Agentic AI — The 2026 Frontier

Standard AI tools, the ones covered in Lessons 13.2 through 13.4, respond to a single input and return a single output — an NLP model reads a verbatim term and returns a MedDRA code suggestion; the human decides what happens next. Agentic AI is different in kind, not degree: it receives a goal — "process this adverse event report" — and then plans and executes the entire sequence of steps required to achieve it, autonomously, using whatever tools are available: NLP for reading, database APIs for searching, validation logic for checking, routing protocols for escalating anything uncertain to a human queue.

This is genuinely shipping in production, not a 2026 talking point about the future. Five named agent types are running inside real PV platforms today. An Intake Agent reads an adverse event email or document end to end, extracts minimum case elements, proposes seriousness and expectedness, checks for duplicates, and creates a pre-populated ICSR shell — routing anything uncertain to a human reviewer queue rather than guessing. A Coding Agent reads the verbatim description, searches the MedDRA hierarchy directly, and proposes an LLT with a linked PT, a confidence score, and ranked alternatives, flagging low-confidence cases for senior review — this is the same MedDRA discipline Module 4 built, now with an autonomous agent doing the initial search.

A Literature Agent runs scheduled searches across PubMed, Embase, and other defined databases on its own, classifies every retrieved abstract, pulls full text for anything potentially relevant, and drafts an ICSR shell for qualifying cases without a human initiating each individual search. A Distribution Agent manages the genuinely complex logistics of multi-jurisdiction E2B(R3) submission — calculating jurisdiction-specific deadlines from Day 0 (Module 8’s data-lock-point discipline, automated), generating submission files, routing to the correct regulatory gateway, and flagging any acknowledgment rejection for immediate human attention. And a Signals Agent runs continuous disproportionality analysis and ML clustering across the company’s entire safety database, generating a candidate signal list with supporting statistics and even a preliminary narrative draft for the medical reviewer to work from.

The single most important thing to understand about all five is captured in one line worth memorising precisely: "agentic" does not mean "autonomous final decision-maker." Every one of these agents, in every production platform running them, operates inside a defined human oversight framework specifying exactly where the agent handles the mechanics of a multi-step workflow and exactly where a human must confirm before anything proceeds. An intake agent creates a case shell; a human authorises its creation. A coding agent proposes a code; a human makes the final selection, always. A distribution agent generates a submission file; a human authorises every expedited transmission before it goes out. Treating "agentic" as synonymous with "the AI decides" isn’t just conceptually sloppy — it’s a governance error, and it’s exactly the kind of imprecision an inspector, or an interviewer, will catch immediately.