A plain-English AI glossary for the people who run clinical trials.

AI vocabulary moves fast, and most of it was written for engineers. This reference translates the terms you will actually hear from vendors, regulators, and colleagues into language grounded in clinical operations. Each definition favors what the term means for a trial over how the technology works under the hood.

Use it to follow an AI conversation, question a vendor claim, or brief your team. The final panel collects the terms Tilda uses to describe AI-native clinical operations.


THE BUILDING BLOCKS

Artificial intelligence (AI). Software that performs tasks normally requiring human judgment, such as reading a document, pulling out data, or deciding what to do next.

Machine learning (ML). A way of building AI where the system learns patterns from examples rather than following rules a person wrote by hand.

Neural network / deep learning. The layered mathematical structure most modern AI is built on, loosely inspired by the brain. “Deep” refers to many layers, which is what lets these systems learn complex patterns.

Large language model (LLM). An AI trained on very large volumes of text that can read, summarize, and generate language. It is the engine behind most current AI tools, but on its own it produces text, not verified work.

Small language model (SLM). A smaller, more focused language model. Cheaper to run and easier to specialize for a narrow task, at the cost of broad general knowledge.

Foundation model. A large, general-purpose model (an LLM is one kind) trained broadly, then adapted for specific uses. Vendors build on top of these rather than starting from scratch.

Generative AI. AI that produces content such as summaries, drafts, images, or code. It answers and drafts. By itself, it does not carry out a multi-step process or own an outcome.

Conversational AI. AI that responds when asked, then waits. It answers questions, summarizes, or explains, and relies on a person to decide what happens next. Useful for thinking, not for moving work forward.

Agentic AI (AI agent). AI that executes work. It plans steps, uses your systems, and completes a task end to end (classify a document, extract metadata, run QC, file it), flagging only what needs a person. This is the distinction that matters most in clinical operations: conversational AI tells you something, agentic AI does the work.

Natural language processing (NLP). The part of AI focused on understanding human language. It is what lets a system read a protocol or a site email and make sense of it.

Rules-based system. Traditional software that does exactly what it is explicitly programmed to do, such as EDC edit checks. Precise and predictable, but it does not infer intent or adapt to situations no one anticipated.


HOW THESE SYSTEMS ARE BUILT AND BEHAVE

Training / training data. The examples a model learns from. The quality and relevance of that data shape how well the system performs on your specific work.

Fine-tuning. Adapting a general model to a specific domain by training it further on specialized examples, such as clinical documents and regulatory standards. It is how a general model becomes useful for trial operations.

Prompt / prompt engineering. The instruction given to an AI, and the practice of wording that instruction to get reliable results.

Retrieval-augmented generation (RAG). A method where the AI looks up relevant source material before answering, then bases its response on those documents rather than memory alone. It reduces guesswork and lets the system point to where an answer came from.

Inference. The moment a model is actually running and producing output, as opposed to being trained. “Running inference” just means the system is doing its job.

Context window. How much information a model can hold in view at once. A larger window lets it work across a longer document or more records in a single pass.

Token. The unit an LLM reads and generates, roughly a word or part of a word. Usage and cost are often measured this way.

Knowledge graph. A structured map of how things relate: sites, studies, documents, people, deviations. It gives a system memory and context that compounds as more work flows through it.

Emergent behavior. Capabilities that appear once a model has seen enough examples, even though no one programmed them directly. It is why modern AI can generalize to documents and studies it was never explicitly told about, much like the intuition experienced staff build over time.

Deterministic vs. probabilistic. A deterministic process gives the same result every time for the same input. Generative models are probabilistic and can phrase things differently on each run. In regulated work, systems are engineered to make outcomes consistent and checkable rather than left to chance.

Structured vs. unstructured data. Structured data sits in fields and tables (an EDC entry). Unstructured data is free-form (a PDF, an email, a scanned form). Most clinical documents are unstructured, which is why reading them reliably is hard.

Multimodal. A system that handles more than one type of input, such as text and images. In practice, reading a scanned, handwritten, or foreign-language document.

Data quality (“garbage in, garbage out”). The principle that AI output is only as good as the data it learns from and works on. Missing or inconsistent operational data limits what any system can do, no matter how capable the model.


WHERE IT GOES WRONG, AND HOW IT IS MANAGED

Hallucination. When an AI produces something that sounds plausible but is wrong or unsupported. The real questions for clinical work are how often it happens, how it is caught, and when a human is pulled in.

Model drift. A gradual drop in performance as the real world changes and the model’s assumptions go stale. It is why validated systems need ongoing monitoring, not one-time sign-off.

Bias. Systematic skew in output caused by skew in the training data. Worth asking about wherever representativeness matters.

