Pharma AI glossary
Agent / agentic AI
AI that doesn't just answer questions but plans and executes multi-step tasks using tools: querying systems, drafting and filing documents, triggering workflows.
Definition
AI that doesn't just answer questions but plans and executes multi-step tasks using tools: querying systems, drafting and filing documents, triggering workflows. The step-change in risk is autonomy: an agent's mistake propagates through several actions before a human sees it, which is why agent governance lags adoption badly (Deloitte: ~75% of enterprises plan agent deployment within two years; only 21% have mature agent governance). In GxP settings, an agent that writes to a validated system is itself part of the validated process; its actions need the same audit trail, access control, and review any human actor would face. Neither FDA nor the EU has agent-specific guidance yet; both frameworks' logic (risk-based evidence; qualified human oversight) applies unchanged.
Beyond chatbots
Agentic AI plans and executes multi-step work through a harness: allowed tools, scoped permissions, and logged actions. In quality, that might mean retrieving batch context from LIMS, compiling CAPA evidence for review, searching change-control packages, or preparing a QMS draft only after approval. The regulated object is the agent system (orchestration, gates, audit trail), not the underlying model in isolation.
In GxP settings
An agent that can write to a validated system is part of the validated process. Governance defines which workflows it may touch (investigations, CAPA, batch documentation, change control, training lookups); the harness enforces tool allowlists, identity, and stop points; auditability captures prompts, retrieved sources, tool calls, and human approvals before anything becomes a GMP record.
Frequently asked questions
How is an AI agent different from an LLM chatbot?
Chatbots return text in a session; agents run inside a harness that selects tools, calls APIs, and chains steps toward a workflow goal. For GxP, the harness (permissions, logging, human gates) is what you govern and validate, not the chat interface alone.
Can agentic AI support deviation investigations?
Yes, when deployed as a governed agent, not a free-form model. The harness must scope which systems it can read, block unapproved writes to QMS, route drafts through qualified human review, and log retrieval, tool use, and sign-off in an audit trail. Governance sets allowed use cases; workflow defines where the agent stops; auditability proves what happened before a record is official.
Where else can governed agents assist in GMP?
Common patterns include CAPA and change-control document prep, batch record compilation for human sign-off, LIMS/MES data pulls, supplier quality file review, and environmental or equipment alert triage routed to staff. Each workflow needs its own harness rules: read-only vs draft vs write, qualified reviewers, and logs tied to the official record.
Related terms
Benchmark your AI literacy
Free 20-question assessment, eight modules, and instant scoring, built for pharma quality and manufacturing teams.
Open the Literacy CenterBrowse all terms
See the full 71-term glossary index or search interactively in the Literacy Center.
Back to glossary hub →Educational content only, not legal or regulatory advice. Regulatory guidance cited here includes drafts (FDA AI credibility guidance; EU GMP Annex 22) as of August 2026; verify against final texts before relying on them in submissions. Company-reported figures (Merck CSR timings, Sanofi results) are labeled where used. MIT's ~95% pilot figure carries its own caveat (~150 interviews, contested definitions, not peer-reviewed). Re-check sources on module finalization.