Clinical AI: Imaging, Drug Discovery, Trials, Scribes, and Patient Chatbots

Treat "clinical AI" as one buying decision and you will make five bad ones. A radiology triage model is a regulated medical device; a protein-structure model is a research accelerant; a trial-matching engine is an NLP pipeline over your EHR; an ambient scribe is workflow software a clinician must still sign behind; a patient-facing chatbot is a triage liability question. Each carries a different evidence bar, a different regulatory pathway, and a different integration cost — and the vendors selling them rarely volunteer which lane they are in. The job of the platform and clinical-informatics leader is to sort every proposal into its lane before the demo, not after the deployment.

Five workloads, one regulatory spine

The single most useful sorting question is: does this system's output reach a patient-affecting decision, and does it do so as a device, as a clinician aid, or as pure back-office automation? The FDA now maintains a public list of AI-enabled medical devices authorized for marketing in the United States — devices that met premarket requirements including "a focused review of the device's overall safety and effectiveness" — and updates it periodically; a scan of the list's advisory-panel column shows radiology entries dominating it, with cardiovascular and neurology following well behind.^[4] That distribution is a market signal: the workloads with the clearest device pathway and the cleanest ground truth industrialized first. The table below is the lane map this guide follows.

Imaging: the lane with real randomized evidence

Imaging is where clinical AI stopped being a promise and became a procurement category. The pattern across radiology, pathology, and cardiology is consistent: models detect, quantify, or prioritize findings inside an existing reading workflow — flagging a suspected hemorrhage for earlier review, quantifying a cardiac measurement, screening slides or studies so human attention lands where it matters. The buying failure mode is equally consistent: a hospital accumulates a drawer of single-finding point solutions, each with its own PACS integration, server, and invoice, because each was bought by the department that wanted that one finding.

The evidence bar here is now genuinely high, which is good news for buyers who use it. The MASAI trial in Sweden randomized 80,033 women to AI-supported mammography screen reading versus standard double reading: the AI-supported arm detected 6.1 cancers per 1,000 participants screened against 5.1 in the control arm, kept the false-positive rate identical at 1.5%, and cut screen-reading workload by 44.3%.^[3] That is what a load-bearing claim looks like — a prespecified safety analysis in a randomized trial, not a retrospective AUC on a curated dataset. Cardiology has begun producing the same class of evidence outside the reading room: in a prospective Mayo Clinic trial of 1,003 patients, an AI algorithm applied to routine ECGs stratified patients such that atrial fibrillation was subsequently detected in 7.6% of the high-risk group versus 1.6% of the low-risk group.^[11] When a vendor's flagship citation is instead a retrospective single-site study, that is not disqualifying — but it prices the product differently, and it should price your rollout differently too.

Regulatorily, imaging AI lives on the FDA's AI-enabled device list, and two recent moves matter for how you contract. First, the agency's January 2025 draft guidance proposes total-product-lifecycle recommendations — design, development, maintenance, documentation — including transparency and bias-mitigation expectations and postmarket performance monitoring.^[1] Second, the finalized Predetermined Change Control Plan (PCCP) guidance lets a manufacturer pre-specify planned model modifications and the methods to validate them, then ship those updates without a new marketing submission, across the 510(k), De Novo, and PMA pathways.^[12] A vendor with an authorized PCCP can retrain and improve legally and quickly; a vendor without one is frozen at its cleared snapshot. Ask which one you are buying — it determines whether model drift gets fixed in months or in years.

Drug discovery: AlphaFold compressed one step, not the pipeline

The structure-prediction breakthrough is real and publicly verifiable. Before AlphaFold, decades of experimental effort had determined the structures of around 100,000 unique proteins, a small fraction of known sequences, because solving a single structure could take months to years; the 2021 Nature paper demonstrated the first computational method that regularly predicts protein structures with atomic accuracy even where no similar structure is known, validated in the blind CASP14 assessment with accuracy competitive with experimental structures in a majority of cases.^[6] DeepMind then released predictions at scale: the AlphaFold Protein Structure Database now offers over 200 million predicted structures covering nearly all cataloged proteins known to science, and reports more than 3 million users in over 190 countries; AlphaFold 3 extends prediction to the structure and interactions of other molecule types.^[2]

