Agentic AI in Pharma

RAG is impressive — you ask, it retrieves, it answers. Agentic AI is a different proposition. You give it a goal, and it figures out the steps to achieve it.

Welcome back to this series on decoding AI language for people who have better things to do than read AI language. Last time we unpacked RAG — Retrieval-Augmented Generation — and why it matters for teams working with large, complex document libraries. The short version: RAG gives an AI agent access to your own documents, so it answers from your approved content rather than from the internet or its own imagination. Both of which, in pharma, are fairly alarming alternatives. This time: Agentic AI. Where things get genuinely interesting — and where the questions I'm getting asked start to carry a little more urgency.

From answering to acting. RAG is impressive. You ask a question, the agent retrieves the right documents, you get a well-grounded answer. It is, at its core, still a sophisticated question-and-answer system. You are always in the loop. Agentic AI is a different proposition. An agentic AI system does not wait to be asked a question. It is given a goal — and it figures out the steps to achieve it. It plans, it retrieves, it makes decisions, it takes actions, it checks its own work, and it moves on to the next step. All without a human approving each move.

Think of the difference this way. RAG is a very well-prepared research assistant who hands you a briefing document. An agentic AI is a very capable junior colleague who you brief on Monday morning, and by Wednesday afternoon they've completed the project, sent three emails, updated the CRM, and flagged two compliance issues they spotted along the way. That is either enormously useful or slightly alarming, depending on how well you've defined the guardrails. Probably both.

What this looks like in pharma marketing. Consider a medical affairs team preparing for a product launch. Today, the pre-launch process involves dozens of people coordinating across regulatory submissions, medical information updates, competitive landscape reviews, HCP engagement planning, and promotional material sign-off. Each workstream moves at human speed, with handoffs, waiting, and version control chaos. An agentic AI system — properly designed and governed — could monitor the regulatory submission status, automatically flag when documents need updating across dependent workstreams, draft the updated materials from approved templates, route them to the right reviewers, and update the project tracker. Not sequentially. Simultaneously.

Or consider pharmacovigilance signal review. An agent could continuously monitor incoming safety data, cross-reference against existing literature in your approved library, draft a preliminary signal assessment, and escalate to the appropriate medical officer — before a human has opened their inbox. The capability is real. It is being deployed today, in less regulated industries, at scale. The question for pharma is not whether agentic AI will arrive. It is whether your organisation is ready for what it brings with it.

And here is where the cloud conversation becomes unavoidable. When an AI agent is taking autonomous actions — retrieving data, generating content, updating systems, sending communications — it is doing all of that compute work somewhere. On a server. In a data centre. Possibly in another country. For pharma, that is not an abstract technical detail. It is a legal and regulatory question that sits squarely on the desk of your marketing and commercial leadership, not just your CTO.

Data residency: your clinical data may not be allowed to leave. Every time an agentic AI system processes a document, a trial summary, a pharmacovigilance report, a regulatory submission, that processing happens on compute infrastructure somewhere in the world. GDPR is the obvious framework. But pharma operates under additional layers: EMA guidelines, country-specific data protection laws, and in some markets explicit requirements that certain categories of health data never leave national borders. A public cloud provider running your agentic AI workloads on servers in Virginia or Dublin may or may not satisfy those requirements. The agent does not know this. It will process what it is given, wherever it is told to run. This is a governance decision that needs to be made before deployment, not discovered during a regulatory audit.

Hybrid cloud: not everything can go to the same place. The practical solution most regulated organisations are moving toward is a hybrid cloud model. Some things run in a private, controlled environment — on-premises or in a sovereign cloud — and some things run in a public cloud. For an agentic AI system, this creates a design challenge. The agent needs to orchestrate across both environments. The retrieval of a sensitive regulatory document might need to happen in the private environment. The generation of a draft promotional email from that document might be able to happen in the public cloud. The routing logic between those two things needs to be explicit, auditable, and consistent. Instruction is not a control. Architecture is a control.

Regulatory risk: who is liable when the agent gets it wrong? When a human employee sends a non-compliant promotional email, the liability chain is relatively clear. When an agentic AI system sends a non-compliant promotional communication — autonomously, at scale, across multiple markets simultaneously — the liability question is considerably less settled. Current regulatory frameworks in pharma were not written with autonomous AI agents in mind. The EMA, FDA, and national medicines agencies are actively developing guidance, but as of now the answer to "who is responsible when the agent acts incorrectly" is largely: whoever deployed it. That is your organisation.

The costs that don't appear in the vendor slide. Agentic AI is significantly more compute-intensive than a chat AI or a RAG system. An agent that is planning, retrieving, reasoning, acting, checking, and iterating is running many more processing cycles than an agent that is simply answering a question. At small scale this is invisible. At enterprise scale — running agents across a commercial organisation, processing thousands of documents, orchestrating across multiple systems — the compute costs become a meaningful budget line. And if those agents are running on public cloud infrastructure, those costs scale with usage in ways that are not always intuitive to forecast.

What good governance actually looks like. It starts with a very clear definition of what the agent is allowed to do autonomously, what it must route to a human before acting, and what it is never permitted to do regardless of instruction. These are not technical settings. They are business decisions that need to be made by commercial, medical, legal, and compliance leadership together. It includes an architecture that enforces data residency requirements structurally — not through policy documents that the agent cannot read. It has a complete audit trail. Every action, every retrieval, every decision point, logged and accessible. It is tested, hard, before deployment — not on easy scenarios, but on the kind of edge cases that will absolutely occur in the real world. And it starts small. One workflow. One market. Measurable outcomes. Before you scale.

The opportunity is real. So is the complexity. Agentic AI will change how pharmaceutical commercial organisations operate. The organisations that benefit most will be those that engage seriously with the governance questions now, before the pressure to deploy quickly overwhelms the discipline to deploy carefully. The language is new. The TLA is new. The underlying dynamic is not. Every significant operational technology that has entered regulated industries has required the same thing: clear ownership, strong governance, and a healthy scepticism of anyone who tells you the hard questions can be sorted out later.