Sovereign automation lexicon

Artificial intelligence (AI)

AI is an umbrella word. It covers older techniques like machine learning and newer ones like generative AI. For a business leader, the useful question is not "do we use AI" but "which problem does it solve, and what is it worth". Many problems are solved faster with simple automation and no AI at all.

At IAC: IAC starts from one business problem, then checks where AI or plain automation pays.

Machine learning (ML)

Machine learning is behind fraud detection, demand forecasts and ticket classification. It needs good historical data, and its quality depends on that data. It predicts and sorts well, but it does not write text like generative AI does.

At IAC: In Fast Discovery for IT, ML algorithms and other techniques look for patterns, noise and meaning in your ticket data.

Generative AI (GenAI)

Generative AI is what most people mean when they say AI today. It is good at drafting, summarizing and answering questions. It can also be wrong with full confidence, so it needs checks before its output drives real decisions.

Large language model (LLM)

An LLM works by predicting the next piece of text, again and again. Products like chat assistants and AI agents are built on top of one or more LLMs. Models change fast, so your design should let you swap one model for another.

At IAC: IAC is tech-agnostic and works with models such as Claude, GPT and Gemini, chosen to fit your architecture.

Context window

A token is a small piece of text, often a word or part of a word. The context window holds your instruction, any documents you add and the answer. Many vendors price usage per token, so long inputs cost more. When the window is full, older content drops out or must be cut.

Hallucination

Hallucinations happen because a model produces likely text, not checked facts. You cannot remove them fully, but you can reduce them: give the model approved sources, test it, and keep a person on high-impact decisions. Treat any unchecked AI answer as a draft.

At IAC: In IAC builds, answers come from your approved data, and people stay in the loop for high-impact or unclear decisions.

Chatbot

Older chatbots follow scripts and decision trees. Newer ones use an LLM, so they answer in free text. A chatbot is useful for questions and simple requests, but the work after the answer is often still manual.

Copilot (AI assistant)

A copilot helps one person go faster, for example drafting an email or a report. The human stays in charge of every action. The gain is real but often hard to measure, because it depends on how each person uses it.

At IAC: Advisory on demand compares options such as Copilot Studio, Agentforce or your own stack, on your architecture, data and costs.

Agentic workflow

In an agentic workflow, the AI decides how to handle a case, uses tools, and passes the case on or escalates it. The fixed parts stay fixed. This fits work with many variations that rules alone cannot cover.

At IAC: IAC moves rules-based bots to AI-driven, agentic workflows, in waves, each with a way back.

Model Context Protocol (MCP)

Anthropic introduced MCP in November 2024. Before it, each AI tool needed its own custom connector to each system. With MCP, a system can be exposed once and used by any AI application that supports the protocol.

At IAC: IAC connects AI agents on Claude, GPT and Gemini to your systems through MCP.

Tool calling

The model does not run the tool itself. It sends a structured request, the application runs the tool, and the result goes back to the model. This is what lets an AI agent look up an order or update a ticket, and it is where most of the risk sits.

Orchestration

Orchestration is where the business rules of an AI solution live. When several specialized agents work together, it is called a multi-agent system, and orchestration coordinates them. Whoever owns this logic controls how the work is done.

At IAC: IAC builds multi-agent systems with tools such as LangGraph and CrewAI, and in its model the client owns the orchestration logic.

Human in the loop

Human in the loop is used for high-impact, unclear or regulated decisions. It slows things down a little but keeps control and accountability with people. The hard part is choosing which steps need a person, and which can run alone.

At IAC: In IAC builds, people stay in the loop for high-impact or unclear decisions, and nothing reaches production without your go or no-go.

Retrieval-augmented generation (RAG)

RAG lets a model answer with your current policies, products or procedures, without retraining it. It can also show which source it used, so people can check. The answer is only as good as the documents behind it.

At IAC: IAC builds knowledge and document AI: answers and document processing grounded in your own data.

Grounding

Grounding is the main way to reduce made-up answers in business use. It usually means RAG, access to your systems, or both. It also helps audit, because you can trace an answer back to a source.

At IAC: In IAC builds, answers come from your approved data.

Fine-tuning

Fine-tuning helps when you need a fixed format, tone or a very specific task done the same way each time. It costs time and needs good example data, and you may need to redo it when the base model changes. For facts that change often, RAG is usually the better choice.

Embeddings

Embeddings let a system find content by meaning, not only by exact words. A search for "refund" can then also find "money back". They are a key building block of RAG and semantic search.

Vector database

Vector databases are used in RAG to pick the right passages before the AI answers. Many existing databases now offer vector search too, so you may not need a new product. What matters more is keeping the content current and access rights correct.

Robotic process automation (RPA)

RPA is fast to deploy on old systems with no API. It is also fragile: when a screen or a field changes, the bot can break. Large bot estates need support, monitoring and change control, or the cost of keeping them alive grows.

At IAC: IAC moves rules-based bots to AI-driven, agentic workflows, in waves, each with a way back.

Intelligent automation

Rules-based automation handles the clear cases. AI reads documents, classifies requests or decides the next step when the case is less clear. Together they cover more of a process from end to end.

At IAC: IAC.ai stands for Intelligent Automation Company.

Process mining

Process mining shows the real paths, delays, loops and exceptions, not the process on paper. It is a strong way to find what to automate and to set a baseline before you change anything.

