Routing is often mistaken for "pick the best model", but in legal work it is closer to choosing the right path for the task.
That path might include a smaller model, a frontier model, a private deployment, retrieval, a human review step, a policy check, an evidence check, an execution gate or a blocked state. The route is the combination of controls and capabilities used to move the work forward, not a single vendor SKU.
This matters because legal AI is no longer a single tool problem. Firms are already dealing with general-purpose assistants, document review systems, contract tools, research products, drafting tools, workflow platforms, client portals and internal experiments. Each one may make its own decisions about model choice, data handling, logging, review and output release.
If the firm does not own the routing layer, coordination does not disappear, it becomes implicit, inconsistent and fragmented across prompts, tools and shadow workflows.
Routing is not model selection
Model selection matters, but it is only one part of routing.
A model may be capable, but still inappropriate for the work because the matter is sensitive, the client terms restrict processing, the output is intended for external use or the evidence base is not strong enough. Another model may be less capable in general terms, but more appropriate for a narrow, repeatable task running inside a controlled environment.
The routing layer should decide where the work goes based on the task, not just the model leaderboard. It should take account of context, sensitivity, cost, latency, reliance, review requirements and the destination of the output.
In practice, the question is not simply "which model is best?", but "which route is acceptable for this work, with this context, for this intended use?"
The route is the whole path
A route is not only the model call, but the sequence of decisions and controls around the work.
For one task, the right route might be retrieval against approved sources, a smaller model for extraction, human review and an audit record. For another, it might be a frontier model with strong logging and a mandatory approval step before anything reaches a client. For highly sensitive material, the correct route might be private inference or no AI use at all.
That makes routing a practical orchestration problem because it coordinates models, tools, humans, policies and gates into a controlled path rather than letting each system make isolated decisions.
Why firms need to own the layer
Firms need to own routing because they own the professional risk. Vendors can provide models, tools, infrastructure and assurances, but they cannot fully decide how a firm should handle a specific matter, client instruction, jurisdictional constraint or reliance scenario.
That judgement sits with the firm, because the firm is the party accountable for how work is handled, reviewed and ultimately relied on.
If routing is left to individual tools, each product becomes its own decision island. One system may use a hosted model by default. Another may retain logs differently. Another may allow outputs to be exported without review. Another may apply a generic policy that does not reflect client-specific terms.
None of those choices may be unreasonable in isolation, but together they create a weak audit story. The firm may have many AI tools, but no single way to explain why a particular piece of work went down a particular path.
What routing decisions should consider
A useful routing layer should consider at least:
- the task type and intended output
- the matter type and current matter state
- the client, jurisdiction and data handling constraints
- the sensitivity of the documents or facts involved
- the permitted inference environments
- the model or tool capability required
- the cost and latency tolerance for the task
- the level of reliance being placed on the output
- the destination of the output
- the review, approval or evidence required before use
These factors do not need to produce the same route every time, but they should produce a defensible route each time, with enough consistency that legal, risk and engineering can understand why the work moved the way it did.
Where routing breaks today
Routing breaks when decisions are hidden inside prompts, product defaults or informal human workarounds.
A lawyer may know not to paste certain material into a general tool, but that knowledge does not become a system rule. A workflow may require partner review before client release, but the AI tool may not know the output is intended for a client. A vendor may offer strong security terms, but the system may not distinguish between low-risk drafting and privileged investigation material.
The result is not always obvious failure. More often, it is inconsistency: similar work being handled differently, sensitive work moving through convenient routes, review happening too late and audit trails showing generation without explaining use.
Over time, that inconsistency becomes a governance gap rather than a speed advantage.
What good routing records
Good routing should create a record of the decision, not just the output.
It should capture what work was requested, what context was considered, which route was selected, which policy version applied, which evidence supported the route, what review was required and whether any gate approved, blocked or escalated the output.
This matters because prompt logs alone are too thin. They can show what was asked and what was generated, but not why that path was allowed, whether the output was permitted for its destination or whether the route matched the firm’s policy and the client’s terms.
A routing record gives legal, risk and engineering a shared object to inspect. That is important because AI governance cannot live only in policy language or only in code. It needs a bridge between the two.
The governance gap
The routing layer will exist whether firms design it or not. Work will still move between models, tools, humans, documents, policies and client outputs.
The choice is whether that movement is controlled and inspectable, or whether it is improvised across product defaults, prompt habits and individual judgement.
Owning the routing layer does not mean building every model or tool internally, but it does mean owning the decision logic for how legal work moves through the AI stack.
That is where coordination becomes governance.