Menu
Close
.svg)
On August 24th, Tempus AI announced that the FDA had cleared an artificial-intelligence tool designed to detect signs associated with pulmonary hypertension from a standard 12-lead electrocardiogram (ECG). Tempus ECG-PH produces a binary output for healthcare providers (HCPs) evaluating symptomatic patients aged 40 or older. It is neither a standalone diagnostic nor a tool for monitoring patients over time, two caveats worth retaining before the exuberance begins. Even so, an ordinary ECG may now become an earlier checkpoint on the pathway to a notoriously elusive diagnosis.
For the pharmaceutical industry, and especially companies working in pulmonary hypertension and other diagnostically elusive conditions, this news is more than an intriguing regulatory milestone. It affords a glimpse of how AI-enabled diagnostics may reconfigure what we call the “buying process”: the sequence through which symptoms are noticed and investigated, a diagnosis is reached, treatment options are weighed, therapy is initiated and the next clinical action is determined. By intervening at multiple points along this journey, AI may make the process earlier, faster and more precise.
The buying process was never the clean, orderly sequence portrayed in brand plans. Patients present late, symptoms masquerade as other conditions, referrals are delayed and clinical decisions inherit all the oddities of human behaviour. Yet pharma has generally concentrated its marketing research efforts on the familiar human and institutional participants: HCPs, patients, caregivers, office staff, payers and health systems. Companies such as Tempus AI, Guardant Health, Caris Life Sciences, Natera and Foundation Medicine are introducing another actor, one that is not human at all. A widening array of AI-enabled systems can now influence what is noticed, which evidence becomes salient and when clinical action is triggered.
The distinction matters. Precision oncology itself is hardly novel: oncologists have used molecular profiling to match tumors with targeted treatments for years. What is changing is the nature of the intelligence surrounding those tests. AI systems can now interrogate routine ECGs, pathology slides, molecular results and longitudinal clinical records, uncovering patterns that were previously difficult, or impossible, to discern. AI need not prescribe a medicine to reshape its market. It can operate much earlier, determining whether a patient enters the relevant clinical pathway at all. A signal surfaced from an ECG may prompt an echocardiogram, referral or catheterization. Opportunities for treatment increasingly germinate upstream of the prescription.
Tempus illustrates the leap from molecular testing to multimodal inference. Its research-stage PRISM2 foundation model integrates computer vision and language models to interpret ordinary pathology slides, predict biomarker status, and explore patient prognosis. Rather than awaiting a purpose-built test for every clinical question, such models may extract new meaning from materials already produced during routine clinical care. The slide is no longer merely an image to be inspected; it becomes a computational terrain from which additional signals can be excavated.
Guardant offers another snapshot of this emerging continuum. Its testing estate -- Shield for screening, Guardant360 for treatment selection and Reveal for residual-disease and recurrence monitoring -- generates information across successive stages of cancer care. Its newer InfinityAI platform is designed to harmonize genomic, epigenomic, transcriptomic and longitudinal clinical data, helping identify new biomarker-defined populations and follow tumor evolution over time. The novelty is not simply that Guardant offers several tests. It is that screening, selection and surveillance can become connected stations along the same digital carriage, with intelligence accumulating as the patient moves between them.
Caris is making the transition still more tangible. MI Clarity employs computational pathology to analyze digitized routine tissue slides alongside clinical inputs, estimating both early and late distant-recurrence risk in certain patients with early breast cancer, all without requiring genomic sequencing. The system is intended to detect subtle histological patterns that conventional assessment may not readily quantify. Caris describes a larger ambition to move from early detection to “early interception,” a phrase that deserves attention. Detection observes disease; interception seeks to alter its trajectory.
Natera shows where such intelligence can ultimately lead. In May, the FDA approved Signatera CDx as a companion diagnostic for adjuvant atezolizumab in muscle-invasive bladder cancer. The assay identifies patients who are molecular-residual-disease (MRD) positive and may benefit from treatment, while MRD-negative patients may be spared therapy unlikely to help them. Signatera is not, strictly speaking, an AI product; and that is precisely why it belongs in this discussion. It demonstrates how a machine-readable biological signal can become a formal checkpoint in the treatment journey. The diagnostic does not merely decorate the decision with more information. It helps determine whom to treat.
Roche-owned Foundation Medicine is also moving beyond the conventional genomic report. Through a planned integration with Roche’s Navify Clinical Hub, it intends to link test results with AI-powered clinical-trial matching and patient education, alongside publication searches and digitised guideline pathways. The report, once a comparatively static document, is becoming an interactive guide to what the HCP might do next.
