AIHA Office Agentic AI & Hyperautomation Implementation Office · demo

Knowledge & community

Nomenclator EN/FR

Bilingual definitions of agentic AI, hyperautomation, rBMS, MCP, HITL, grounding score and the other terms used in this office.

Terms

  • Agentic workflow: An autonomous software pipeline driven by large language models that can reason, plan, execute multi-step tools and adapt strategies to reach predefined goals.
  • Hyperautomation: The integration of AI, machine learning, process mining, intelligent document processing (IDP) and RPA to automate complex end-to-end processes.
  • Results-Based Management System (rBMS): A management framework aligning resources, activities and agent outputs with explicit outcome metrics and long-term institutional impact.
  • Model Context Protocol (MCP): An open standard for connecting AI models to external tools, databases and enterprise applications safely.
  • Human-in-the-loop (HITL): A control pattern that pauses automated execution for human validation or authorisation when risk thresholds are exceeded.
  • Grounding score: A metric (0.00–1.00) of whether generated output is supported by the retrieved source context.
  • ABAC: Attribute-based access control: access decided by a deterministic policy engine from user, resource, action and context attributes.
  • Air-gapped: A system with no network path to the internet or other untrusted networks.
  • Autonomy level (L0–L4): How much an agent may act without a human: from suggestion only (L0) to fully autonomous (L4, tier 1–2 only).
  • CAB: Change Advisory Board: approves releases of models, prompts, tools, adapters and bots.
  • CES: Customer Effort Score on a 1–7 scale; target ≤ 2.0, alert > 3.5.
  • Critic model: An independent model that checks another model's output against sources to find contradictions.
  • CSAT: Customer satisfaction score; target ≥ 88%, alert < 85%.
  • Data diode: A one-way transfer device: data can enter an enclave but nothing can leave through it.
  • DPO (preference tuning): Direct preference optimisation: tuning a model on pairs of preferred and rejected outputs, here from HITL corrections.
  • Evaluation set: Labelled cases used to measure quality before and after every change.
  • FinOps: Financial operations for cloud and AI: cost visibility, budgets, optimisation.
  • Guardrail: A control applied to inputs, outputs, actions, runtime or data to keep an agent within policy.
  • IDP: Intelligent document processing: extraction of structured data from documents.
  • LoRA: Low-rank adaptation: a small adapter trained on top of a base model.
  • NPS: Net Promoter Score (−100 to +100); target ≥ +45, alert < +30.
  • Open-weight model: A model whose weights can be downloaded and run on your own infrastructure.
  • Prompt injection / poisoning: Instructions hidden in input or stored content that try to change an agent's behaviour.
  • RAG: Retrieval-augmented generation: the model answers from retrieved sources, with citations.
  • Risk tier (1–4): Classification driving controls: internal draft, internal decision support, customer-facing, regulated or financial decision.
  • RPA: Robotic process automation: software bots operating applications through their user interface.
  • SDLC: Software development life cycle; here extended with evaluation, red teaming and CAB release control (AI-SDLC / LLMOps).
  • V&V: Verification (built right) and validation (right thing built), with evidence linked to gates.
  • WORM: Write once, read many: storage where records cannot be altered or deleted.

Public demo with fictional data (Example Group). © 2025–2026 Russo-Got Marius Mihail & PIR2-IT. All rights reserved.

AIHA Office

More in this demo