Your Scan Is Done, Your Report Is Not

In 2024, 976,000 scan reports in England took longer than 28 days to be written. The Royal College of Radiologists described this as the worst performance on record. You are not inside those numbers. You are underneath them, waiting for a voice on the phone or a letter through the door telling you what the machine saw three weeks ago when you lay still on the table.

The AI triage platform that flagged your CT as routine is running in production at half the trusts in England now. More than 4 million patients have received a faster diagnosis or all-clear for lung cancer by improving patient care routes. Early data shows the technology helps radiologists analyse scans in an average of just 4 days, compared to 8 days for the most complex cases previously. You do not know if you were in the pool the algorithm sorted, or whether your file is still sitting in a different queue, waiting for a pair of human eyes that will not arrive for another week.

Behind the operational standard nobody meets

As of early 2026, approximately 1.8 million people in England are waiting for a diagnostic test, with around 447,000 waiting beyond the six-week standard. The NHS has not met that standard since November 2013. The target is administrative. The wait is yours.

The diagnostic imaging backlog is the product of a structural workforce gap that has widened for years, with demand for CT and MRI scans growing at more than twice the rate of workforce growth and a 29% national shortfall of consultant radiologists in England projected to reach 39% by 2029 without intervention. AI enters this gap not as a choice but as the only mechanism that keeps the pipeline moving. Laboratories are no longer evaluating AI as a future capability, they are deploying it now to preserve diagnostic throughput.

The trust that performed your scan may have contracted with Aidoc, whose CT triage platform received CMS reimbursement approval in August 2026. It may have installed one of the AI-powered X-ray tools the government committed $20 million to roll out by 2029. Or it may be using nothing at all, because 57% of AI diagnostic studies in a June 2026 scoping review did not report a deployment platform, suggesting most remain at the algorithm development stage without a defined path to clinical implementation. You do not know which, and the consent form you signed before the procedure did not name the software.

Sensitivity, specificity, and what goes unsaid

A systematic review of AI in digital pathology reported a mean sensitivity of 96.3% and mean specificity of 93.3%. 99% of studies identified for inclusion had at least one area at high or unclear risk of bias or applicability concerns. The performance threshold a radiologist would require to trust an AI system has not been explicitly addressed in the clinical literature. Studies show that algorithm-assisted pathologists decreased the human error rate by almost 85%, with sensitivity of 91% versus 83% for pathologist or algorithm alone. But the median discrepancy rate between primary pathology reports and secondary review was 18.3%, with a rate of major discrepancies of 5.9%, and IHC tests have had error rates up to 10 times higher than other clinical tests.

The model that triaged your scan may have been trained on datasets you will never see. California's AB 2013, which took effect 1 January 2026, requires developers of publicly available generative AI systems to disclose training data on their websites, but diagnostic triage tools are not always classified as generative systems, and your trust is in England, not California. The vendor contract sits with the NHS, not with you. The liability sits with the clinician who countersigns the report, not the vendor.

There is no federal law that shifts malpractice liability from a clinician to an AI tool or its developer. Under current law in most states, the physician remains responsible for any recommendation that makes it into patient care, and state medical boards and legislatures have been moving toward formalizing that responsibility. The radiologist who writes your report has accepted liability for a finished output she may have had 90 seconds to review. Real review needs a framework, the records, enough time to compare them, and the authority to make your own decision; glancing at something and signing it is not enough, so when a tool gives you a finished note or a scored recommendation and asks you to take responsibility, you are being given a reviewer's liability without the conditions a real reviewer would have.

Automation bias wears a white coat

When leading large language models were tested with physician-validated vignettes containing even one incorrect detail, hallucination rates reached 50-82%. Unlike traditional AI systems that provide discrete classifications with confidence scores, LLMs generate narrative recommendations that appear highly sophisticated yet may contain subtle but clinically significant errors. The triage algorithm sorting your scan into a priority queue does not hallucinate in the generative sense, but it does misclassify, and the introduction of AI tools will inevitably introduce novel errors that are, most fundamentally, misclassifications made by a computational algorithm, and understanding of how these translate into clinical impact on patients is often lacking, meaning true reporting of AI tool safety is incomplete.

You do not see the triage score. The system presents a worklist to the radiologist ranked by clinical urgency. A study at University Hospital Aachen found that patients support deploying AI in diagnostic imaging but balk at using it for triage decisions. The researchers noted that "uncertainty towards the mechanisms of triage, combined with uncertainty towards AI, amplifies disapproval in this field." You were not asked.

The AI-driven lab automation market is valued at $4.19 billion in 2026 and anticipated to grow to $19.23 billion by 2035, driven by pharmaceutical and diagnostic sectors that need throughput the current workforce cannot deliver. Proscia raised $50 million in March 2025 (bringing total funding to approximately $130 million) to scale its Concentriq platform; PathAI secured $165 million in Series C financing and extended its partnership with Labcorp in 2026 to deploy the FDA-cleared AISight Dx platform across Labcorp's pathology network. The capital is moving faster than the regulatory framework that would tell you whose liability you are sitting inside when the algorithm gets it wrong.

You signed nothing that named the model

The consent form you were handed before your scan covered radiation exposure, contrast reactions, claustrophobia protocols. It did not name the software that would read the image. It did not disclose whether an AI system would determine the order in which a human being looked at your file. It did not specify the error rate of the model, the training dataset, the version number, or the threshold the algorithm used to decide your scan was routine and someone else's was urgent.

AI tools displaced electronic health record usability as the top technology priority for practice leaders for the first time in a January 2025 Medical Group Management Association poll, with 32% naming them as their leading focus, up from just 13% in late 2023. The American Medical Association's 2026 Physician Survey on Augmented Intelligence reports that 81% of medical providers are now using AI in their practices, more than double the 38% reported in 2023. The shift happened while you were waiting for the appointment.

The government's 10-Year Health Plan commits to restoring the NHS constitutional standard of 92% of patients beginning elective treatment within 18 weeks by 2029, with diagnostic imaging identified as a priority area and plans to deploy validated AI reporting tools across the NHS from 2027. Validated is the word doing work in that sentence. Details on selection of cases, division of model development and validation data, and raw performance data were frequently ambiguous or missing in the pathology AI literature. The timeline says deployment comes first, validation catches up later, and you are already in the system.

Your scan is done. The pixels are stored. The model has run, or has not. A radiologist will write a report when the queue reaches your name, incorporating or ignoring what the algorithm flagged, and you will receive the result by post or by phone. If the system missed something, you will learn that in three months or six months or a year, when the symptom that should have been caught presents again, louder this time. If the system over-triaged and you were escalated unnecessarily, you will never know, because the counterfactual does not generate a letter.

Behind each one is a person waiting to understand what is happening inside their body, whether that is a possible diagnosis, a cancer staging result, or simply the reassurance that something does not need treatment. The model does not wait. The radiologist is underwater. The operational standard has not been met since 2013. You are one of 1.8 million, and the AI that may or may not have looked at your file first is the reason the system is still moving at all.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.