The enterprise read requires precision about what changed. Structure prediction compresses target characterization — one early step in a pipeline whose cost and failure risk are dominated by wet-lab validation, safety, and clinical trials. Generative-chemistry platforms that propose novel molecules sit in the same position: they widen the top of the funnel and shorten design cycles, but every candidate still has to survive synthesis, assays, and the clinic. So the honest business case for discovery AI is cycle-time and option-value on early-stage programs, not a headline reduction in total R&D spend. Pharma and biotech platform teams should also note what the AlphaFold database's openness does to build-vs-buy: when 200 million structures are free,^[2] paid platforms must justify themselves on proprietary chemistry, integration with your assay data, and workflow — not on access to structures.

The regulatory perimeter is also no longer theoretical. In January 2025 the FDA issued draft guidance — its first on AI in drug and biologic development — proposing a risk-based credibility assessment framework for AI models whose outputs support regulatory decision-making about safety, effectiveness, or quality, organized around defining the model's specific context of use.^[5] If your discovery or trial-analytics models will ever touch a submission, the practical consequence is documentation discipline now: model provenance, training-data lineage, and validation evidence tied to each context of use, kept audit-ready rather than reconstructed later.

Trial matching: the strongest near-term ROI in clinical operations

Patient recruitment is a chronic bottleneck: eligibility criteria are semantically complex, buried in free-text EHR notes, and traditionally screened by manual chart review that is slow and misses eligible patients. This is a language problem, which is why large language models moved the needle here faster than in most clinical domains. The best public benchmark is TrialGPT, an NIH-built LLM framework for patient-to-trial matching evaluated in Nature Communications: its retrieval stage recalled over 90% of relevant trials while examining less than 6% of the trial collection, its criterion-level eligibility predictions reached 87.3% accuracy — close to expert performance — and in a user study it reduced screening time by 42.6%.^[7]

Those numbers also define the correct operating model: an accuracy in the high-80s is transformative for a pre-screen and unacceptable for a final decision. Every serious deployment therefore runs AI as a funnel-narrowing stage with human confirmation of eligibility — the model ranks and explains, coordinators verify. Evaluate vendors accordingly: demand criterion-level accuracy on your own historical recruitment data, explanation output a coordinator can check against the chart, and EHR integration that respects your privacy architecture. Two risks deserve standing attention. Bias: models trained on historical enrollment can reproduce historical exclusion, so track the demographics of AI-surfaced candidates against your catchment population. And privacy: trial matching means mining identifiable clinical records, which puts it squarely under HIPAA and, for multinational sponsors, GDPR — the governance groundwork is mapped in /guides/personal-data-protection-ai.

Ambient scribes: the fastest-spreading clinical AI, and the least regulated

Ambient scribes listen to the clinical encounter and draft the note. Microsoft's Dragon Copilot documentation is usefully explicit about the shape of the product category: ambient conversation capture, draft document generation "for a clinician's review," recommendations for discrete orders and clinical data, and an instruction that organizations obtain patient consent before recording an encounter; deployment splits between standalone apps, where finalized content is transferred into the EHR manually, and partner-embedded integrations where capture, editing, and signing all happen inside the EHR.^[8] Those documented details are the evaluation criteria: the draft-for-review workflow is where accountability lives, the consent step is a workflow you must build, and the embedded-vs-standalone choice determines both clinician friction and who supports the integration.

The clinical literature, young as it is, is directionally consistent. A Stanford Health Care pilot with 48 physicians over three months found large, statistically significant reductions in documentation task load (−24.42 on the NASA-derived scale used) and burnout score (−1.94), with usability improving significantly as well.^[13] A 2025 systematic review found eleven implementation studies — ten of them published in 2024, and the Dragon Ambient eXperience family accounting for seven of the eleven — with nine of ten studies reporting improvement on at least one efficiency metric and seven of ten reporting positive effects on clinician wellness or burnout; it also found that accuracy varied, notes frequently required manual edits, and error concerns persisted.^[9] Translation for buyers: the burnout benefit is the best-evidenced claim in the category, the time-savings claim is usually real but smaller than the demo implies, and note accuracy is the metric you must audit yourself because the literature will not settle it for you.

