Glossary · Cheat sheet
Agentic AI on one page. What an agent is, when you need one and what keeps it in check. Six blocks to save and pass on, each with a path into the full glossary.
DWG BF-CHEATSHEET 6 blocks · as of 09/2026 Sheet 1/1 · Vienna
It is given a goal instead of a sequence of steps. It works in a loop until the goal is reached or a rule stops it. Chatbot answers · copilot suggests · agent executes . The test: who presses “execute”? If the system does it itself, it is an agent. 01 Perceive 02 Reason 03 Plan 04 Act 05 Verify GOAL In the glossary AI agent Agent vs. chatbot vs. copilot
Stable interface, fixed rules Script or API No interface, the same screens every time, rare exceptions RPA Many variants, unstructured documents, case-by-case judgement Agent No agent as long as there are no real cases to test against or nobody owns the exceptions.
In the glossary RPA Process mining
Model Understands language, but guarantees nothing.
Context & RAG Your documents as evidence instead of guesswork.
Tools & MCP The agent's hands. What does not exist as a tool, it cannot do.
Evals Quality as a measurement on real cases, not a gut feeling. In the glossary LLM RAG Tools & function calling Evals
Identity One identity per agent, short-lived tokens, never a shared service account.
Policy Limits live in code and the tool layer, not in the prompt.
Audit Every decision is logged immutably, with its reasoning.
Oversight Exceptions land in a queue with an SLA; every agent has a stop switch. Agent own identity Policy ≤ €10,000 autonomous €12,400 › limit Oversight queue · SLA 4 h ✓ approved Posting ERP · payables ✓ posted AUDIT LOG 14:02:07 Invoice RE-2231 detected · €12,400 14:02:08 Policy: €10,000 limit exceeded 14:02:08 To payables queue · SLA 4 h 14:31:55 Approved and posted Agent own identity Policy ≤ €10,000 autonomous €12,400 › limit Oversight queue · SLA 4 h ✓ approved Posting ERP · payables ✓ posted AUDIT LOG Invoice RE-2231 detected · €12,400 Policy: €10,000 limit exceeded To payables queue · SLA 4 h Approved and posted In the glossary Agent identity Guardrails & policy Audit trail Human-in-the-loop
L0 Manual work L1 RPA: fixed scripts L2 ML inside the flow L3 Agent proposes, a human confirms L4 Agent acts within limits, humans step in on exceptions Move up a level only when evals, the audit trail and the exception rate justify it.
In the glossary Levels of autonomy L0–L4
06 EU AI Act: where agents sit Prohibited High-risk Transparency obligations Minimal risk Most back-office agents sit in the lower two tiers. They move up as soon as their decisions directly affect people: pre-screening job applications is high-risk.
since 2 Aug 2026 Article 50 transparency obligations 2 Dec 2026 End of the marking grace period 2 Dec 2027 High-risk under Annex III 2 Aug 2028 High-risk under Annex I Reflects the Digital Omnibus; not legal advice.
In the glossary Risk classes The timeline
