Contract Redliner

Contract Redliner routes all document processing through model endpoints with contractual zero-retention terms — the enterprise tiers of Anthropic or Azure OpenAI, where prompts and completions are not used for model training and are not retained beyond the request lifecycle. That zero-retention data processing agreement is part of every engagement file before any contract touches the system. The redline output and risk summary are generated entirely inside your environment. Nothing leaves to a third-party storage layer unless you've explicitly connected a document management system you already own and control — Google Drive, NetDocuments, iManage. If the firm requires on-premises processing with no data leaving the network perimeter, we scope the deployment accordingly. The build documentation is written for review by the firm's general counsel or ethics committee before go-live, and final sign-off authority on any reviewed contract stays with the licensed attorney. The system flags and summarizes; it does not approve.

The three rules that come up in every law firm conversation on this topic are Rule 1.1 (competence), Rule 1.6 (confidentiality), and Rule 5.3 (supervision of nonlawyer assistance). None of them prohibit AI-assisted review. Rule 1.1 actually cuts the other direction — staying current with technology relevant to legal practice is part of competent representation under the 2012 comment amendment, and several state ethics opinions have since affirmed that AI assistance in document review, when supervised, is consistent with competence obligations. Rule 1.6 is addressed through the zero-retention data processing terms and the access controls described above — confidential client information doesn't flow to a model that retains it for training. Rule 5.3 is addressed by design: the system produces a draft redline and a risk summary; a licensed attorney reviews both before anything is sent or acted on. The attorney is supervising the output, not delegating final judgment. We build the workflow so that the human review step is structural, not optional — it's not possible to route the system's output directly to a counterparty without an attorney touching it first.

There is no fine-tuning of a shared model on your playbook. Your preferred positions, fallback language, and dealbreakers are encoded as structured instructions — a prompt layer and a rules document — that travel with every request the system makes. That means your playbook never becomes part of any shared model's weights, and it's not accessible to other clients. It also means your playbook can be updated at any time: when your standard terms change, when a new practice group is added, or when you want to add or remove positions based on what you've learned from a year of redline logs. You own the playbook document. We configure the system to reference it. If the engagement ends, the playbook document stays with you and the system is decommissioned. Retainer scope includes periodic playbook reviews — typically quarterly — to keep your preferred positions current as deal terms and market standards shift.

Accuracy on straightforward deviations from standard commercial paper — a limitation-of-liability cap that's below your threshold, an auto-renewal clause with a non-standard notice window, an IP assignment that conflicts with your preferred position — is high. The system is comparing text against a defined playbook, and that's a pattern-matching task it handles reliably. Where accuracy degrades is in contracts with unusual structure, heavy defined-term cross-referencing, or clauses where the risk is contextual rather than textual. A clause that reads within-playbook in isolation but creates risk because of how it interacts with a carve-out three sections earlier is harder to catch. So is a contract where the risk is what's missing — an absent limitation clause, for example — rather than a clause that deviates from a known position. The risk summary flags missing standard provisions as a separate check, but that list is only as comprehensive as the playbook we configure it against. Senior review exists for exactly these cases. The system earns its keep on the first pass; it doesn't replace the partner on the close read.

Honest answer: it doesn't catch what it wasn't taught to look for. If a risk stems from a regulatory change that happened after the playbook was last updated, the system won't flag it. If the risk is in the commercial context of the deal rather than the contract language — counterparty creditworthiness, market standard shifts in your practice area, a client relationship that makes a standard position inappropriate — that's outside what any contract-review system surfaces. Jurisdictional nuance is another edge: the system can flag that a governing-law clause selects an unfavorable forum, but it doesn't perform choice-of-law analysis or assess how a specific state court has interpreted a clause your playbook accepts as standard. The right mental model is that the system handles the first pass your most careful junior associate would handle on a good day, every time, with no variance. The attorney who reviews the flagged output is doing the work that requires judgment, experience, and knowledge of the client relationship — which is the work that actually justifies the rate.