2. Elias Tharakan, LinkedIn — “We bought AI to fix clinical trial enrollment. Instead, we automated the excuses.”
3. Elias Tharakan, LinkedIn — How Walgreens turned standing patient trust into a distribution advantage
4. Andrea Hartman-Gonzalez, LinkedIn — Why Cost Per Enrollment, not Cost Per Lead, decides a trial’s fate
5. Andrea Hartman-Gonzalez, LinkedIn — The sponsor / CRO / site enrollment standoff