Glossary

Enterprise AI & integration terms, as we use them.

Practitioner-written definitions for the terminology of enterprise AI, integration, BPMS, and digital transformation as we use it in our delivery work. Not encyclopedic; matched to how the terms come up in real engagements.

Agentic AI: AI systems where an LLM-driven agent coordinates work — calling functions, reading data, taking actions — toward a goal. In production, agentic systems are typically bounded supervisor-worker patterns with deterministic hand-offs and human checkpoints on consequential actions.
API Gateway: An infrastructure layer that sits in front of backend services, handling authentication, authorisation, rate limiting, routing, transformation, observability, and policy enforcement. For LLM workloads, an internal API gateway pattern is increasingly used to centralise model access.
BPMS (Business Process Management Suite): Platforms that model, execute, monitor, and improve business processes using BPMN 2.0 notation. Modern BPMS integrates AI for decision augmentation, intelligent routing, and exception handling alongside conventional workflow orchestration.
Chain of Thought: A prompting technique where the model is encouraged or required to produce intermediate reasoning steps before its final answer. Reasoning models internalise this; non-reasoning models can be prompted for it.
Embedding: A numerical vector representation of content (text, image, audio) that places semantically similar content in nearby positions in vector space. The basis for similarity search, retrieval-augmented generation, and semantic operations.
Enterprise Integration: The discipline of connecting disparate enterprise systems through governed, observable, audit-ready interfaces. Platforms include webMethods, MuleSoft, Boomi, Apache Kafka, Azure Integration Services. Foundational to AI workloads in 2025+.
Fine-Tuning: Updating a pretrained model's weights using a specific training dataset to specialise its behaviour. For enterprise workloads, fine-tuning is most appropriate for narrow high-volume tasks; retrieval is generally preferred for adding knowledge.
Function Calling: An LLM capability where the model can produce a structured call to a function exposed by the application, instead of free-form text. The application executes the call and returns the result to the model. The basis for tool-using and agentic systems.
Governance (AI): The policies, processes, technical controls, and audit mechanisms ensuring AI workloads are compliant, auditable, bounded, aligned with risk appetite, reviewable, and improving. Encoded in the AI platform rather than enforced through manual reviews in mature deployments.
Guardrails: Input and output filters that constrain AI behaviour — PII detection, prompt injection screening, content policy enforcement, schema validation, citation checking. Layered defences; no single guardrail is sufficient.
iBPMS: AI-native Business Process Management Suite. Combines BPMN 2.0 orchestration with AI augmentation for decision-making, routing, and exception handling. Intellectual's iBPMS product is one of the four products in the Intellectual Engine.
iPaaS (Integration Platform as a Service): Hosted enterprise integration platform delivering connector-based integration, transformation, routing, and governance. Intellectual's iPass product is an enterprise iPaaS.
Intelligent Document Processing (IDP): Pipelines that extract structured information from unstructured or semi-structured documents using a combination of OCR, layout analysis, classification, and LLM-based extraction. Validation against business rules is essential; the LLM is the most flexible step, not the most important.
LLM (Large Language Model): Foundation models trained on large corpora of text to predict the next token in a sequence. Current frontier models include GPT, Claude, Gemini families; competitive open-weight options include Llama, Mistral, Phi, Qwen.
MCP (Model Context Protocol): A protocol published by Anthropic in late 2024 for connecting AI clients to servers exposing tools, resources, and prompts. Gaining adoption as a standard for AI-to-system integration; reduces bespoke integration work where servers exist.
Model Risk Management: Banking framework (SR 11-7 in the US and similar elsewhere) governing the development, validation, monitoring, and governance of models. AI/ML models fall under model risk management; the framework applies without modification.
Multimodal: AI models that handle multiple input modalities (text, image, audio) natively in a single model call. Enterprise applications include document processing for complex layouts, field operations with photographic evidence, and mixed-content workflows.
Prompt Engineering: The discipline of constructing prompts that produce reliable outputs from LLMs. For enterprise integration workloads, schema-first design, decomposition, few-shot examples, and defensive prompting are core techniques.
Prompt Injection: An attack where user input or external content contains instructions intended to override the AI system's intended behaviour. Direct (in user input) and indirect (in content the model processes) variants. Layered defences required; no single mitigation is sufficient.
RAG (Retrieval-Augmented Generation): An LLM pattern where a retrieval step surfaces relevant content from a knowledge base; the content is included in the prompt; the model generates grounded in the retrieved material. The dominant enterprise LLM pattern for knowledge tasks.
Reasoning Models: LLMs that internalise chain-of-thought reasoning, spending tokens on intermediate steps before producing an answer. Examples in 2025 include OpenAI o1/o3, Claude reasoning variants. Higher cost and latency; useful for multi-step problems where accuracy matters more than speed.
Semantic Layer: A curated description of business entities, attributes, relationships, and definitions sitting between the physical data model and AI workloads. Critical for text-to-SQL and conversational BI. The investment in the semantic layer is the unlock for production accuracy.
Sovereign AI: AI capability deployed inside an institutional or national boundary — in-country compute, in-boundary data, accountable governance. Common in government and regulated industry contexts where data residency and compute sovereignty are constraints.
Tool Use: An AI capability where the model can invoke functions, APIs, or other tools as part of its reasoning. The basis for agentic systems. In production, tool catalogues are governed surfaces with permission, validation, and audit at the execution layer.
Vector Database: A database optimised for storing embeddings and answering approximate nearest-neighbour queries. Production options include Pinecone, Weaviate, Milvus, Qdrant, pgvector, and the major clouds' vector services. The foundation of retrieval-augmented systems.
webMethods: Enterprise integration platform originally from Software AG (now part of IBM after the 2024 acquisition). One of the most widely deployed iPaaS platforms in regulated enterprise and government. Intellectual has 15+ years of webMethods delivery experience.

Enterprise AI & integration terms, as we use them.

More to read.