Glossary
The vocabulary.
Definitions for the terms that decide how AI programs are funded, governed, and measured — written the way we would explain them to an operating committee, not the way a vendor deck would.
Measurement
8 terms
AI evaluation (eval)
An AI evaluation, or eval, is a repeatable test that measures whether an AI system produces a better outcome than the alternative it replaced, against a defined baseline.
Read the definitionEval discipline gap
The eval discipline gap is the distance between knowing that an AI system is being used and knowing whether it produced a better outcome than what it replaced.
Read the definitionCalibration data
Calibration data is the labelled record of how your organization actually decided real cases, used to tune and test whether an AI system reaches the same judgment for the same reasons.
Read the definitionOutcome metric
An outcome metric measures the result an AI system was built to change — decision quality, cost, cycle time, leakage — as opposed to an activity metric that only counts usage.
Read the definitionBaseline (and counterfactual)
A baseline is the alternative an AI system is measured against — ideally the cheapest unglamorous approach that mostly works, not the absence of any system at all.
Read the definitionGolden dataset
A golden dataset is a fixed, representative set of real cases with agreed correct outcomes, held constant so that performance can be compared across releases.
Read the definitionPrecision and recall
Precision is how often the system is right when it flags something; recall is how much of what it should have flagged it actually caught. Improving one usually costs the other.
Read the definitionAutomation rate
Automation rate is the share of cases completed without human intervention — a useful operating number and a dangerous headline when quoted without its quality partner.
Read the definitionAutonomy
6 terms
Agentic AI
Agentic AI describes systems that plan multi-step work, call tools and external systems, and take actions on their own rather than only returning text to a person.
Read the definitionDelegation boundary
A delegation boundary is the explicit line between what an AI system may decide on its own and what it must hand back to a person, expressed in actions and value limits rather than in intent.
Read the definitionTool use (function calling)
Tool use is the mechanism by which a model stops producing text and starts taking action — querying a system, writing a record, sending a message.
Read the definitionModel Context Protocol (MCP)
The Model Context Protocol is an open standard for connecting AI systems to tools and data sources through a consistent interface rather than bespoke integrations.
Read the definitionAgent orchestration
Agent orchestration is the layer that decides which model, tool, or human handles each step of a task, and what happens when a step fails.
Read the definitionCopilot vs. agent
A copilot proposes and a person disposes; an agent decides and acts. The distinction is not sophistication — it is who is accountable for the outcome.
Read the definitionGovernance
8 terms
Prompt injection
Prompt injection is an attack in which instructions hidden inside content an AI system reads — a document, an email, a web page — are followed by the model as if they came from its operator.
Read the definitionShadow AI
Shadow AI is the unsanctioned use of AI tools by employees outside approved channels — and, read correctly, an unusually honest requirements document.
Read the definitionAI governance
AI governance is the set of controls, records, and accountable owners that determine which AI systems may operate, on what data, with what authority, and who answers when one is wrong.
Read the definitionEU AI Act
The EU AI Act is the European Union's risk-tiered regulation of AI systems, imposing the heaviest obligations on high-risk uses — which include several core insurance functions.
Read the definitionGuardrails
Guardrails are the deterministic checks around a probabilistic system — input filters, output validators, scope limits, and refusal paths that do not depend on the model behaving well.
Read the definitionExplainability
Explainability is the ability to state, in terms a regulator or a customer accepts, why a specific decision came out the way it did.
Read the definitionModel risk management (MRM)
Model risk management is the established discipline of inventorying, validating, and independently challenging models in use — a framework most financial institutions already have and most AI programs quietly bypass.
Read the definitionData lineage
Data lineage is the recorded path from a source system to a decision — what data was used, where it came from, when, and under what permission.
Read the definitionOperations
9 terms
Human in the loop
Human in the loop means a person reviews or approves an AI system's output before it takes effect — meaningful only when that person has the time, information, and authority to disagree.
Read the definitionStraight-through processing (STP)
Straight-through processing is the share of transactions — claims, submissions, endorsements — completed end to end with no human touch.
Read the definitionModel drift
Model drift is the gradual decline of an AI system's real-world performance as the world, the data, or the underlying model changes beneath a deployment that has not itself been altered.
Read the definitionLarge language model (LLM)
A large language model is a statistical system trained on vast text that predicts the most plausible continuation of an input, which is why it is fluent by construction and correct only by supervision.
Read the definitionHallucination
A hallucination is a fluent, confident output that is not supported by any source — not a malfunction, but the predictable behaviour of a system built to produce plausible text.
Read the definitionRetrieval-augmented generation (RAG)
Retrieval-augmented generation is the pattern of fetching the relevant source documents first and asking the model to answer only from them, so the answer is traceable to evidence.
Read the definitionContext window
The context window is the amount of text a model can consider at once — the working memory that bounds what any single call can take into account.
Read the definitionException queue
The exception queue is where cases the system will not or should not decide are routed for human judgment — and the truest indicator of whether an automation program is working.
Read the definitionFine-tuning
Fine-tuning adapts a base model to your style, format, or task using your own examples — powerful for consistency, largely useless for supplying facts.
Read the definitionEconomics
4 terms
Inference cost
Inference cost is the per-request compute cost of running an AI system in production — the number that determines whether an automated decision is cheaper than the human one it replaced.
Read the definitionToken
A token is the unit of text a model reads and writes — roughly three quarters of a word — and the unit in which almost all AI usage is priced.
Read the definitionCost per decision
Cost per decision is the fully loaded cost of getting one operational decision made — model calls, retries, tooling, and the human review time that remains — expressed per claim, submission, or file.
Read the definitionTotal cost of ownership (AI)
Total cost of ownership is the full multi-year cost of an AI capability — inference, integration, evaluation, governance, review labour, and the re-work triggered every time the model underneath changes.
Read the definitionGo deeper
Ask a question instead.
Our research assistant answers from the full Lanthos library and cites the articles it drew from, so you can check the reasoning yourself.
Ask the Insights