Legal AI Moat Shifts to Workflow
Spellbook
The real moat in legal AI is shifting from having a slightly better model to owning the workflow, the data, and the place where lawyers already work. Spellbook wins today by living inside Microsoft Word, comparing language against more than 10 million contracts, and letting teams encode negotiation rules through Playbooks, but Harvey already abandoned its own legal tuned model once frontier models caught up, showing how fast model level advantage can disappear.
-
Spellbook is already pushing beyond pure text generation into workflow software. Review flags missing definitions and bad cross references, Benchmarks shows whether a clause is market standard, Associate handles multi document edits, and Playbooks applies company specific fallback rules to incoming contracts. Those features are harder to copy than drafting alone.
-
Harvey is the clearest proof that legal fine tuning is not a permanent moat. By June 2025, frontier reasoning models had improved enough that Harvey dropped its proprietary legal model and moved to orchestrating multiple general models and tools inside legal workflows instead.
-
The companies with the strongest protection are building around proprietary data and system level control, not just prompts. Luminance trains on more than 150M legal documents and sells a broader document system that spans review, redlining, and negotiation, while DraftWise mines a firm's own deal history for clause suggestions.
Going forward, pricing power will come from becoming the default operating layer for legal work inside Word and across contract systems. That favors products that bundle drafting with review, benchmarks, precedent content, playbooks, and multi document automation, because once general models reach parity, customers will pay for embedded workflow and trusted firm specific context, not the raw model itself.