Patient chatbots: scope is the whole game

Patient-facing conversational AI spans three very different jobs: administrative tasks (scheduling, reminders, intake), post-visit engagement (follow-up instructions, medication adherence nudges), and symptom triage. The first two are the safe beachhead — they automate phone-tag, their failure modes are inconvenience rather than harm, and they belong in the same governance bucket as the revenue-cycle automation covered in /use-cases/healthcare-ai-administration. Triage is categorically different, and the audit literature explains why.

The benchmark study in the BMJ ran 45 standardized patient vignettes through 23 symptom checkers: the correct diagnosis appeared first in 34% of evaluations, appeared in the top 20 in 58%, and triage advice was appropriate in 57% overall — 80% for emergent cases but only 33% for cases where self-care was reasonable, because the tools skewed risk-averse and pushed users toward care they did not need.^[10] That study predates LLM-based systems, and modern engines are stronger conversationalists — but it remains the canonical caution: triage accuracy is hard, asymmetric, and must be demonstrated, not assumed. Risk-averse triage also has a business cost sellers rarely mention: a bot that over-refers to the emergency department transfers cost from the call center to the most expensive care setting in the system.

The deployment rules follow directly. Constrain the bot's scope in writing — what it may advise, what it must escalate, what it must never say. Ground clinical content in your organization's approved protocols rather than open-ended generation. Build the escalation path to a human as a first-class feature and measure its latency. Log every conversation for review, and monitor triage dispositions against clinician judgment on a sample. And because these systems collect health information from identified patients, the privacy posture — business associate agreements, data retention, secondary-use limits — has to be settled before launch; see /guides/personal-data-protection-ai for the framework and /guides/sectoral-ai-regulation-regtech for where regulators are moving.

Honest objections

The skeptic's case against clinical AI procurement deserves a fair hearing, because parts of it are correct. First: outcome evidence is thinner than performance evidence. MASAI showed detection and workload effects in a randomized design,^[3] but most products on the FDA's list were cleared on standalone or reader-study performance, and very few AI deployments of any kind have shown mortality or morbidity benefit. If your health system demands outcome evidence before adoption, most of this market is not ready for you — that is a defensible institutional posture, not Luddism.

Second: the workload story has a rebound problem. A 44.3% reading-workload reduction^[3] or a 42.6% screening-time reduction^[7] only becomes value if the freed capacity is redeployed deliberately; otherwise it is absorbed invisibly. Third: automation bias is real — a clinician who stops reading past the AI flag, or signs scribe drafts unread, converts a decision-support tool into an unsupervised decision-maker, and the scribe literature's persistent accuracy caveats^[9] show why that matters. Fourth: models drift and populations differ; a system validated elsewhere degrades quietly on your data unless someone owns monitoring. None of these objections argues for abstention. All of them argue for the same thing: buy fewer systems, validate each locally, instrument everything, and staff the monitoring function before go-live rather than after the first incident.

The read: how to decide

For a health-system CIO or clinical-informatics leader, the portfolio logic falls out of the lane map. Ambient scribes and trial matching are the near-term wins: the evidence base is young but consistent, the regulatory burden is manageable, and the ROI mechanism — clinician time and recruitment velocity — is measurable within quarters. Imaging AI is the mature regulated market: buy against trial-grade evidence, prefer vendors with PCCPs^[12], consolidate integrations onto a platform strategy, and fund local validation as a line item. Patient chatbots should launch administrative-first, with clinical triage gated behind protocol grounding, escalation, and monitoring. Drug-discovery AI concerns a narrower buyer set, and for them the question is what proprietary value a platform adds on top of free, Nobel-recognized public infrastructure.^[2]

Across all five lanes, the connective tissue is governance you build once: an intake process that classifies each proposal by lane and risk, a validation standard scaled to patient impact, a monitoring function with named owners, and documentation discipline aligned with where the FDA is heading on lifecycle management^[1] and model credibility.^[5] Health systems that run that spine can say yes quickly to good proposals in every lane. Health systems that lack it either say no to everything or — worse — say yes without knowing which lane they just entered.