Skip to content

Glossary

Entries are sorted alphabetically. Each one carries a short definition and a cross-reference to the BAGF section where the concept is used in context.

Every AI feature has a single named human who carries final responsibility for it. See: Pillar III §1.

An AI system that can plan, use tools (call APIs, query databases), and chain decisions to complete a complex goal. See: Pillar IV §11.

Bizzi’s five-stage process for taking an AI agent from design to production operation. See: Pillar IV §5–10.

Real-time tracking of an agent’s reasoning trace, latency, and token usage. See: Pillar IV §14.

Bizzi’s central AI engineering team. See: Pillar I §2, §6.

The routing and security layer in front of LLMs, with automatic fallback routing. See: Pillar IV §9.

The top-level body, composed of the CEO, CPTO, and Legal. See: Pillar I §3.

An immutable record of every system action, traceable back to model version, dataset, and parameters. See: Pillar III §11.

An assessment of whether a system discriminates against any defined group. See: Pillar III §2.

Synchronises changes from OLTP to OLAP and vector stores in near real time. See: Pillar IV §1.

See: Kill-switch.

A columnar database used for OLAP workloads. See: Pillar IV §1.

The model’s reported confidence in a result. surfaced through colour-coding in the UI. See: Pillar III §7.

The cost KPI. AI cost per document processed. See: Pillar I §12.

A document describing a dataset. source, size, and labelling methodology. See: Pillar III §6.

Public / Internal / Confidential / PII. each level has its own handling rules. See: Pillar IV §2.

The condition where production data distribution diverges from the training distribution. See: Pillar IV §10.

A risk assessment for personal-data processing activities. See: Pillar II §8.

The requirement that data be stored and processed within a defined jurisdiction. See: Pillar II §4.

A request by an individual to access, correct, or delete their personal data. See: Pillar II §3.

A DoS variant. an attacker forces the AI vendor to charge fees beyond budget. See: Pillar V §10.

OWASP LLM08. an agent granted more agency than the task requires. See: Pillar V §9.

The KPI for AI field-extraction accuracy. target greater than 99% for critical fields. See: Pillar I §12.

The principle that an AI system does not produce unjustified disparate outcomes across protected groups. See: Pillar III §2.

Additional training of a base model on domain-specific data.

A class of AI capable of producing new content.

The degree to which an AI answer is anchored in the supplied source documents. See: Pillar IV §7.

The requirement that AI cite its sources when reaching a decision. See: Pillar III §10.

The frequency with which an AI fabricates information not present in the source data. See: Pillar I §12.

The governance model. a central CoE plus peripheral squads. See: Pillar I §2.

The AI requires human approval before acting. See: Pillar III §3.

The AI acts on its own. a human supervises and can intervene. See: Pillar III §3.

The process of a trained model producing a prediction or response.

A filter or scanner that runs on input before the LLM sees it. See: Pillar V §3.

Bypassing an LLM’s safety constraints.

The mechanism to cut every AI feature in an emergency while keeping the manual flow available. See: Pillar V §11.

Key Performance Indicator. Bizzi tracks STP rate, accuracy, hallucination rate, cost per transaction, and others. See: Pillar I §12.

An observability platform for LLM and agentic workflows. See: Pillar IV §14.

Using an independent LLM to score the output of another LLM. See: Pillar IV §7.

A class of algorithms that learn from data rather than from explicit programming.

A document describing a model’s purpose, training data, and limitations. See: Pillar III §6.

An open standard that lets LLMs access system data safely. See: Pillar IV §12.

Decline in model prediction quality over time. See: Pillar IV §10.

An architecture in which multiple AI agents coordinate. See: Pillar IV §11.

Vietnam’s Personal Data Protection Decree (Decree 13/2023/NĐ-CP). See: Pillar II §1, §3.

The NIST AI Risk Management Framework. See: Appendix A.

Online Analytical Processing. systems designed for large-scale analytical queries. See: Pillar IV §1.

Online Transactional Processing. systems designed for high-frequency transactional workloads. See: Pillar IV §1.

The list of the ten most common risks for LLM applications, maintained by OWASP. See: Pillar V §2.

Data that identifies an individual. See: Pillar II §3.

Removing or masking identifying data before processing or storage. See: Pillar II §3, Pillar V §7.

The structured review conducted after an incident. See: Pillar I §11.

The open-source relational database used for OLTP. See: Pillar IV §1.

OWASP LLM01. An attack that inserts hidden instructions into the LLM’s input. See: Pillar V §3.

Versioned prompt management. See: Pillar IV §6.

A standard metric for distribution drift. See: Pillar IV §10.

Per-tenant and per-IP quota enforcement. See: Pillar V §10.

The inventory of personal-data processing activities. See: Pillar II §8.

A team that simulates attacks to find weaknesses. See: Pillar I §10.

Supplying an LLM with relevant documents before it generates a result.

A data subject’s right to demand an explanation of an automated decision. See: Pillar III §12.

Access control granted by role. See: Pillar V §9.

An isolated environment that runs untrusted code. See: Pillar V §8.

The incident severity scale. See: Pillar I §11.

A new model runs in parallel but does not return results to the user. See: Pillar IV §8.

The mapping to ISO/IEC 42001, NIST AI RMF, EU AI Act, OWASP LLM Top 10, and Decree 13/2023. See: Appendix A.

The KPI for the share of transactions processed by AI without a human touch. Target greater than 85%. See: Pillar I §12.

A third party that processes data on behalf of the primary processor. See: Pillar II §9.

The LLM parameter that controls randomness.

The pattern where an agent inherits the user’s access token. there is no super-admin. See: Pillar V §9.

The ability to trace a decision or result back to its origin. See: Pillar III §11.

OWASP LLM03. corruption of the training dataset. See: Pillar V §4.

The quarterly customer report covering KPIs, incidents, and model changes. See: Pillar I §13.

A database specialised for storing embeddings for similarity search. See: Pillar IV §1.

A numerical representation of text or images in a high-dimensional space.

The discipline of evaluating and monitoring third-party risk. See: Pillar II §6.

Web Content Accessibility Guidelines. the accessibility standard Bizzi adopts. See: Pillar III §4.

A vendor commitment not to retain Bizzi data after processing. See: Pillar II §6.