The largest randomized trial of AI-assisted cancer diagnosis found that AI cut radiologist reporting time by twenty-eight percent and reduced time to diagnosis by zero. The bottleneck was never the radiologist.
The LungIMPACT trial enrolled ninety-seven thousand chest X-rays across five NHS trusts over eighteen months. It is the largest randomized controlled trial of AI-assisted cancer diagnosis ever conducted. The AI — Qure.ai’s qXR system — flagged suspicious images and pushed them to the front of the radiologist’s worklist. The technology worked exactly as designed. Reporting time dropped from forty-seven hours to thirty-four. The radiologists saw the flagged X-rays faster, read them sooner, filed their reports more quickly.
Time from chest X-ray to cancer diagnosis: forty-four days with AI, forty-six without. The difference was not statistically significant. The primary endpoint — time from X-ray to CT scan — was fifty-three days in both arms. Five hundred and fifty-eight patients were diagnosed with lung cancer across the trial. The AI changed nothing about how quickly they learned.
The System That Didn’t Move
The lead author, Nick Woznitza, described the result precisely: “The bottleneck isn’t the reporting; it’s everything that happens next: telling the patient, the CT appointment, the clinic slot, the multidisciplinary meeting.”
The AI accelerated the fastest step in a multi-step process. The radiologist’s report was never the constraint. The constraint was the booking system that schedules CT scans, the referral pathway that routes patients to specialists, the clinic capacity that determines when someone is seen, the multidisciplinary team meeting that decides treatment. None of these moved.
This is not an AI failure. It is a systems failure — or rather, a failure to understand that “the system” extends beyond the component you optimized. The AI vendor delivered exactly what was purchased: faster prioritization of suspicious images. The NHS purchased a local optimization and received a local optimization. The error was expecting local improvement to propagate.
The Theory of Constraints in Practice
Eliyahu Goldratt formalized this in 1984. Any improvement not at the constraint is an illusion. Speeding up a step that feeds into a bottleneck just builds inventory in front of the bottleneck faster. The flagged X-rays arrived at the CT booking queue sooner — and waited there the same amount of time as every other referral.
The two-day difference in diagnosis time (forty-four versus forty-six days) is exactly what the theory predicts. It is the reporting time savings (thirteen hours) expressed as a fraction of the total pathway duration (roughly fifteen hundred hours). The signal vanishes into noise because the accelerated step was less than three percent of the total elapsed time.
Professor David Baldwin, chair of NHS England’s Clinical Expert Group for Lung Cancer, stated the implication directly: “If we want AI to make a real difference to lung cancer outcomes, we need to redesign how the NHS responds when the AI raises an alert, and that means better coordination, more resource, and a genuine commitment to same-day action.”
The Pattern
This journal has tracked the same structural failure across industries. Fifty-six percent of CEOs report zero financial returns from AI. Over forty percent of agentic AI projects will be canceled by 2027. The enterprise software market lost two trillion dollars in thirty days. In each case, the technology performed. The system around it did not change.
The LungIMPACT trial is the cleanest evidence yet because it is a randomized controlled trial — the gold standard of causal inference — applied to a real deployment at scale. It eliminates the usual confounders. The AI was not misconfigured. The sample was not too small. The implementation was not botched. The result is structural: optimizing one node in a bottlenecked system produces zero system-level improvement.
The paper’s own conclusion is unequivocal: “CXR AI deployments should not include worklist prioritization in this context.” Not because the AI failed to prioritize. Because prioritization was never the problem.
What Changes the Outcome
The bottleneck in lung cancer diagnosis is not information. It is capacity and coordination. A flagged X-ray that arrives at the front of the reporting queue thirteen hours earlier still enters the same CT booking pipeline, the same referral pathway, the same weekly MDT meeting schedule. Same-day action — Baldwin’s phrase — requires restructuring the entire downstream pathway: pre-authorized CT slots for AI-flagged cases, direct specialist referral without GP intermediation, dedicated clinic capacity for urgent findings.
This is expensive. It requires institutional change, not software deployment. And it confronts the uncomfortable reality that most AI procurement is premised on the opposite assumption: that technology can improve outcomes without changing the organization that deploys it.
The two trillion dollars currently being spent on AI infrastructure is largely a bet on the node, not the system. LungIMPACT is the first rigorous, large-scale evidence of what happens when that bet resolves. The AI works. The system-level outcome does not change. The bottleneck was never where the vendors said it was.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)