Glossary — Pharma AI Collective

agent

An AI system given a goal and tools that chooses its own next steps, acts through those tools, checks the result and repeats until it judges the goal met or hits a limit.

AI literacy (EU AI Act, Article 4)

The EU AI Act's duty on providers and deployers of AI systems to take measures to build their staff's understanding of the AI they use. No specific level is mandated.

AI steward

A named person in a function who owns how AI is used there: keeps test cases current, watches quality, collects issues and decides with the business owner when to change or stop.

AI-native redesign

Redesigning a process around what an agent and a person each do best, often moving the human decision to a different step, instead of adding an agent to today's steps (a bolt-on).

application programming interface (API)

The documented way one piece of software asks another to read or change something. AI tool connections, including MCP servers, usually end up calling a system's API.

audit trail

A secure, time-stamped record of who created, changed or approved what, and when. In this course, Meridian also records which model version and sources produced an AI draft.

automation bias

The tendency to accept a system's output with too little scrutiny, especially when it looks polished. It can turn a human approval step into a rubber stamp. Experienced reviewers are affected too.

autonomy level

How far an agent is designed to act without its user. One published scale names five user roles, from least to most agent autonomy: operator, collaborator, consultant, approver, observer. A design choice, not a capability.

baseline

A comparable reference for a metric, measured before a change or in an unassisted group during the same period. Measure it the same way as the result, with a similar case mix, so you can judge what changed.

CAPA (corrective and preventive action)

The actions a quality system requires to fix the cause of a problem and stop it recurring, tracked to closure and checked for effectiveness.

carrying cost

The yearly cost of holding inventory, as a percentage of its value: storage, insurance and write-offs, and often a charge for the capital tied up. Count that capital charge once, never on top of released cash.

CDMO

Contract development and manufacturing organisation: a company that develops or makes products for others under a quality agreement that sets who does what.

centre of excellence (CoE)

A small central team that sets AI standards, runs the shared foundation and helps functions deliver. It enables when functions own the outcomes; it gatekeeps when every request queues behind it.

change control

The applicable regulated quality-system process for proposing, assessing, approving and documenting changes to a system, process or document before use. A vendor model update is assessed for impact; the required testing and sign-off depend on that assessment.

chunking

Splitting documents into passages, typically a few hundred tokens each, so they can be searched. A poor split can separate a rule from its exception or a table from its header.

computer software assurance (CSA)

FDA's risk-based approach to gaining confidence in production and quality-system software: more rigorous testing where failure matters most, lighter methods elsewhere. Written for medical-device makers.

connector

Software that lets an AI system read from or write to another system, such as the eQMS or LIMS. One built to an open standard may be reusable across tools; access controls and data mappings still need checking.

context engineering

Deciding what a model sees when it answers: instructions, documents, examples, tool results. The aim is the smallest set of material that gives it what it needs, and nothing that distracts it.

context of use

One precise statement of what an AI model is used for: the question it helps answer, its role in the decision, and what other evidence sits beside it. Risk is judged for that use, not for the tool.

context window

The model's working memory, measured in tokens: the conversation, supplied or retrieved material and its own output, including reasoning steps for some models. Accuracy tends to fall as it fills up.

cost of goods (COGS)

What it costs to make the products sold in a period: materials, manufacturing labour and overheads. Some companies also report amortisation of acquired intangibles here. An annual expense.

cost per successful task

Total monthly running cost (model, tools, review, rework and fixed run costs) divided by the tasks whose output was usable. The number to fund against, rather than price per token.

cost pool

A body of spend, or of risk, with one accountable owner: for example, deviation-investigation labour or inventory. An AI case should name the pool it moves and by how much.

critical GMP application (draft EU Annex 22)

In the draft EU GMP annex on AI: a use with direct impact on patient safety, product quality or data integrity. LLMs should not be used there, even with human review. A draft, not in force.

data lineage

The trail showing where a data value came from and what happened to it on the way: source system, record version, transformations. It is how you trace an AI answer back to the record behind it.

data product

A dataset packaged for reuse, with a named owner, documentation, access rules and quality commitments, such as a batch-genealogy product used by several teams and tools. A table without an owner isn't one.

decision consequence

How serious the harm would be if the decision the AI feeds turned out wrong, judged on the decision itself, whatever tool is used.

embedding

A list of numbers representing what a passage means, so a search can find passages with similar meaning even when the wording differs. It can miss exact codes and numbers.

eQMS (electronic quality management system)

The system that holds deviations, investigations, CAPAs, change controls and, at many companies, batch disposition decisions.

