设计合规的智能体人工智能——企业面临的关键法律问题.pdf

1、Designing Compliant Agentic AI Key Legal Issues for Companies.Session 1Dr Tina Gausling,LL.M.(Columbia University)Counsel(Munich),Digital BusinessSimmons&Simmons GermanyAndrew JointPartner(London),Digital BusinessSimmons&SimmonsWhat are AI AgentsWhat are AI Agents01 What are AI Agents?New dimension

2、of autonomyWorking definition:Working definition:An AI agent is a system that perceives its environment,makes decisions,and takes actions autonomously to achieve goalsgoals often using reasoning,memory,and external tools.BEFOREBEFOREGenerative AI produce contentGenerative AI produce contentReactive,

3、single-step:Reactive,single-step:input produces output in one cycleHuman-in-the-loop:Human Human-in-the-loop:Human decidesdecides how to use the output AI remains a toolLimited external reach:Limited external reach:typically no access to third-party systemsExamples:Examples:ChatGPT conversations,Cop

4、ilot suggestionsNOWNOWAgentic AI acts autonomouslyAgentic AI acts autonomouslyCore loop+layersCore loop+layers:perceive-decide-act,with planning,tools,and memory(often LLM-orchestrated)Tools:Tools:APIs,RAG systems,browsers;advanced agents can control software via UI automation(sometimes termed“LAMsL

5、AMs”)Risk profile=tools autonomyRisk profile=tools autonomyExamples:Examples:Claude Code,claims-handling agent,recruiting agentAUTONOMY GRADESAUTONOMY GRADESThe legally decisive triadThe legally decisive triadHITL:Human in the Loop:HITL:Human in the Loop:every action requires user approvalHOTL:Human

6、 on the Loop:HOTL:Human on the Loop:human supervises and can intervene beforebeforeexecutionHOOTL:Human out of the Loop:HOOTL:Human out of the Loop:no prior approval or controlTrade-off:Trade-off:more autonomy=higher productivity,but stricter duties(e.g.Art.14(2)AI Act)Consequence:Consequence:AI age