Confidence score. A number expressing how sure the system is about an output. Low confidence can be used to route a task to a person automatically.

Ground truth. The verified correct answer a system is measured against. Accuracy claims only mean something relative to a credible ground truth.

Failure mode. The specific ways a system can go wrong. A mature vendor can name theirs and show how each one is caught.


TRUST, OVERSIGHT, AND GOVERNANCE

Human-in-the-loop. A design where people review, approve, or handle the exceptions the AI flags rather than checking everything. In practice, this looks like a “needs attention” queue for the cases the system cannot resolve on its own.

Guardrails. Built-in limits that keep a system from acting outside its intended scope.

Explainability. The ability to show why a system produced a given output. On an inspection day, this is the difference between “the AI did it” and being able to account for every decision.

Auditability / audit trail. A complete, time-stamped record of what the system did and what a person changed. Required for regulated work under standards such as 21 CFR Part 11.

Validation / CSV. Documented evidence that a system does what it is supposed to, reliably, before it is trusted with real work. In regulated settings, this is formalized as computer system validation (CSV).

Accuracy, precision, recall. Different measures of performance. Accuracy is overall correctness, precision is how often the system’s positives are right, and recall is how many of the real cases it catches. Always ask which one a vendor is quoting.

AI governance. The policies, reviews, and oversight that keep AI performing safely over time: performance monitoring, exception reporting, and ongoing validation. Often run by a council spanning Clinical Operations, QA, IT, and Compliance.

Agent-to-agent (A2A). AI agents on different sides (site, sponsor, CRO) exchanging status, follow-ups, and reconciliation directly, with people stepping in for judgment or escalation. It cuts the email volume between organizations.


COMPLIANCE AND SECURITY

GxP. The umbrella of “Good Practice” quality regulations in life sciences (GCP, GMP, GLP, and others). “GxP alignment” is shorthand for meeting the ones relevant to your work.

GCP in an AI context. Good Clinical Practice still applies when a system does the work. The question becomes how oversight, documentation, and accountability are preserved.

21 CFR Part 11. The FDA regulation governing electronic records and electronic signatures. Any AI touching regulated records has to meet it.

ALCOA+. The data-integrity standard (Attributable, Legible, Contemporaneous, Original, Accurate, plus later additions). A useful lens for judging whether AI-handled records hold up.

SOC 2 (Type 2). An independent audit of how a vendor handles data security and controls. Type 2 tests that those controls held up over a period of time, not just at a single moment. A common security gate in procurement.

GDPR / HIPAA. Data privacy regulations (EU and US) governing how personal and health data is handled. Relevant wherever clinical documents contain patient data, sometimes referred to as PHI (protected health information).


ADOPTION AND EVALUATION

Proof of concept / pilot / production. The usual ladder for deploying AI. A pilot is a scoped, low-risk test with defined success metrics; production is live, ongoing use. A pilot works best when designed as a controlled preview of production, not a standalone experiment.

Baseline. The current performance of a workflow before AI (cycle time, error rate, rework, human hours). Without a baseline, you cannot show what AI actually changed.

Benchmark. A reference figure used to compare systems or track progress, ideally tied to your specific use case rather than a generic claim.

Time to value. How quickly a system delivers meaningful operational impact after deployment, not just eventual cost savings. A short time to value lowers adoption risk.

ROI, tangible and intangible. Tangible return shows up quickly as reduced spend, fewer manual hours, and faster cycles. Intangible return (higher quality, greater consistency, more predictable execution) is harder to measure but compounds over time.

Copilot vs. agent. A copilot suggests and drafts alongside a person who stays in control. An agent takes the task and finishes it, reporting back. The difference sets expectations about who is doing the work.

Orchestration. Coordinating multiple steps, systems, or agents so a whole process runs end to end rather than in disconnected pieces.

Integration / API. The technical connections that let an AI work inside your existing systems (EDC, CTMS, eTMF) rather than as a separate island.

System of record vs. system of execution. A system of record stores information (for example, Veeva or Medidata). A system of execution does the work on top of it. Knowing which one a vendor is selling tells you what you are actually buying.

Intelligent document processing (IDP). Reading a document and turning it into usable, structured data: classifying it, pulling metadata, and checking it. Core to TMF work.

Optical character recognition (OCR). Turning an image or scan of text into machine-readable text. The first step before an AI can read a scanned document.

Metadata extraction / classification. Identifying what a document is and pulling its key fields (sponsor, study, version, date).


TILDA’S FRAMEWORK LANGUAGE

Terms Tilda uses to describe AI-native clinical operations. These are our framing, not industry-standard definitions.

Now you're all set to understand, speak, and act with AI fluency.