Glossary of AI for pharma and medical devices

54 plain-English definitions of AI, regulatory, quality and business terms for pharmaceutical companies and medical device manufacturers.

5 whys: Root cause analysis method: ask “why?” several times in a row to get back to the root cause.
A2A (agent-to-agent): Exchange standard that lets agents communicate and hand tasks to each other.
Agent: AI program that carries out a multi-step task using tools.
Agent harness: Everything built around the model (instructions, tools, context, execution loop) that turns it into an agent.
Annual operating plan: A product’s or team’s objectives, actions and budget for the year.
Autonomy: The degree to which an agent acts without human intervention. The riskier the context, the stronger the supervision must be.
CAPA: Corrective and preventive actions: dealing with the cause of a problem and preventing it from happening again.
CNEDiMTSMedical devices: The French National Authority for Health (HAS) committee that assesses medical devices for reimbursement.
Connector, MCP: Standard way of connecting an agent to tools, data or systems. Without a connector, the agent cannot reach the right data.
CRM: Customer relationship management software: contacts, calls, call reports.
CRO: Contract research organisation: a company to which a pharmaceutical company outsources the conduct of clinical trials.
Deployer, provider: Under the AI Act: the deployer uses an AI system, the provider develops it and places it on the market.
Determinism: The same starting point always gives the same result. Essential for regulated processes.
DPO: Data protection officer: makes sure the company complies with the GDPR.
Enterprise solution: A tool designed to run in production for a team or an organisation, according to development and risk standards.
ERP: Enterprise resource planning software: purchasing, inventory, manufacturing, finance.
Evaluations (evals): Structured, repeatable tests that measure the quality of an agent’s outputs.
Extensibility: How easily tools, data or agents can be added without rebuilding everything.
FSN (field safety notice)Medical devices: Notice sent to the users of a medical device to inform them of a risk and the actions to take.
Gateway: Single, controlled entry point for all calls from agents to tools: authentication, rules, limits, traceability.
GMPMedicines: Good manufacturing practice for medicines.
Guardrails: Rules and limits that define what an agent is allowed to do.
GxP: Regulated good practices: manufacturing, laboratory, clinical, distribution, pharmacovigilance (BPx in French).
HDS: Health data hosting: certification required in France to host health data on behalf of a third party.
Human in the loop: Mandatory human approval at certain steps of an agent’s work.
ISO 13485, IEC 62304, ISO 14971Medical devices: Reference standards for medical devices: quality management system, software lifecycle, risk management.
IVDMedical devices: In vitro diagnostic medical device, governed by the IVDR.
KAM: Key account manager, for example for healthcare facilities.
Large language model (LLM): AI model trained on large amounts of text, at the heart of AI assistants.
Marketing authorisationMedicines: Authorisation to place a medicine on the market (AMM in French).
MDCGMedical devices: Medical Device Coordination Group (EU). It publishes guidance on applying the MDR and IVDR.
MDR, IVDRMedical devices: EU regulations on medical devices (MDR) and in vitro diagnostic medical devices (IVDR).
Memory: What the agent retains during a task (short term) or from one task to the next (long term). Long-term memory avoids redoing work but exposes data more.
Notified body: Designated organisation that assesses the conformity of certain medical devices before they are placed on the market.
Observability: The ability to see how an agent works: logs, steps, decisions. Used to correct abnormal behaviour.
Orchestration: System that coordinates tools, steps and models so that the right action happens at the right time.
Planning: An agent’s ability to break a goal down into steps. Makes it possible to trace an unexpected result.
PMSMedical devices: Post-market surveillance of a medical device.
POC (proof of concept): A limited trial to check that an idea works before investing.
Pre-mortem: Exercise that consists of imagining a project has failed in order to list the likely causes and prevent them.
Prompt: Written request to the AI: role, context, task, data, format and checks (chapter 6).
Prompt injection: Malicious instructions hidden in content (document, web page) to hijack the AI.
PRRCMedical devices: Person responsible for regulatory compliance at a medical device manufacturer.
RAG: Retrieval-augmented generation: the AI relies on retrieved documents to answer instead of guessing. Improves reliability.
Registration: Recording a tool, agent or skill in a register to make it visible and usable.
Runtime environment: Where the agent acts: isolated test environment or production. It is best kept close to the data.
S&OP: Sales and operations planning: a monthly process that aligns sales forecasts, manufacturing and inventory.
Semantic search: Search by meaning rather than by exact words. Weak search is a frequent cause of poor answers.
Shadow AI: Use of AI tools not approved by the company.
Skill: Reusable capability that an agent can call, for example a standard document review. Avoids rebuilding the same logic.
SmPCMedicines: Summary of product characteristics: the reference document for a medicine (RCP in French).
State management: How the agent tracks progress on a task, including the restart point if it is interrupted.
Token: Fragment of a word. The unit in which models read and are billed.
Transparency CommitteeMedicines: The French National Authority for Health (HAS) committee that assesses medicines for reimbursement: actual clinical benefit and improvement in actual clinical benefit.