Harvey's Orchestration Reliability Risk

This reveals that Harvey is no longer just selling a smart model, it is selling an orchestration layer that has to keep a whole chain of models, tools, and firm systems working together on every task. Once a due diligence review or drafting workflow spans OpenAI, Anthropic, Google, retrieval, and firm data sources, failures can come from the model choice, the routing logic, latency spikes, or broken integrations, which makes legal grade debugging and uptime much harder than in a single model product.1 2 3 4 5

• Harvey explicitly expanded from OpenAI to Anthropic and Google, and says different models win on different legal subtasks. That improves output quality, but it also means one legal workflow may depend on several providers with different speeds, strengths, and failure modes.2 3
• In production, reliability is not just model accuracy. Legal tech operators describe observability, throughput, fault tolerance, and architectural clarity as critical, because when a multi step agent stalls, the problem may sit in routing, rate limits, provider latency, or data handoffs rather than the answer itself.4 5
• Law firms already treat deployment as a long security and workflow exercise, often taking around six months, and they watch usage closely after launch. That raises the bar for Harvey, because every extra external dependency creates more approval work and more places where trust can break during live matters.4 6

The direction is toward tighter control of the orchestration layer. The winners in legal AI will be the vendors that can hide multi model complexity behind stable workflows, clear audit trails, and strong integrations, so lawyers experience one dependable product instead of a fragile stack of external AI services.1 5 6