Specialization is a workflow decision
A tax question and a governance question share language but demand different context. Purpose-specific fine-tuned models may help with taxonomy, extraction or domain reasoning. Their value must be established against a defined task and rights-cleared evaluation set, not assumed from the model’s label.
Choose the pattern that the task needs
A sequential pipeline can classify, retrieve, reason and verify. Parallel specialists can inspect different issues under the same authorized scope. A router can select a model by jurisdiction and task. A reviewer can challenge unsupported assumptions. Each pattern introduces costs and failure modes that should be measured.
Keep the permission service outside the conversation
An orchestrator can request access but cannot grant it. A model’s confidence, a document instruction or an urgent-sounding task must not widen a data scope. The gateway checks each retrieval and tool action against the institution’s current policy.
Build the graph with accountable agents
Agents can help classify internal records and propose evidence-backed graph relationships. Publication of consequential edges remains a controlled change. The institution needs a review queue, version history, rollback and a way to identify every downstream unit affected by a correction.
Evaluate the whole chain
A correct retrieval can still become an unsupported conclusion. A reliable classifier can still route to an unapproved endpoint. Evaluate source support, legal applicability, boundary enforcement and human review across the complete workflow. More agents are not evidence of better legal intelligence.