ERP (enterprise resource planning)

The business system for orders, materials, inventory and finance. It tracks lots and their stock status, such as quarantined or unrestricted.

evaluation (eval) and test set

A controlled set of cases with expected outcomes or agreed scoring criteria, used to assess an AI system for its intended task. Keep test cases separate from tuning where feasible; rerun relevant cases after changes to check for regressions.

fine-tuning

Further training a model on your examples to change how it behaves: style, format, a narrow task. It adds training and hosting cost and is a poor way to keep facts current.

generative AI

AI that produces content such as text, images or code. It can sit inside a chat, a fixed workflow or an agent; what differs between those is who chooses the steps.

See also: agent

grounding (grounded, traceable answer)

Answering from supplied or retrieved sources and pointing to the passage behind each statement. It makes answers traceable, not guaranteed: check the cited passage supports the claim.

GxP

Shorthand for the "good practice" rules and guidelines (manufacturing, clinical, laboratory, distribution and others) that govern work affecting product quality, patient safety and the integrity of the records behind them.

hallucination

Output that reads as fluent and confident but is false or unsupported by its sources, such as an invented section number or citation. It comes from how models generate text and happens even with good documents.

high-risk AI system (EU AI Act)

A legal category in the EU AI Act that carries extra duties, such as AI used in recruitment or qualifying AI in regulated products like medical devices. Separate from GMP criticality.

human in the loop

A design in which the system stops and waits for a person's approval before an action takes effect. Real oversight only if the person has the time, the evidence and the authority to say no.

human on the loop

A design in which the system acts on its own while a person monitors it and can intervene or stop it. Oversight comes during or after the action, not before.

hybrid search

Running meaning-based and keyword search together and merging the results, so a question finds both similar wording and exact matches such as an SOP or deviation number.

inference

Running a trained model to produce an answer. Each request is usually billed by the tokens going in and coming out, so inference cost grows with every use.

intended use

The documented description of what a system or model is for, the tasks it supports and the inputs it should expect. Validation and testing are judged against it.

ISA-95

An international standard (also IEC 62264) for connecting plant-floor systems with business systems. Its level model separates control (Levels 0–2), manufacturing operations such as MES (Level 3) and business planning such as ERP (Level 4).

jagged frontier

Uneven capability: a model can excel at a hard task and fail an easy neighbouring one. Results on one task, or on a leaderboard, don't predict results on yours.

kill criteria

Conditions, agreed before a pilot starts, under which it stops or changes course: a missed target, a safety signal, a cost per task above the value per task, no owner for the run budget.

knowledge graph

The actual facts stored as connected records, usually shaped by an ontology: this batch used that resin lot, ran in that bioreactor and is named in that deviation. Good for questions that follow links.

knowledge management

Acquiring, storing and sharing product and process knowledge so people can find and trust it when deciding. ICH Q10 names it one of two enablers of the pharmaceutical quality system.

large language model (LLM)

Software trained on a vast amount of text, then refined, to generate text that fits what it is given. It writes fluently from learned patterns and doesn't consult a source unless one is supplied or connected.

least privilege

Giving a person or system only the access a task needs, for only as long as it needs it. For an agent: its own identity, narrowly scoped reading rights, and write access only where justified.

lethal trifecta

Simon Willison's term (2025) for an AI system with access to private data, exposure to untrusted content and a way to send data out. With all three, a planted instruction can leak data. Remove one leg.

LIMS (laboratory information management system)

The system that manages samples, tests and results in the quality-control laboratory, including approval of results.

lock-in

The cost and time it would take to leave a supplier. With AI platforms it can sit as much in connectors, agent definitions and the records built up on the platform as in the model.

master data

The shared reference records that many processes use: materials, products, equipment, sites, suppliers. Managed once, with an owner, so every system means the same thing by the same code.

MES (manufacturing execution system)

The system that runs and records production on the floor: it issues the electronic batch record, enforces the steps and captures who did what, when.

metadata

The context needed to understand a data value: what it measures, its unit, where and when it was recorded, and by whom. A number without metadata means nothing, to a person or to an agent.

model context protocol (MCP)

An open standard for connecting AI applications to tools and data: a common plug. It makes connecting easier; it doesn't decide what the AI may do. Governed by the Linux Foundation's Agentic AI Foundation since December 2025.

model influence

How much the AI's output drives a decision compared with the other evidence used. Sole basis for the decision: high. One input among independent checks: lower.

model retirement

A provider withdrawing a model version, often on a published schedule; check each provider's and platform's dates. Afterwards requests to that version typically fail, so systems built on it must move and be re-tested.

multi-agent system

A design in which several agents split a task, often a lead agent directing others in parallel. It can tackle broader tasks and uses many more tokens than a single agent.

