Glossary

The vocabulary, defined plainly.

The terms that come up around production AI and go-to-market, each in one clear sentence, no jargon defending itself with more jargon.

AI governance

The controls, records, and evidence that let an organization account for how an AI system decides and behaves, so that when a regulator, customer, or board asks, there is an answer.

Production-grade AI

An AI system engineered to run unattended under real data, real volume, and real edge cases, with observability and rollback built in, the opposite of a demo that works once.

The last 80 percent

The load-bearing work a prototype skips: validation, retries, permissions, escalation paths, observability, and rollback. It is where most enterprise AI fails to reach production.

Observability

The ability to see what a system is actually doing in production (its inputs, decisions, failures, and rollbacks) as they happen, rather than reconstructing them after an incident.

Human override

A design principle under which an AI system's actions are logged, reversible, and pausable by a person, so control never leaves the human.

SOC 2

An auditing standard for how a company manages customer data across five trust criteria: security, availability, processing integrity, confidentiality, and privacy.

ISO/IEC 42001

The international standard for an AI management system: how an organization governs the AI it builds and operates, from risk to accountability.

EU AI Act

The European Union's regulation that classifies AI systems by risk and attaches obligations that scale with that risk, from transparency duties to strict requirements for high-risk uses.

Grounding

Anchoring a model's output in specific source material, retrieved documents, structured data, or verifiable facts, so the response can be audited against the source rather than trusted on its confidence alone.

Retrieval Augmented Generation (RAG)

A pattern where a model is given relevant documents retrieved from a knowledge base at query time, so its answer is grounded in that material instead of only its trained weights.

Prompt injection

An adversarial input that tricks a model into ignoring its original instructions and following the attacker's instead. A core threat model for any AI system that reads untrusted content.

Guardrail

A runtime control that constrains what an AI system is allowed to say or do. Unlike a policy document, a guardrail intervenes at the moment of the decision.

Evaluation harness

The tests an AI system is graded against continuously, not once. Catches regressions when a prompt, model, or dependency changes.

Model drift

Slow degradation in a model's outputs over time as the world it sees shifts away from the world it was trained on. Detected by continuous evaluation, not by intuition.

Model card

A short, structured document that describes an AI system: what it does, what it was trained on, how it was evaluated, what it should not be used for, and who owns it. The starting point of a governable system.

Red teaming

Structured attempts to break an AI system before real users do: adversarial prompts, edge cases, and abuse scenarios, run as part of building the system rather than after an incident.