Category · agentic systems
Agentic legal AI in France: choose the autonomy level before the product
A legal AI agent does more than answer: it may plan several steps, call tools, inspect results and sometimes trigger actions. In a French-law workflow, every additional permission increases the need for boundaries, logs and approval. Start read-only, enforce sources and action budgets, and place every external communication, modification or decision behind explicit human approval.
Scope: French law · information, not legal advice
Autonomy–guardrail ladder
Moving up a level is justified only when control and reversibility increase with it.
- 01Level 0 · answer
The tool responds without acting; a user checks sources before use.
- 02Level 1 · planning
The agent proposes a plan and subtasks but waits for approval before execution.
- 03Level 2 · tool-assisted reading
It searches an authorised read-only corpus with query logs and scope limits.
- 04Level 3 · internal production
It creates reversible drafts or classifications isolated from production systems.
- 05Level 4 · controlled action
Every external send, modification or commitment requires dedicated human approval.
- 06Emergency stop
An owner can suspend the agent, inspect traces and return to a known state.
Choose a starting point without entering personal facts
Select a category only. No identity, date, amount, document or free-text question is included in analytics events.
Level 0 · answer
The tool responds without acting; a user checks sources before use.
Keep the output, verify it against the listed sources, then obtain human review for any material decision.
What is an agentic legal AI system?
It is a system that can chain steps and use tools towards a defined legal-work objective under permissions and controls. Check the source, its date and the full document before using this point in a decision.
How do agents differ from copilots and chatbots?
A chatbot converses, a copilot supports a human workflow, and an agent may plan and execute multiple actions within granted rights. Check the source, its date and the full document before using this point in a decision.
Can an agent send a legal letter?
It may be technically possible, but external action should be blocked by explicit approval, especially when it can create an effect or deadline. Check the source, its date and the full document before using this point in a decision.
How should tool access be limited?
Apply least privilege, allowlisted actions, read-only defaults and volume and time limits. Check the source, its date and the full document before using this point in a decision.
Which traces should be retained?
Retain the objective, version, tool calls, sources, outputs, approvals and errors without adding unnecessary personal data. Check the source, its date and the full document before using this point in a decision.
How do I test a legal AI agent?
Use an isolated environment, adversarial scenarios, simulated actions and stop criteria before real access. Check the source, its date and the full document before using this point in a decision.
Does the EU AI Act ban legal AI agents?
Not as a category; obligations depend on the system, its role, intended use and risks under applicable provisions. Check the source, its date and the full document before using this point in a decision.
Does Julie act autonomously?
Julie is described here as an assistant; do not attribute external actions or human decisions that the product does not claim. Check the source, its date and the full document before using this point in a decision.
I want to scope an AI agent for a French legal workflow. Help me choose an autonomy level, permissions, approvals and an emergency stop without assuming matter data.