Module 1 — The Oncology AI Landscape

Course: AI in Oncology Practice Foundations Format: Educational reading (~1,100 words) + worksheet + 10-item assessment Time to complete: ~25 minutes Prerequisites: None beyond Module 0 syllabus walkthrough.

Educational content only. This module is not CME, not accredited, not medical advice, and not patient-specific guidance. The Certificate of Completion is not a professional credential. Module 1 teaches you to recognize categories of AI tools used in oncology; it does not authorize, endorse, or validate any specific product.

Learner outcome

By the end of Module 1, you can map any oncology AI tool you encounter into one of six recognizable categories — and explain, in plain language, what kind of work it is doing and what kind of clinical and operational scrutiny that work attracts.

You will not finish this module a buyer, a deployer, or an ML engineer. You will finish it as a clearer reader of the next vendor pitch, journal article, conference demo, or hallway pilot proposal that lands on your desk.

Why this module exists first

Most AI-in-oncology conversations collapse into a single bucket called "AI." Once everything is "AI," nothing has different rules. The same person who would never let a billing macro touch a clinical note will sometimes let an LLM-generated draft assessment slip into a chart note unreviewed — because both are described as "AI," and the distinction between them gets lost.

This module restores the distinction. We sort the field by what the tool is doing in your workflow, not by the model architecture underneath. The model architecture is the engineer's question. The workflow location is yours.

A practical taxonomy beats a precise one. The six categories below are not exhaustive and they do overlap. Some tools sit in two categories. Some hide in one category but quietly do the work of another (a "documentation" tool that also performs triage; a "patient education" tool that quietly answers clinical questions). When that happens, the dangerous thing is usually the unlabeled one. Learning to name the second category is the skill this module builds.

The six categories

1. Ambient documentation and communication support

What it is: tools that capture, transcribe, draft, summarize, or restructure clinical visit content, patient messages, internal communications, and chart notes.

What it is doing: speech-to-text, text-to-text, and structuring. A scribe-like layer that produces editable drafts, not final clinical decisions.

Where the risk lives: the gap between "draft" and "signed." A note marked "AI-drafted" that is signed without meaningful review converts into your attestation. Consent for ambient capture, retention of raw audio, and downstream use of the captured content (vendor model training, secondary analytics) are all live questions.

Editorial note: by the author's read of the 2026 oncology vendor landscape, this category appears among the most commercially mature in oncology today. It is also the category where the economic pressure to under-review is highest, because the value proposition is time saved.

2. Imaging, pathology, and multimodal analysis

What it is: tools that interpret radiology images, pathology slides, genomic sequences, or combinations across modalities — flagging findings, quantifying tumor burden, computing biomarkers, or proposing classifications.

What it is doing: pattern recognition over high-dimensional clinical data, often with a regulated decision-support framing.

Where the risk lives: this category is the one most likely to interact with FDA's Software-as-a-Medical-Device framework. Validation evidence, the population the tool was trained on, calibration in your patient population, and silent-failure detection (the tool is wrong but confident) are the load-bearing questions.

Editorial note: much of the published clinical evidence in oncology AI to date appears concentrated in this category, plausibly because SaMD-adjacency forced earlier validation work. The strongest safety concerns also concentrate here. The two patterns coexist by design — stronger evidence tends to be generated where the safety stakes forced it.

3. Trial matching and eligibility

What it is: tools that match patients to clinical trials based on demographics, diagnosis, biomarker status, prior treatment, performance status, and other eligibility criteria — sometimes by parsing the chart, sometimes by parsing the protocol, often both.

What it is doing: structured-data extraction, eligibility logic, and ranked recommendation.

Where the risk lives: false negatives (a patient who was eligible but was not surfaced), false positives (a patient who is surfaced but who fails screening, wasting trial-coordinator time), and equity drift (does the tool surface trials evenly across the patient population it sees, or does it concentrate on the chart shapes it knows best).

Editorial note: the operational return on this category is large in cancer centers with active trial portfolios; the individual-patient return is uneven and depends heavily on the trial mix at the institution.

4. Prior authorization and administrative automation

What it is: tools that assemble, draft, submit, track, or appeal prior authorization requests; that reconcile claims; that triage referrals; that draft administrative letters; that fill structured forms.

What it is doing: structured-data extraction, document generation, and workflow automation against payer or operational systems.

