The language of AI governance.
Plain definitions of the terms behind enforced AI - and how each one works in practice.
A working vocabulary for teams shipping AI under scrutiny: what each term means, why it matters, and where it shows up in the control layer.
AI Fundamentals
Machine Learning (ML)
A branch of AI in which systems learn patterns from data instead of following hand-coded rules - the foundation large language models and modern AI are built on.
Large Language Model (LLM)
A machine-learning model trained on massive text to predict the next token, letting it generate and reason over language - the engine behind chat assistants and AI agents.
Token (LLM)
The unit an LLM reads and writes - a word or word-piece. Prompts and responses are measured, priced, and rate-limited in tokens.
Token Economy
The enterprise shift in which AI cost, capacity, and risk are all denominated in tokens - making token consumption a first-class thing to budget, optimize, and govern.
Token Optimization
Reducing the tokens a workload consumes - through prompt design, caching, retrieval, and model routing - to cut cost and latency without sacrificing quality.
AI Inference
Running a trained model on new input to produce an output - the live serving step where AI actually does its work, and where governance has to act.
Fine-Tuning
Further training a pre-trained model on a narrower dataset to specialize its behavior - an alternative or complement to prompting and retrieval.
AI Hallucination
When a model produces confident, fluent output that is factually wrong or fabricated - a core reliability risk of generative AI.
Foundation Model
A large model pre-trained on broad data at scale that can be adapted to many downstream tasks - the general-purpose base layer of modern AI.
Frontier Model
The largest, most capable AI models at the leading edge of the field - the highest-performing and the most expensive to run per token.
Context Window
The maximum amount of text, measured in tokens, that a model can consider at once - its working memory for a single request.
System Prompt
The standing instruction that sets a model's role, rules, and tone before any user input - the model's brief for a conversation.
Governance
AI Governance
The practice of directing and controlling how AI systems are built and used - setting policies, enforcing them at runtime, and proving compliance with evidence.
Control Layer
A dedicated layer in the AI stack that inspects, authorizes, and records every model, agent, and tool interaction - the way identity or TLS sits as infrastructure.
AI Policy (Policy-as-Code)
A machine-readable rule that defines what an AI system may or may not do, expressed as code so it can be versioned, tested, and enforced automatically.
Shadow Mode
Running a new policy against live or replayed traffic without enforcing it, so you can measure its impact before it can block anything.
Audit Ledger
A tamper-evident, queryable record of every AI decision - what was requested, which policy applied, and what was allowed, blocked, or redacted.
Data Governance
The overall management of data availability, quality, integrity, and security across an organization - the discipline AI governance extends to models and agents.
AI Observability
Monitoring how AI systems behave in production - traces, metrics, and logs of prompts, responses, and cost. Necessary, but not the same as enforcement.
Human in the Loop (HITL)
A design where a person reviews or approves an AI action before it takes effect - the escalation path for high-stakes or low-confidence decisions.
Enforcement & Security
Encryption
Encoding data so only authorized parties can read it - at rest and, via TLS, in transit. A baseline security control that AI governance is layered on top of.
TLS (Transport Layer Security)
The protocol that encrypts data in transit between clients and servers - on by default across the modern web, and the model for governance becoming infrastructure.
Post-Quantum Cryptography (PQC)
Encryption and signature algorithms designed to resist attack by quantum computers - now NIST standards (ML-KEM, ML-DSA), replacing the math a quantum computer could break.
Role-Based Access Control (RBAC)
Granting permissions by role rather than to individuals - the access-control model that AI systems extend to models, agents, and tool calls.
AI Guardrails
Controls that constrain what a model or agent can receive or produce at runtime - blocking unsafe inputs and outputs as they happen.
Runtime Policy Enforcement
Applying governance rules inline on every live request, so a policy decision actually shapes the response instead of just being reported.
Adjudication
Deciding the outcome of a governed request by combining fast deterministic checks with model-based judgment only when needed.
Prompt Injection
An attack where malicious instructions hidden in input get a model to ignore its rules - the AI equivalent of an injection vulnerability.
LLM Firewall
A runtime filter for traffic to and from large language models - inspecting prompts and responses and enforcing policy at the boundary.
Zero-Trust AI
Applying zero-trust principles to AI - never trusting a model, agent, or tool call by default, and authorizing every action explicitly.
Deny by Default
A security posture where every action is refused unless a policy explicitly allows it - the safe state is the closed state.
Attack Surface
The total set of points where an untrusted actor could reach a system - every input, endpoint, and pathway that could be exploited.
Defense in Depth
A security strategy that layers independent controls so that if one fails, others still stand between an attacker and the asset.
Software Supply Chain Security
Assurance that the software you run is exactly what its authors built - via signed artifacts, a bill of materials (SBOM), and verifiable build provenance.
AI Gateway
A runtime layer in front of AI models that mediates every request - routing, authentication, rate limiting, and policy enforcement at the boundary.
Shadow AI
Employee use of AI tools outside sanctioned, governed channels - a browser tab or personal account the enforcement boundary never sees.
LLM Jailbreak
An input crafted to make a model bypass its own safety instructions or restrictions - talking the system out of the rules it was given.
LLM-as-a-Judge
Using a language model to grade or classify another model's output against a rubric - a flexible but costly evaluation step, used sparingly in enforcement.
AI Red Teaming
Deliberately attacking an AI system - with adversarial prompts, jailbreaks, and edge cases - to find failures before real users or attackers do.
Data Loss Prevention (DLP)
Controls that detect and stop sensitive data from leaving an organization - extended, for AI, to the prompts and responses crossing the model boundary.
