The AI moat in legal, tax, health, and insurance isn't the model — it's the domain structure

The structure is the product

Harvey launched Harvey Graph, which organizes legal and matter data and maps how it connects to give agents relational context — and in the company's Agent Bench experiments, agents using the context graph improved rubric pass rates by up to 30% while reducing task costs by up to approximately 50%. Harvey's US case-law knowledge source uses a graph to map citations and relationships across American case law, with precedent flags and cite-checking planned. The 30% is a company benchmark, not independent replication — but the direction of travel is the point: the context the agent can traverse beats the model it runs on. Harvey II's Spaces give each firm a closed, private graph of its own matter data and connections, and the company plans to let firms define their own legal ontology on top of their documents — the graph is becoming the firm's proprietary asset, not just a feature. m:…7684

Legora announced a partnership to bring Blue J's AI tax research into the Legora platform, connected via the Model Context Protocol. "Inline citations in the integration allow lawyers to verify reasoning before using it in advice" — citations as the trust interface, not marketing. Blue J is used by more than 6,000 organizations and has answered over 14 million tax questions to date, and general availability is planned for November. When a domain has an answer key, the winning product exposes the receipts. m:…9093

Evals, clinicians, and narrow models

The Numan case study, published by LangChain, is the week's best object lesson in eval discipline. Numan found that single-turn evaluation fixtures failed to capture the conversational nature of coaching and masked compliance failures. Its answer: evaluation criteria must be authored and signed off by a clinician, and the evaluation harness must be the same graph used in production to prevent drift. Numan's AI coach Nu is deliberately designed as a wellness product rather than a medical device to avoid regulatory constraints on diagnosis — the product scope is itself an eval decision. m:…9986

EvenUp's engineering post shows what happens when the domain work is done first: a fine-tuned Qwen model improved total match rate by 6% over the company's frontier baseline, with a 50% reduction in cost per case and roughly 68% lower p50 latency — for insurance billing extraction. The fine print matters as much as the headline: the model is task-specific and will not generalize to adjacent tasks without additional fine-tuning, and the team calls labeled training data a significant engineering challenge. Narrow beats frontier when the data is curated; the cost is that curation. Under the hood: the team used the Qwen3.5-397B-A17B model as the base, fine-tuned with LoRA on Nvidia B300 GPUs, and optimized inference by removing the vision tower, applying FP8 quantization, and serving with SGLang — the company says self-hosting means managing quantization, serving, auto-scaling, and utilization yourself. m:…8376

The floor beneath it all

Invisible Technologies makes the base-case argument explicitly: "Enterprise AI deployments fail primarily because data is fragmented, inconsistent, and impossible to unify at runtime." True enterprise AI integration requires a semantic layer above the data infrastructure that models can query with governed results, the company argues — and, in its words: "Agentic AI is less tolerant of data fragmentation than traditional analytics because errors compound across multi-step tasks." Organizations with modern data warehouses can finish integration work for a focused use case in weeks; those with legacy system debt typically measure it in months. Every graph, citation, and eval in this story sits on that timeline. m:…4747