Where the risk lives: errors here look administrative but become clinical when they delay or deny care. The accountability question is who owns the output: a prior-auth letter that misrepresents medical necessity is the prescriber's problem regardless of which tool drafted it.

Editorial note: this is the category where agentic AI — tools that act, not just summarize — appears to be advancing most visibly through 2026, based on observation of the prior-auth automation news cycle. It is also the category where regulators and payers are most actively re-evaluating their own use of automation, which makes the rules a moving target.

5. Precision oncology and knowledge retrieval

What it is: tools that pull guidelines, drug interactions, biomarker-treatment associations, sequencing reports, molecular tumor board references, or literature into the clinical encounter on demand.

What it is doing: retrieval, synthesis, and sometimes generation. The tool's answer quality depends on its underlying knowledge sources, the freshness of those sources, and how it cites (or hides) them.

Where the risk lives: fluent wrong answers. A retrieval-augmented model that pulls a stale guideline, a misindexed reference, or an unrelated passage and then writes a confident clinical-sounding paragraph around it is the failure mode that matters here. The safest tools in this category cite their sources and tolerate the answer "I don't know."

Editorial note: this category is the one most often demoed in conference talks and most often misunderstood in adoption. The demo answers tend to come from the questions the tool is best at; the production answers come from the questions you actually have.

What it is: tools that talk to patients or caregivers — answering general questions, summarizing handouts, scheduling, navigating appointments, supporting symptom triage, or offering education between visits.

What it is doing: conversational generation, retrieval, and routing — usually in plain language and outside the chart.

Where the risk lives: the boundary between general health education (lower-risk) and patient-specific clinical advice (high-risk, regulated, and often unsafe outside a BAA-covered, validated path). Tools that drift across this boundary — even unintentionally — are a category-defining failure mode. Trust, scope, escalation paths, and what the tool does when it does not know are the load-bearing questions.

Editorial note: this is the category where oncology has the most to lose and the most to gain. Many patients may already be using general-purpose chatbots between visits to ask about their cancer, and the live question for the field is whether oncology provides a safer alternative or cedes the ground.

Cross-category patterns to remember

Same model, different category: the same large language model can power Categories 1, 4, 5, and 6. Category labels are about workflow location, not about underlying technology. A vendor that sells you "AI" without telling you the workflow location is selling you ambiguity.
Combined-category tools attract combined-category scrutiny. A tool that documents a visit and drafts the prior-auth letter is doing both Category 1 and Category 4 work. It must satisfy both sets of questions.
Risk is a property of the workflow, not the technology. A simple template-based summarizer in a high-stakes oncology decision is more dangerous than a complex generative model in a low-stakes scheduling task. Sort by risk only after you have sorted by category.
Hidden category-2 work is the failure mode. When a Category 1 tool quietly performs Category 2 work (e.g., flags an imaging finding inside a draft note), the quiet path is the dangerous one. Modules 2 and 3 go deeper into this pattern.

What Module 1 does not do

Module 1 does not evaluate, endorse, validate, or rank specific vendors or products.
Module 1 does not give clinical advice or patient-specific guidance.
Module 1 does not certify anyone to deploy AI tools in their clinic.
Module 1 does not replace governance, compliance, IRB, or institutional review.

Module 1 gives you the map. Modules 2 through 8 give you the toolkit for what to do once you can read the map.

What's next

Module 2 — What AI Is Good At, Bad At, and Dangerous At zooms into the failure modes that show up across all six categories: hallucination, automation bias, data shift, silent failure.
Module 3 — Clinical Safety and Validation Floor sets the minimum evidence bar before a tool from Categories 2, 3, 5, or 6 should touch a clinical workflow.

Worksheet — AI Tool Category Recognition

Pick one AI tool you have actually encountered — a vendor demo, a journal article, a conference poster, a pilot proposal, a tool a colleague mentioned, or a tool you have used. Use it as your case for the worksheet. Do not include patient identifiers, real case details, or protected health information of any kind.