Data Exfiltration
The unauthorized transfer of data out of a system - by an attacker, a compromised agent, or a careless prompt. The failure mode data controls exist to prevent.
Identity and Access Management (IAM)
The systems and policies that manage digital identities and control what each one can access - the identity foundation AI governance binds every model and agent action to.
Secrets Management
The secure storage, distribution, and rotation of credentials - API keys, tokens, certificates - so applications and agents get access without secrets sprawling.
Mutual TLS (mTLS)
TLS in which both sides of a connection authenticate each other with certificates - not just the server - so services can trust who they are talking to.
Agents & MCP
AI Agent
An AI system that can take actions toward a goal - calling tools, APIs, and other systems - not just generating text.
Model Context Protocol (MCP)
An open standard that lets AI models connect to external tools and data sources through a common interface.
MCP Gateway
A control point for Model Context Protocol traffic - an approved server registry, scoped credentials, and policy checks on every tool call.
Tool Call
A request from a model or agent to run an external function - a search, a database query, an API action - and return the result.
Least Privilege
Granting an AI agent only the minimum access it needs for a task, so a compromise or mistake has limited reach.
Blast Radius
The maximum damage an AI action could cause if it goes wrong - the scope a single mistake or attack can reach.
Agentic AI
AI systems that pursue goals autonomously - planning, using tools, and taking multi-step action - rather than only responding to a single prompt.
Multi-Agent System
Several AI agents working together on a task - a planner delegating to workers, or agents calling other agents - multiplying both capability and attack surface.
Function Calling
A model capability that lets an LLM emit a structured call to a developer-defined function - the mechanism underneath tool use and agents.
Data
PII Redaction
Detecting and removing or masking personally identifiable information before it reaches a model or leaves your environment.
Data Egress
Data leaving your controlled environment - in AI, the sensitive content that flows out in a prompt to an external model.
Synthetic Data
Artificially generated data that mirrors the statistical shape of real data without containing real records.
Differential Privacy
A mathematical guarantee that the output of an analysis reveals almost nothing about any single individual in the underlying data.
Retrieval-Augmented Generation (RAG)
A technique that retrieves relevant documents at query time and feeds them to a model so its answer is grounded in your data.
Embeddings
Numeric vector representations of text or other data that place similar meaning close together - the substrate of semantic search and retrieval.
Vector Database
A database that stores embeddings and retrieves them by similarity - the engine behind semantic search and retrieval-augmented generation.
Personally Identifiable Information (PII)
Any data that can identify a specific person, directly or in combination - the category most privacy law and AI redaction turns on.
Data Pipeline (ETL)
The automated flow that moves and transforms data from sources to destinations - the ETL/ELT plumbing that feeds analytics, warehouses, and AI.
Data Warehouse
A centralized system optimized for querying and analyzing large volumes of structured data from across an organization - a primary source AI reaches into.
Data Lineage
The traceable record of where data came from, how it was transformed, and where it flowed - the map that makes data auditable and trustworthy.
Deployment & Infrastructure
Container (OCI Image)
A lightweight, isolated unit that packages a process and its dependencies to run on a shared host kernel - the standard way self-hosted software like DataStrict is deployed.
Private Cloud (BYOC)
A single-tenant deployment that runs inside your own cloud account or datacenter, so software runs within your trust boundary instead of a vendor's multi-tenant service.
Single-Tenant Architecture
A deployment model where one customer gets a dedicated instance - no compute, database, or memory shared with anyone else - eliminating the cross-tenant blast radius.
Air-Gapped Deployment
Running software on a network with no route to the public internet - zero egress, no phone-home - for classified, sovereign, or strictly regulated environments.
Confidential Computing
Protecting data while it is being processed by running workloads inside a hardware-encrypted enclave, so even the host operator cannot read memory in use.
On-Premises (On-Prem)
Software that runs on hardware an organization owns and operates in its own datacenter - the maximum-control alternative to a vendor's cloud service.
Virtual Private Cloud (VPC)
A logically isolated, private network within a public cloud account - the boundary a bring-your-own-cloud deployment runs inside.
Key Management Service (KMS)
A service that creates, stores, and controls cryptographic keys - so an organization holds the keys that protect its own data, separate from the software that uses them.
Compliance
EU AI Act
The European Union's risk-based regulation of AI, setting obligations for high-risk systems including transparency, oversight, and record-keeping.
NIST AI RMF
A voluntary US framework for managing AI risk across four functions: Govern, Map, Measure, and Manage.
Data Residency & Sovereignty
Requirements that data be stored and processed within a specific jurisdiction, and remain subject to that jurisdiction's laws - central to GDPR and public-sector rules.
SOC 2
An independent auditor's attestation that a service organization's controls for security, availability, and confidentiality are designed and operating effectively - the default enterprise trust bar.
HIPAA
US law governing the privacy and security of protected health information - mandatory for anyone handling US healthcare data, with real penalties for breaches.
GDPR
The EU's General Data Protection Regulation - the comprehensive law governing how EU residents' personal data is collected, processed, and stored.
ISO 27001
The international standard for an Information Security Management System - a documented, audited approach to managing information risk that global and regulated buyers often require.
Protected Health Information (PHI)
Individually identifiable health data covered by HIPAA - a high-sensitivity subset of PII that AI systems must detect and guard.
Data Minimization
The principle of collecting and processing only the data actually needed for a purpose - a GDPR requirement that maps directly onto scoping what AI can see.
Govern AI like infrastructure.
Talk to our team about deploying DataStrict across your enterprise stack.