That is the more consequential shift. These companies are not simply producing better diagnostics. They are constructing an interpretive layer around the clinical journey, one capable of finding new patients, extracting new signals from familiar materials, predicting what may happen next and placing new choices before HCPs. The buying process is no longer merely being measured by technology. Increasingly, it is being shaped by it.
The seminal change is not any single test. It is the accumulation of an infrastructure around the patient journey. Each clinical encounter can generate another parcel of molecular, imaging, pathology, phenotypic or outcomes data, subject of course to consent and governance. Those data can improve models, support new diagnostic applications and refine the questions placed before HCPs. A test that begins as a snapshot can become part of a learning system.
Guardant provides an unusually clear example. Patients undergoing colorectal-cancer screening with Shield may opt to receive results covering nine additional cancers, provided they join the company’s data initiative and authorize access to their medical records. Guardant claims that a majority of PCPs ordering Shield are opting into these multi-cancer reports. With between 270,000 and 285,000 colorectal-cancer tests expected this year, routine clinical use could rapidly generate a substantial database of real-world performance, one that Guardant hopes will support an eventual FDA submission expanding Shield from colorectal screening into multi-cancer detection.
Adoption creates evidence; evidence supports expansion; expansion creates further adoption. The flywheel, in this instance, begins with a blood draw.
Caris reports a similar compounding effect across more than 1.13m patient profiles, with AI trained on the resulting molecular depth. The company says these data power Caris Detect and its ambition to convert an early cancer signal into personalized immune targets. Tempus, meanwhile, delivered a foundation model to AstraZeneca that was tested on its ability to predict treatment response in public and blinded clinical trials. Its proposed acquisition of Personalis would add MRD data that may reveal recurrence months before a conventional scan. In each case, information generated at one checkpoint is intended to improve the intelligence available at the next.
Natera demonstrates how such a loop can eventually redraw the treatment pathway itself. An initial Phase III trial of atezolizumab in muscle-invasive bladder cancer failed its primary endpoint in an all-comer population. But results among Signatera-positive patients helped germinate a second PhIII restricted to patients with detectable MRD. That study subsequently led to FDA approval of Signatera as a companion diagnostic and its inclusion in treatment guidelines. The data did not merely improve the test; they helped redefine who should receive the medicine.
None of these loops turns automatically. Natera notes that prospective studies can take years to mature. Guardant still requires patient authorization, sufficient evidence and regulatory acceptance. Reimbursement, guidelines, workflow integration and HCP trust remain formidable stage gates; and healthcare has never suffered from a shortage of ingenious technologies awaiting budget and/or formulary approval.
Nor will HCPs become passive passengers in this evolving clinical theatre. They will interpret, challenge and sometimes override these outputs, as they should. But it would be equally foolish to assume that clinical autonomy renders the surrounding information architecture irrelevant. Decisions remain human; the environment in which they are made is becoming computational.
All of this leaves pharmaceutical companies with a strategic blind spot. A brand team may conduct an exquisite set of studies that capture prescribing attitudes, quantify preference, and test a score of message effectiveness...yet fail to examine the machinery determining whether the patient ever reaches the prescribing occasion in the first place. If an algorithm alters who is screened, how disease is classified, which biomarker is reported or when recurrence is suspected, the buying process has changed before the sales representative, advertisement or visual aid has had its at bat.
Treatment-journey research must therefore thoroughly examine and map three interlocking systems:
Pharma needs to understand where these technologies enter the journey; which patients they reveal, overlook or misclassify; what new triggers they create; how HCPs learn to use and trust them; when HCPs question or override their outputs; and how access, guidelines and reimbursement govern their adoption. It must also follow the data: what data are collected, where they travel, how they are interpreted and how they may influence the next generation of diagnostics and therapeutics.
This is not an argument for replacing traditional buying-process research with a guided tour of fashionable technology and digital algorithms. Beliefs, habits, evidence, emotion and clinical experience continue to matter enormously. It is an argument for widening the aperture. The treatment journey is becoming simultaneously more personalized while at the same time more systematized -- a neat contradiction of the sort healthcare produces with admirable regularity.
The next pharmaceutical buying process will not be designed solely based upon what happens in the exam room, the guideline committee meeting, or the brand-planning workshop. Part of it will be written in code. Companies that understand how that code interacts with clinical judgement will see where new treatment opportunities are forming. Those that do not may win the battle for brand preference and still lose the patient before the brand ever enters the conversation.