Tool name (or generic descriptor if not a vendor product)
In one sentence, what does the tool do in the workflow?
Primary category (Module 1 categories 1–6)
Secondary category, if any (most real tools straddle two)
Is the tool's primary category the one the vendor describes, or a different one you noticed? If different, why?
Who in your setting would be the workflow owner if this tool went live (not who buys it — who is accountable for its output)?
What is the worst plausible failure mode of this tool, in one sentence?
What evidence would you need to see before allowing the tool into a clinical (Categories 2, 3, 5, 6) versus administrative (Categories 1, 4) workflow?
Does this tool quietly perform any work in another category that the vendor does not foreground? (E.g., a documentation tool that also flags imaging findings; an education tool that drifts into clinical advice.)
What single question would you ask the vendor next, given your answers above?

Self-check: if your answers to 4 and 9 are both "no" and you spent under two minutes on the worksheet, you have probably under-read the tool. Real oncology AI tools almost always touch more than one category, and the unlabeled category is where the risk lives.

Assessment — 10-item category-and-risk sort

For each item, pick the best primary category (1–6) and assign a risk tier (Low / Medium / High). Risk tiers refer to the risk to clinical safety and patient trust if the tool fails silently — not the commercial or strategic risk to your institution.

A microphone in the exam room captures the visit, transcribes it, and returns an editable draft note to the clinician inside the EHR before the day ends.
A pathology-lab tool quantifies tumor cellularity on whole-slide images and produces a numeric score that influences downstream molecular testing decisions.
A workflow tool reads new oncology consults in the EHR overnight and surfaces a ranked list of open clinical trials each consult might be eligible for.
A back-office tool drafts prior-authorization letters by extracting diagnosis, prior therapy, and supporting evidence from the chart and submitting via a payer portal.
A point-of-care tool answers free-text clinician questions about NCCN-style guideline recommendations, citing the source guideline section in its answer.
A patient-facing chatbot on the cancer center website answers general questions about side effects, appointment logistics, and medication refills.
An imaging-AI module flags small pulmonary nodules on routine surveillance CTs and adds an annotated overlay to the radiology read.
A "scribe" tool that, in addition to drafting the visit note, autonomously messages the patient with a treatment-plan summary unless the clinician opts out within four hours.
A precision-oncology tool ingests next-generation sequencing reports and proposes a ranked list of candidate therapies with literature citations.
A patient-portal companion that, when a patient describes a new symptom, advises the patient on whether to come to the ED, call triage, or wait until the next appointment.

Answer key (educational reference; not clinical guidance):

#	Best category	Risk tier (rationale, in brief)
1	1 Ambient documentation	Low–Medium. Risk lives in the unsigned-vs-signed boundary and consent for capture.
2	2 Imaging/pathology/multimodal	High. Quantitative output influences downstream clinical decisions; classic SaMD-adjacent territory.
3	3 Trial matching and eligibility	Medium. Operational benefit is real; clinical risk is in missed eligibility (false negatives) and equity drift.
4	4 Prior-auth and admin automation	Medium–High. "Administrative" output crosses into clinical when it misrepresents medical necessity or delays care.
5	5 Precision oncology and knowledge retrieval	Medium. Citation discipline reduces risk; fluent-wrong answers without citation raise it sharply.
6	6 Patient-facing education and navigation	Medium. The fact pattern stays on the lower-risk side of the boundary unless the tool drifts into individualized clinical advice.
7	2 Imaging/pathology/multimodal	High. Same SaMD-adjacent posture as item 2; missed nodules and false positives both have downstream clinical cost.
8	1 + 6 (combined)	High. The "scribe" frame masks Category 6 work; the four-hour opt-out is itself a design decision with patient-trust consequences. The combined-category tool inherits the higher risk tier.
9	5 Precision oncology and knowledge retrieval	High. The "ranked list of candidate therapies" framing approaches clinical decision support; literature-citation discipline and human-in-the-loop posture determine whether it stays in Category 5 or crosses into regulated territory.
10	6 Patient-facing education and navigation	High. This is the category-defining example of drift into individualized clinical advice; the safer design returns to general education plus an escalation path.

Passing score for this module: 7 of 10 correct on category, with directional agreement (low/medium/high) on risk tier. Disagreement with the answer key on a single risk tier is acceptable when you can defend your reasoning in one sentence — the goal is category recognition with explicit reasoning, not memorization.

Editorial-judgment lines in this module marked "Editorial note" are held under the Lumen Course Claim Ledger v1 for external marketing. They may remain in the lesson body as labeled judgment but must not be repeated verbatim in marketing copy, sales pages, OG cards, or paid ads until backed by a dated source row.