What is Legal Decision Intelligence?

A category that is still taking shape

Legal Decision Intelligence is not currently a settled legal standard or regulatory category. The term is used for different approaches, including internal-knowledge activation, risk analysis, scenario modelling and judicial-decision analysis. A useful enterprise definition should therefore state which decisions are covered, which evidence is used and which role remains with people.

Mirmi uses the term for a governed decision loop. A new question is compared with the organisation’s real practice, complemented by useful external legal context when the scope requires it, and routed to a permitted action. The final decision enriches memory without automatically turning an exception into a rule.

Four pillars of a reusable decision

An answer becomes decision intelligence when it shows not only an outcome, but also what supports it, the limits of that support and what the user can do next.

• Internal memory: contracts, positions, policies, approvals and exceptions remain connected to context;
• Evidence: each conclusion points to the passages, dates and sources actually used;
• Governance: roles, thresholds and permissions determine who may decide, ask for advice or escalate;
• Learning loop: the human decision, corrections and scope are retained for future situations.

How it differs from a chatbot, CLM and legal research

A chatbot provides a question-and-answer interface. A CLM manages the contract lifecycle. A document repository retrieves files. Legal research identifies legislation, rulings or external analysis. Each can contribute to a decision, but none is Legal Decision Intelligence on its own.

The decision layer connects these systems without erasing their responsibilities. The CLM or SharePoint can remain the system of record, external research retains provenance and the workflow continues to carry permissions. Value comes from continuity between evidence, precedent, action and recorded decision.

Example: handling a contract change

When a counterparty changes a liability clause, Mirmi retrieves comparable agreements and decisions, exposes contextual differences and qualifies alignment with the approved position. If evidence and threshold cover the situation, the team sees the actions permitted by its workflow. If the matter is new or sensitive, acceptance remains unavailable and Legal receives the issue with its sources.

After the decision, the result joins the history with its author, date, context and status. An exception can remain retrievable without becoming a standard for every agreement.

Minimum requirements for a reliable platform

Model quality is not enough. A Legal Decision Intelligence platform should control data, access, provenance and behaviour when evidence is insufficient.

• permissions inherited from source systems and customer isolation;
• passage-level citations and retained provenance;
• separation of internal practice, external research and generated content;
• abstention or escalation when coverage is insufficient;
• a record of consulted sources, scores, actions and human corrections;
• documented location, data flows and responsibilities for the selected deployment.

How to deploy an initial scope

Start with one frequent decision family that is costly to reconstruct and sufficiently documented. Measure the baseline, qualify authoritative sources, define covered cases and exceptions, then open the service to an identified user group.

Scaling should depend on observed quality, not imported volume. Every extension needs an explicit owner, permissions, coverage threshold and escalation route.

• average internal time per request;
• share of requests resolved without new research;
• rate and reason for escalations;
• coverage of answers by reviewable evidence;
• human corrections and decision consistency over time.

Hosting and control of data

Legal decisions, agreements and corrections are sensitive assets. In Mirmi SaaS, the application infrastructure and customer data are hosted in France by Scaleway, in the fr-par region. Customer data is isolated and is not used to train or enrich another customer’s answers.

A dedicated VPC or on-premise deployment can be assessed when the scope requires it. In every model, integrations, specialised services, data flows and responsibilities should be documented for the selected architecture.