How we structure matter-scoped AI context

Legal AI fails when context leaks, a draft that cites the wrong precedent, or a summary that blends two clients’ facts. Every Largence AI surface inherits the active matter’s ACL: documents, messages, and templates the user may already access. Regulators on both sides of our markets, ICO and NITDA, expect purpose limitation; matter scope is how we operationalise it.

type MatterContext = {
  matterId: string;
  documentIds: string[];
  templatePackId?: string;
  jurisdiction: 'NG' | 'UK';
};

async function buildCustoContext(matterId: string): Promise<MatterContext> {
  const acl = await resolveMatterAcl(matterId);
  return {
    matterId,
    documentIds: acl.readableDocuments,
    templatePackId: acl.activeTemplatePack,
    jurisdiction: acl.jurisdiction,
  };
}

Agents extend the same envelope. An automation cannot attach to a channel or document the initiating user could not open manually. That constraint is enforced server-side, not prompt-side.

Governance belongs in the platform layer, not in the hope that the model behaves.