At IAC: The Capability academy trains builders in process and task mining, hands-on, on your own tools and data.

Intelligent document processing (IDP)

IDP turns documents into clean data that other systems can use. It handles different layouts and sends unclear cases to a person. Quality depends on the document types and on good exception handling.

At IAC: IAC builds knowledge and document AI: answers and document processing grounded in your own data.

Integration platform as a service (iPaaS)

iPaaS tools let you link systems like a CRM, an ERP and an ITSM tool without writing all the code yourself. They often include workflow features too. They are a common base for automation that runs across several systems.

At IAC: IAC works with workflow and integration tools such as Workato, n8n and Power Automate, and recommends what fits your architecture.

IT service management (ITSM)

ITSM is the process side of IT support, and ServiceNow is one well-known ITSM tool. Because it runs on tickets, it holds a lot of data on what breaks and what users ask for. That makes it a natural place to start with automation.

At IAC: In IT ticket automation, classification, routing, approvals and fulfilment run as workflows in your ITSM tool.

Ticket deflection

Deflection usually uses a self-service portal, a knowledge base or an AI agent that can answer and act. It cuts volume and gives users an answer at any hour. It only works if the answers are right, or users come back and open a ticket anyway.

At IAC: In IT ticket automation, AI agent self-service lets users solve common requests themselves, at any hour.

Self-healing automation

Self-healing works on known, repeat problems with a clear fix, like restarting a service or clearing a full disk. Every automatic fix should be logged, so people can see what happened and remove the root cause for good.

At IAC: In IT ticket automation, self-healing fixes known issues automatically, before a ticket is raised.

AIOps

AIOps groups related alerts, spots unusual behavior and can trigger automated fixes. The goal is fewer incidents and faster resolution. It needs clean monitoring data to work well.

At IAC: IAC uses AIOps for alert correlation and automated fixes, as part of optimizing what already runs.

MLOps and LLMOps

MLOps came from machine learning, LLMOps applies the same ideas to applications built on LLMs. They cover release steps, testing, monitoring and roll-back. Without them, each model update is a risk.

At IAC: IAC hosts, runs, supports and improves your AI and automation after go-live.

Model drift

A model that worked at launch can slowly give worse results as customers, products or processes change. With LLM applications, a provider update can also change behavior. You only see drift if you keep measuring after go-live.

Evaluations (evals)

Evals turn "it looks good in the demo" into a measured result. A good set uses real cases, including hard and unusual ones. Run them again when you change the prompt, the data or the model.

At IAC: At IAC, every deliverable is tested against its acceptance criteria before release.

Observability

For AI, observability also means seeing the inputs, tool calls, outputs and costs of each run. It is how you explain a bad answer, find a failing step, and prove to an auditor what happened.

At IAC: IAC puts logging, audit trails and monitoring in every automation.

Service level agreement (SLA)

An SLA makes service quality measurable and gives you a remedy when it is missed. Good SLAs measure what users feel, not only what is easy to count. Review them regularly, as the service changes.

At IAC: IAC managed services run your AI and automation under an SLA, a unit price, or outcome-based.

Level 2 and level 3 support (L2, L3)

Level 1 takes the first contact and solves simple cases. Level 2 troubleshoots further and escalates. Level 3 fixes the root cause, handles major incidents and changes code or configuration when needed.

At IAC: IAC provides named L2 and L3 engineers who run and fix your AI and automation on your hours, up to 24/7.

EU AI Act

The AI Act entered into force on 1 August 2024 and applies in phases. Bans on some practices and AI literacy came first, in February 2025. Under the current timeline, rules for high-risk AI systems land in December 2027, and August 2028 for AI built into regulated products. It can apply to companies outside the EU when their AI is used in the EU.

At IAC: IAC classifies each use case by risk, including its EU AI Act category.

High-risk AI system

Examples include AI used in hiring, credit decisions, education or critical infrastructure. These systems need risk management, good data, logging, human oversight and technical documentation. Under the current timeline, the high-risk rules land in December 2027, and August 2028 for AI built into regulated products.

At IAC: The AI Enablement Framework sets risk tiers and controls, ready for the EU AI Act.

Responsible AI

Responsible AI is not a policy document on a shelf. It shows up in daily practice: risk checks, access rules, tests, human approval and audit trails. It is cheaper to build in from the first use case than to fix later.

At IAC: IAC applies two controls at every step, from the first idea to the solution in production.

Guardrails

Guardrails can block certain topics, hide personal data, check outputs or stop an agent from acting above a set limit. They work best in layers. No single guardrail catches everything.

At IAC: At IAC, each agent gets clear limits: what it may do, what it may not, and when it hands over to a person.

Shadow AI

Shadow AI often starts with good intent, like pasting a document into a public chat tool to save time. The risk is data leaks and decisions nobody can trace. Bans rarely work alone, approved options and clear rules work better.

At IAC: The AI Enablement Framework gives every AI initiative one intake, risk tiers and one front door across the group.

Prompt injection

Prompt injection matters most when an AI agent can read outside content and take actions. There is no complete fix today. You reduce the risk by limiting what the agent can do, checking inputs and outputs, and asking a person to approve high-impact actions.

Model lock-in

Models improve and prices change fast. If prompts, orchestration and data are built around one model, you cannot move when a better or cheaper option comes. A good design keeps the model a replaceable part.

At IAC: IAC designs for no model lock-in: platforms follow your architecture, not a partner list.