The words of agentic AI, defined and sourced.
Each term in 40 to 60 words, with the source that backs it and the Scalt page that puts it to work.
Updated
Governed AI agent
An AI agent is a system built on a large language model that perceives a context, plans and acts with tools to reach a goal. It is governed when its access rights are bounded, when its sensitive actions wait for a human approval, and when every run leaves a verifiable trace.
Sources : Wang et al., “A Survey on Large Language Model based Autonomous Agents”, 2023 · Regulation (EU) 2024/1689 on artificial intelligence, EUR-Lex
In Scalt : GovernanceSovereign AI
An AI is called sovereign when the organisation using it keeps control of its data, of the models it calls, of the hosting and of the law that applies. In France, ANSSI grants the SecNumCloud qualification to cloud services that offer this level of trust; sovereign AI relies on those guarantees or on an installation on the organisation’s premises.
Source : ANSSI, “SecNumCloud”
In Scalt : Security and sovereigntyShadow AI
Shadow AI is the use of AI tools by employees without their organisation’s approval or oversight: a consumer assistant open in a browser, internal documents pasted into a prompt. According to the 2024 Work Trend Index by Microsoft and LinkedIn, 78 % of AI users at work bring their own tools.
In Scalt : Security and sovereigntyHuman approval
Human approval is the step where an authorised person approves, corrects or rejects an action proposed by an agent before it runs: sending an email, changing a case, committing a spend. The European AI Act makes human oversight a requirement for high-risk systems (Article 14).
Source : Regulation (EU) 2024/1689 on artificial intelligence, EUR-Lex
In Scalt : Inbox & CasesAudit log
The audit log is the time-stamped record of what an agent did: sources consulted, tools called, decisions taken, approvals received, results produced. It lets you reconstruct an incident, answer an audit and prove who authorised what. The European AI Act requires this automatic logging for high-risk systems (Article 12).
Source : Regulation (EU) 2024/1689 on artificial intelligence, EUR-Lex
In Scalt : GovernanceAgent register
The agent register is the up-to-date inventory of the AI agents deployed in the organisation: purpose, owner, data and tools it can access, risk level, planned approvals, review date. The NIST AI Risk Management Framework makes such an inventory a basic governance mechanism (GOVERN 1.6).
Source : NIST, “AI Risk Management Framework (AI RMF 1.0)”, January 2023
In Scalt : AI Act kit: the agent registerAgent orchestration
Agent orchestration makes several specialised AI agents cooperate on one task: an orchestrator splits the work, hands out sub-tasks, passes results along and decides when to stop or ask for an approval. Multi-agent systems built on large language models are described in the survey by Guo and co-authors (IJCAI 2024).
In Scalt : Agents & ApplicationsRAG (retrieval-augmented generation)
RAG, for retrieval-augmented generation, has the model search a document base for the relevant passages before it writes its answer, instead of relying on its training memory alone. The answer cites its sources and reflects the current state of the documents. The method was described by Lewis and co-authors in 2020.
Source : Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, 2020
In Scalt : Knowledge Hub