← /editions
#76

The agent screens. The human still signs.

// Agentic AI is landing in oncology clinical trials, and the design pattern surviving regulatory contact is confidence-gated autonomy with a human signature. Epic and Oracle both shipped agent platforms into the EHR this week, and a 3,804-patient study published the ablations showing which architectural choices earn their keep.

$ today_in_10_seconds

Agentic AI is being deployed into the most heavily regulated workflow in existence — oncology clinical trials — and the pattern that's surviving contact with regulators isn't autonomy. It's fast evidence assembly with a hard stop at the human signature. Epic and Oracle both shipped agent platforms into the EHR this week, and a 3,804-patient study just published the ablations showing which architectural choices actually earn their keep.

$ lead_story

How Agentic AI Aims to Streamline Clinical Trial Enrollment

Shaalan Beg, MD, MBA, chief medical officer at ConcertAI and a gastrointestinal medical oncologist, told Targeted Oncology that agentic AI can cut cognitive and administrative burden for oncology teams and expand trial access for patients treated in community settings. His concrete claim: agents collect, synthesize, and surface chart data for the human abstractor in a uniform way, reducing screening activity from an hour or 90 minutes to a few minutes. He frames agents as specialized tools aimed at specific decision points in a clinician's workflow, not as replacements — and he's unusually explicit about what stays human. A person still has to read the records, explain the study to the patient, and obtain a signature.

// why_it_matters: The deployable agent in a regulated domain isn't the one that decides — it's the one that assembles evidence fast enough that a human can decide in minutes instead of hours.

read more →

$ also_worth_attention

  • Epic ships an agent platform into 43.7% of US acute care

    At UGM this week Epic unveiled Agent Factory, an agentic tool embedded in the EHR that lets health systems build and monitor agents that make decisions and automate clinical and administrative workflows — around 120 AI features customers can adapt to their own needs. Epic now holds 43.7% of the acute-care EHR market across more than 3,700 hospitals and 325 million patient records. And after years of refusing customer requests, Epic has built an end-to-end clinical trial management system. That last line is the one to read twice: the platform is absorbing the exact workflow today's lead-story vendors sell into.

    link →
  • Oracle Health expands its Clinical AI Agent the same week

    Oracle announced on August 19 that its Clinical AI Agent now automates professional fee coding for ambulatory workflows, supports direct dictation, and assists chart review by surfacing relevant context across the patient record. Oracle describes the agents as working as a system — using semantic reasoning to share context across workflows and collaborate in near real time while keeping clinicians in control. The company claims 400,000+ physician hours saved since 2024, which is vendor-reported and unaudited. The load-bearing sentence is the compliance one: clinicians remain responsible for reviewing, editing, and signing all documentation.

    link →
  • The ablation study nobody else publishes

    A prospective 12-month evaluation screened 3,804 consecutive oncology patients through a neuro-symbolic multi-agent platform combining LLM extraction agents, an oncology knowledge graph, a recommendation engine, and human-in-the-loop review. Headline metrics: sensitivity 0.8375, specificity 0.8359, precision 0.8121, F1 0.8246. The interesting part is what they tore out. Removing knowledge-graph grounding dropped F1 from 0.8246 to 0.79 and produced hallucinated matches violating numeric or temporal criteria; collapsing extraction and matching into a single LLM call dropped F1 to 0.78. Decomposition and symbolic grounding bought measurable points — that's rare, citable evidence for an architecture choice.

    link →

$ bigger_picture

The trend behind the headlines

Four stories, one design pattern: confidence-gated autonomy with an audit trail. Epic's autonomous coding agent routes on its own confidence — auto-code when high, flag for human review when low. That's not a product decision, it's a regulatory one. The FDA's 2026 clinical decision support guidance hinges non-device status on whether a clinician can independently review the basis for a recommendation, and states that software for time-critical decisions generally fails that test because clinicians lack time to review. Explainability isn't a feature here; it's the thing keeping the product out of device regulation entirely. And note why the knowledge graph earned its points in the ablation — it worked by blocking logically inconsistent matches, not by being smarter. If you're building agents into any regulated workflow, that's the transferable lesson: the constraint layer is what makes the reasoning layer shippable.

$ tool_of_the_day

TrialGPT

An end-to-end framework for zero-shot patient-to-trial matching, built by researchers at the NIH's National Library of Medicine, structured as three modules: retrieval, criterion-level matching, and ranking. Physician-annotated evaluation put criterion-level accuracy at 87.3% against expert performance of 88.7–90.0%, with a user study showing 42.6% reduction in screening time. Worth reading even outside healthcare — it's a clean, public reference implementation of decompose-retrieve-score-explain. Two caveats: it was evaluated on synthetic patients, and it requires an OpenAI or Azure OpenAI key to run.

explore TrialGPT →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime