Pilot5 Legal

Pilot5 Legal is a deliberative AI platform built for legal work. Rather than returning a single fluent answer, it puts one legal question in front of five independent frontier AI models. Each model forms its own position without seeing the others, then the models critique one another, answer objections, and revise under pressure before the panel converges on a structured recommendation. The product is aimed at lawyers and legal teams who need an answer they can inspect, challenge, and verify rather than simply trust, and it is organised around five concrete legal workflows: statutory review, contract understanding, settlement range analysis, statutory authority lookup, and firm playbooks. The product exists because a fluent answer is easy to produce, while a defensible one needs the source, the opposing view, and a record of how the conclusion was reached. A single answer leaves things out: the adverse case is not argued, the authority behind a claim cannot easily be checked, and the limits of the analysis stay invisible. Legal work that will be challenged requires a different standard of answer, and that standard is the design brief behind every part of the product. Pilot5 is described as relaunching for legal work, adding primary-source research, citation verification, contract analysis, and transparent reasoning so that lawyers can inspect, challenge, and verify AI output. The premise is that authority, challenge, and uncertainty should stay attached to the conclusion rather than being stripped away. The statutory review workflow checks a separation agreement against selected statutory requirements and shows the provision behind each finding, so a lawyer can find what the agreement is missing and see the authority behind it, with an annotated agreement delivered in Word. Contract understanding is built for getting oriented in an unfamiliar contract: it produces a clause-by-clause map explaining what each numbered clause does while quoting the source wording verbatim, and it identifies skipped material rather than silently passing over it. The stated goal is orientation without losing the original text, so that reading a new agreement starts from a structured map instead of a blank page. The settlement range workflow pressure-tests both sides of a dispute and exposes a settlement range that costs each side something, using five independent positions, statutory claims that are checked, and preserved dissent. Statutory authority lookup retrieves the enacted text: a lawyer pastes a citation and receives the publisher's own words with no generative model involved, handling up to 25 citations at a time and returning the official source and retrieval date, under the explicit rule that no text is generated and anything unresolved stays unresolved. The firm playbook workflow makes the firm's position repeatable by adding approved firm positions to the statutory baseline used in future reviews; human approval is required, the statutory baseline remains in place, and approved positions are reused in later reviews. The defining difference is that the models do not merely answer, they answer one another. The deliberation runs in four stages: blind first analysis, anonymous cross-critique, reasoned revision, and synthesis with a minority report. In the published example, the panel includes The Architect (structure, benchmarks, operational logic), The Counsel (evidence, nuance, legal clarity), The Strategist (long-range value and trade-offs), The Engineer (feasibility and failure conditions), and The Contrarian, who is required to make the strongest case against consensus. In a worked Chapter 7 preference dispute, the Engineer initially proposed building and litigating only after a six-week forensic payment audit and otherwise settling at 35 to 45 cents; the Contrarian challenged the legal premise by warning not to assume the affirmative defenses and instead attacking the trustee's prima facie case; the Counsel argued that a prima-facie-only attack is not a standalone strategy and called for quantifying ordinary-course and new-value evidence; and the Architect revised the approach so that payment timing, new-value offsets, and defense cost became settlement gates rather than a fixed dollar anchor. The final view follows that argument; it is not an average of five first drafts. Every deliberation is designed to remain reviewable. The authority sits beside the claim: primary law is fetched from the publisher's text, and analytical inference is marked separately so that confidence never masquerades as authority. The opposing view is part of the work, and the strongest dissent stays visible rather than being erased. The reasoning remains inspectable, with every source, finding, challenge, and limitation travelling with the outcome, and the lawyer remaining the decision-maker and the reviewer of record. Accepted, rejected, and overridden findings carry into the review output. The product states plainly that silence never means clearance: unresolved and unexamined issues remain explicit, and resolved citations link to publisher text while unresolved ones stay unresolved. Because client matters carry their own duty of confidentiality, Pilot5 Legal publishes the data rules that govern every review. Client data is never a training asset: there is no cross-account use and no model training. Data is encrypted with AES-256 at rest across tables, WAL, and backups, and encryption cannot be disabled. A right to erasure provides hard delete across all tables, and retention is published per data class. Frontier models run on zero-retention endpoints, and models without one are excluded. Identifiers are pseudonymized before inference, tokenized, then re-identified in the response, and traffic uses TLS 1.3 with AES-256-GCM and forward secrecy. A Zero Data Retention mode offers ephemeral processing with no outcome logging; every sub-processor is disclosed with role, region, and transfer basis; in-region deployment with BYOK and sovereign hosting is available on demand; and a DPA under Article 28 can be put in place under NDA. Every review is scoped to United States law, no other jurisdiction is consulted, and the same disclosure travels with every deliberation record. The practical benefit is that the lawyer receives an answer that arrives with its own audit trail. Because the strongest opposing view travels with the result, the thinking has already been stress-tested before it is relied on. Because sources and inference are separated, a reviewer can see which parts of a conclusion rest on authority and which rest on reasoning. Because unresolved issues stay visible, the absence of a finding is never mistaken for a clean bill of health. And because human accept, reject, and override decisions remain on the record, the output fits a review workflow rather than replacing it. Cost transparency is part of the proposition as well: the product cites fifty cents for a statutory review and about four dollars to pressure-test a settlement position. Pilot5 Legal runs where lawyers already work: deliberations can be launched from Claude, ChatGPT, Cursor, Perplexity, Mistral Le Chat, and Microsoft Copilot Studio through an MCP setup, and data can be brought in from sources including Slack, GitHub, Notion, Google Drive, Jira, Confluence, Linear, GitLab, Stripe, Zendesk, Sentry, PostgreSQL, Zotero, Finnhub, EODHD, Twelve Data, Indian Kanoon, OpenCorporates, Docket Alarm, UniCourt, Clio, SharePoint, and NetDocuments. Authority lookup and the firm playbook are free. Every plan includes all five legal workflows and differs only in monthly credit allowance: contract understanding costs 0.2 credits, statutory review 0.5 credits, and settlement range roughly 4 credits. Paid plans are Starter at $29 per month for 30 credits, Pro at $79 per month for 90 credits, Expert at $149 per month for 200 credits, Master at $299 per month for 450 credits, and Business at $999 per month for 1,750 credits. Unused credits roll over for 12 months, the credit estimate is shown before a paid run, any unused reserve is refunded when the run completes, and there are no daily caps or feature-gated plans. Pilot5 Legal's core value proposition is that it treats the record, not the answer, as the product. Five independent models argue the question, one is required to argue against consensus, the authority behind each claim can be checked, and the dissent, limits, and human decisions stay attached to the result. For legal work that will be challenged, that combination turns AI output from something a lawyer has to trust into something a lawyer can review, challenge, and stand behind.