Glossary
Glossary: AI agents, workflows and the EU AI Act
The terms that come up in proposals, control sets and compliance discussions about AI agents. Each entry gives the definition first, then what it means in a running workflow.
Updated 11 October 2026
Agentic AI
Agentic AI describes AI systems that do more than generate text: they plan steps towards a goal, call tools and act in systems.
In a company, embedding matters more than autonomy. An agentic system reads a case, pulls context from the ERP or CRM, prepares a decision and writes the result back, inside rules someone is accountable for.
AI agent, chatbot and copilot comparedAI agent
An AI agent is software that uses a language model to understand input, decide on the next steps and use tools for them, for example a search in the article master or the creation of a draft quote.
Three questions decide whether an agent works in operations: in which system it acts, with which rights, and who releases its output. Without answers it is a demo.
Examples of AI agents in manufacturingAgentic workflow
An agentic workflow is a business process in which one or more AI agents take on defined steps, while rules, release points and the system of record set the frame.
The mechanics are almost always the same: intake, context, proposal, release, write-back. Deterministic steps stay interfaces and rules; the agent only does the reading, weighing and drafting.
Four workflows, step by stepAudit trail
An audit trail is the complete record per case: input, sources used, the agent's output, the release and any changes made by people.
It answers afterwards why a quote, a credit note or a posting came out the way it did. Retention follows the period of the business record.
Baseline
The baseline describes a process before the agent: case volume, handling time per case, error and rework rate, cycle time.
Without a baseline nobody can say after go-live what improved. It is measured rather than estimated, ideally from twelve months of system data plus a stopwatch sample.
Open the AI ROI calculatorConfidence threshold
The confidence threshold is the value below which an agent produces no draft and hands the case to a person instead.
It is set per case category and calibrated in shadow mode. The control set names who may change it, and every change is logged.
Control set
A control set brings together five things per workflow: permissions, thresholds, checkpoints, logging and escalation.
It is the document IT, compliance, data protection and the works council can read in an hour, and the reason a workflow is allowed to run.
Governance inside the workflowDeployer and provider (EU AI Act)
The EU AI Act distinguishes the provider, who develops an AI system or has it developed and places it on the market or puts it into service under its own name (Art. 3(3)), from the deployer, who uses it under its own authority in a professional capacity (Art. 3(4)).
A company using a bought-in system in its own process is a deployer. A company that has an agent built for itself and puts it into service under its own name, even only for its own use (Art. 3(11)), is a provider as well. The classification belongs with your compliance function.
Open the AI Act checkEscalation
Escalation defines what an agent does when it is unsure, when a case is outside its remit or when an input looks hostile.
The right answer is almost always the same: no draft, but a case with full context for a named person. It comes with a fallback that restores manual operation within an hour.
Guardrails
Guardrails are technical limits around a language model that check inputs and outputs, for example for disallowed content, missing sources or values outside limits.
They are part of the control set, not all of it. Business rules such as discount floors belong in the workflow as hard rules, not in the prompt as a request.
Hallucination
A hallucination is a plausible but false statement by a language model, for example an invented article number or contract clause.
In operations it is handled by mechanics rather than hope: every draft cites its sources, values are checked against the system of record, and below the confidence threshold there is no draft.
Human in the loop
Human in the loop means a person reviews and releases every output of an AI system before it takes effect. With human on the loop, the person monitors continuously and intervenes when needed; with human out of the loop, the case runs without them.
Good workflows set the level per case category and change it only on measured acceptance numbers. Human oversight under Art. 14 of the EU AI Act applies to high-risk systems; for everything else it is simply good design.
Human in the loop in detailIntelligent document processing (IDP)
Intelligent document processing reads invoices, delivery notes, order confirmations or certificates regardless of layout, extracts fields and checks them against master data.
The difference from template-based OCR lies in the exceptions: a new layout needs no new template. IDP becomes valuable once the checked data lands in the ERP and deviations reach a person with a reason.
Document processing with AI agentsLarge language model (LLM)
A large language model is a model trained on large amounts of text that understands and generates text; it is the building block underneath chatbots, copilots and agents.
For a company, the model matters less than its surroundings: where it runs, which data it sees, who switches model versions and which test set every change runs against.
Model Context Protocol (MCP)
The Model Context Protocol is an open standard through which AI applications access tools and data sources in a uniform way; Anthropic introduced it in November 2024.
MCP simplifies integration but replaces no permission decision. Every connected server raises the same question as any interface: what may the agent read, and what may it write?
Prompt injection
Prompt injection is an attack that hides instructions in input, for example in an email or a PDF, so that a language model follows them instead of its actual task.
For agents with write access it is the most important risk. The defence lies in the design: restricted rights, writing only in draft status, release for anything with external effect and an escalation rule for suspicious input.
RAG (retrieval-augmented generation)
Retrieval-augmented generation means a language model retrieves relevant documents or records before answering and bases its answer on them.
In service and quoting, RAG is how manuals, past cases and price lists come in. Quality comes less from the model than from maintaining the sources and from every draft citing them.
Shadow mode
In shadow mode the agent drafts against real, live cases while people keep deciding; both results are compared per category.
The comparison is the evidence for every later widening of release rights and the material an auditor will ask for. Acceptance criteria are set beforehand, not afterwards.
The first 90 daysStraight-through processing
Straight-through processing means a case runs from intake to posting without human intervention; the touchless rate is the share of such cases.
It is a result, not a starting point. Only when shadow mode shows for a category that the error rate holds may clean cases within defined tolerances run without release.
Document processing with AI agentsSystem of record
The system of record is the application whose data is binding for a transaction, for example the ERP for orders and prices or the ticket system for service requests.
An agent that does not write back to the system of record creates side work. That is why every serious build starts with the question of which status the agent may write in there, for example draft only.
Workflow owner
The workflow owner is the person in the business who is measured on the outcome of the workflow, for example the head of the service desk or the quote desk.
Without this person you get a pilot nobody runs after the demo. The owner co-decides which categories the agent takes on and when release steps can be dropped.
Who owns the agent