See also: agent

observability (AI)

Monitoring an AI system in use: what it was asked, what it did, what it cost and how good its answers were, so teams can investigate quality and cost trends, not just uptime.

ontology

A formal model of the things in a domain, how they relate and the rules that hold, written so software can follow it. For example: at Meridian, each Tozumab shipment lot comes from one batch; a deviation affects batches.

P&L (profit and loss statement)

The income statement: revenue minus costs over a period, such as a year. A recurring saving shows up here; cash released once from the balance sheet, such as lower inventory, does not.

permission trimming (security trimming)

Removing from search results anything the person asking couldn't open in the source system, before the model sees it. Only as good as the access lists behind it.

pilot purgatory

A pilot that neither scales nor stops: extended because users like it, with no baseline, target or deadline that could prove it worth funding or not.

process historian

A database that stores time-stamped readings from plant equipment, such as temperature, pH or dissolved oxygen, often every few seconds, for trending and investigations.

prompt injection

Instructions planted in content an AI reads, such as an email, document, web page or tool result, that it then follows as if its user had given them. Models can't reliably tell the two apart.

reasoning model

A model that works through intermediate steps before answering, which may take more time and tokens. It can help on complex, multi-step problems and can still be confidently wrong.

red team

People asked to attack a plan before it is approved: to find weak assumptions, missing evidence and ways it could fail. In this course, quality and finance play the role.

reranking

A second pass that re-scores the passages a search returned and keeps only the best few, so the model gets less, better material.

retrieval-augmented generation (RAG)

A design where the system first searches your documents, passes the best passages to the model, and asks it to answer from them and cite them. The usual design behind "chat with our documents".

review by exception

Reviewing only the records or steps a system flags as outside expected limits, instead of every line. In batch-record review it is often built from rules in the MES rather than from AI.

SCQA

Situation, complication, question, answer: a way to order an argument so the reader sees why a decision is needed before the recommendation. A common extension of the situation-complication-question opening in Barbara Minto's pyramid principle.

semantic layer

Agreed business definitions of measures and entities, such as "batches released this quarter", mapped once onto the underlying tables so dashboards, people and AI tools count the same way.

shadow AI

AI tools staff use for work without the organisation's approval or knowledge. Their outputs bypass the company's controls, records and reviews.

shared foundation

Shared components that several AI uses may reuse, such as connectors, retrieval, test sets, logging and access control, each governed to the requirements of its intended uses.

skill (agent skill)

A packaged set of instructions, and sometimes scripts, that an agent loads only when a task needs it. In effect, an SOP for the agent: it needs an owner, review and versioning.

See also: agent

stage gate

A funding checkpoint between phases of a project. Money for the next phase is released only when evidence agreed in advance is shown; otherwise the project is changed or stopped.

static, deterministic model

A model that does not keep learning once in use (static) and gives the same output every time for the same input (deterministic). Typical language-model tools are set up to vary their output.

system of record

The system whose copy of a fact counts when copies disagree, such as the executed batch record in the MES or the signed disposition in the eQMS. Copies elsewhere are for use, not for authority.

taxonomy

A controlled hierarchy of categories, from broad to narrow, such as deviation types or a product list. It names and groups things; it doesn't say how they relate in other ways.

technology transfer

Moving a product's process and the knowledge behind it from development to manufacturing, or between sites, so the receiving site can make it consistently.

time horizon (METR)

The length of task, measured in skilled-human time, that an agent completes at a given success rate, often 50% or 80%. A measure of task difficulty, not of how long an agent can safely run alone.

token

The unit of text a model reads and writes: a word or part of a word. Usage, prices and context limits are counted in tokens.

tool use

A model asking another system to do something for it, such as search a database, run a query or send a message, then reading the result. It is how AI reaches beyond the chat window.

See also: agent

two-speed portfolio

A governance design, not a regulatory category: uses whose output influences no GMP execution, documentation or disposition run under lighter review; uses that do get proportionate, documented controls, testing and QA approval.

unified namespace (UNS)

An architecture pattern, not a product: plant and business systems publish their current state and events to a shared hub, organised by site, area and equipment, instead of many point-to-point links.

value hypothesis

A testable claim for a pilot: which change, for whom, moves which measure from what baseline to what target, by when. If it can't be shown false at the deadline, it isn't one.

vector database

A store for embeddings that finds the passages closest in meaning to a question. One component of a retrieval system, not the whole of it.

workflow (AI)

A path of steps designed in advance by people, possibly with fixed branches, with a model doing one or more steps. The model doesn't choose the path, even when it runs unattended.

See also: agent

working capital

Cash tied up in running the business, mainly inventory and receivables less payables. Cutting inventory releases cash once; it isn't a recurring annual saving.