Across the 2025–2026 documented deployments at AmLaw 100 firms, agentic AI captures durable value in three of the six billable-hour sub-tasks (document review, precedent retrieval, deposition prep) and produces a net malpractice-risk increase in two (legal drafting submitted as final, citation generation) vs a junior-associate-drafted equivalent at the same time-to-delivery; the remaining sub-task (client communication) is bounded by professional-conduct rules, not technology.

Claim created at publish. Three-value sub-tasks grounded in Allen & Overy / Harvey deployment outcomes (FT, Law.com coverage) and Lexis+ AI / Westlaw Precision vendor disclosures. Two-risk sub-tasks grounded in Mata v. Avianca (S.D.N.Y. Jun 2023) and Park v. Kim (2nd Cir. Jan 2024) sanctions, plus Stanford CodeX hallucination-rate data on legal-specific LLMs. The critical comparator in the claim is vs junior-associate-drafted equivalent at the same time-to-delivery, not vs zero; the claim overstates if read without that comparator. Client-communication bound grounded in ABA Formal Opinion 512 (Jul 2024) and Model Rule 1.4. Malpractice insurance angle: e&o carrier tightening on AI-final-drafted work cited as directional; the synthesis is labelled source:our-estimate in the article body. 90-day review cadence. Trigger conditions: (a) e&o carrier publicly underwriting AI-final-drafted filings — collapses the two-risk sub-task finding for that task class; (b) second sanction wave in 2026 H2 — reinforces or extends it; (c) ABA / EU bar guidance shift post-FO 512 that expressly relaxes or tightens the client-communication bound; (d) Stanford CodeX or comparable independent hallucination-rate re-run on a legal-specific LLM showing sustained <5% hallucination on citation-generation tasks.