Agentic vs Generative AI in Hematology: From Trial Summaries to CAR-T Toxicity Management

Agentic AI vs Generative AI: What's Next for Healthcare?

From summarising trials to autonomously managing CAR-T toxicity, AI's generative and agentic paradigms are reshaping hematology care. The goal is simple: reduce cognitive load, shorten time-to-treatment, and raise consistency in high-stakes settings.

Generative AI: What it does well today

• Condenses clinical trials, guidelines, and EHR notes into clear briefs for tumor boards and rounds.
• Drafts consult notes, discharge summaries, patient instructions, and prior-authorization letters.
• Surfaces similar patient cohorts and outcomes to guide case discussions.

Its limits are known: missing context, hallucinations, and inconsistent calibration. Use retrieval from vetted sources, strict prompt templates, citations, and mandatory human review to keep outputs safe and useful.

Agentic AI: Where care is heading

Agentic systems don't just generate text-they take steps toward a goal using tools, protocols, and feedback. Think of them as policy-aware assistants that can watch data streams, trigger workflows, and request approvals.

In hematology, that can mean continuous monitoring after CAR-T, automated grading of CRS/ICANS from vitals and labs, proposing orders (e.g., tocilizumab, steroids) per protocol, notifying the attending, and executing pre-approved steps. Human sign-off stays in loop for high-risk actions.

Hematology use cases you can implement now

• Trial summaries for tumor boards with line-of-therapy fit and eligibility flags.
• CAR-T post-infusion watchlist: real-time CRS/ICANS grading, smart escalation, templated notes.
• Order set guidance: conditioning, antimicrobial prophylaxis, growth factors, transfusion thresholds.
• Sepsis and febrile neutropenia alerts tuned to hematology patients, with pre-built management bundles.
• Transfusion stewardship: dose optimization, single-unit strategies, and indication checks.
• Bone marrow transplant (BMT) pathway automation: day-by-day checklists, labs, and consult nudges.
• Prior auth and financial clearance drafts pulled from EHR facts and guideline references.
• Patient messaging: plain-language education and symptom check-ins with escalation rules.

Safety, governance, and regulation

• Data protection: PHI minimization, encryption, role-based access, and rigorous logging.
• Model oversight: pre-deployment validation, drift monitoring, error review, and bias audits.
• Action controls: tiered autonomy-inform, suggest, execute-with-preapproval, execute-with-hard-stop.
• Documentation: versioned prompts, source attribution, decision logs, and clinician acknowledgments.
• Regulatory alignment: follow good machine learning practices and an enterprise AI risk framework.

Helpful references: NIST AI Risk Management Framework, and ASTCT consensus on CRS/ICANS grading.

Integration checklist (EHR and clinical ops)

• FHIR-based data pipes for vitals, labs, meds, notes, and orders; cut latency to minutes, not hours.
• Protocol encoding: local order sets, dosing rules, and escalation paths represented as machine-readable policies.
• Identity and privilege: who can approve, override, or pause the agent; clear on-call routing.
• Human-in-the-loop UX: one-click acceptance/edits, rationale visibility, and sources on display.
• Simulation first: run in shadow mode on historical and live data before enabling any autonomous step.

Metrics that matter

• Documentation time per case and after-hours burden.
• Time-to-intervention for CRS/ICANS and febrile neutropenia.
• ICU transfer rate post-CAR-T and 30-day readmissions.
• Order set adherence and variance by shift.
• Alert acceptance rate and override reasons (to curb alert fatigue).

Build vs. buy: decide fast with these questions

• Can the vendor ingest your protocols and map to your EHR safely (read-only vs write-back)?
• Are there audited outcomes in hematology or closely related services?
• What's the failure policy? Clear fallbacks, clinician ownership, and uptime SLOs.
• How are prompts, policies, and models versioned and governed?
• Who maintains ongoing validation when guidelines or order sets change?

Clinical guardrails for agentic workflows

• Keep high-risk actions behind explicit approval unless pre-approved under strict criteria.
• Require transparent rationale with links to sources and local protocols.
• Make "pause agent" one click, with auto-notify to supervising teams.
• Run frequent case reviews and publish change logs to staff.

Change management and upskilling

Start with a small, motivated unit. Co-design prompts, thresholds, and escalation rules with the clinicians who will use them. Measure weekly, iterate fast, and expand only after clear wins.

If your team needs structured training on clinical AI workflows and safety, explore role-based learning options such as courses by job at Complete AI Training.

What good looks like in 6-12 months

• Trial summaries and documentation drafts are routine, with citations and high acceptance rates.
• Post-CAR-T monitoring runs in shadow mode first, then progresses to suggest-mode with pre-approved steps.
• Meaningful drops in time-to-intervention and documentation time, without new safety events.
• Governance in place: risk reviews, audit trails, and a simple model update process.

Bottom line: use generative AI to clear clerical work and surface the right information; use agentic AI to run protocol-driven steps with tight oversight. Clinicians set the rules, AI handles the busywork, and patients benefit from faster, more consistent care.