How Marcella compares.
Specific differences. No marketing language.
CoCounsel Legal (Thomson Reuters)
Seat license, plus model spend, plus data processor review, plus integration, plus a third-party custodian
Thomson Reuters announced a Model Context Protocol integration connecting Claude to CoCounsel Legal on May 12, 2026, and has said the next generation of CoCounsel Legal is being rebuilt on Anthropic's Claude Agent SDK. Thomson Reuters also built its own model, Thomson, with its first integration launching August 2026 inside Tabular Analysis. Three parties can sit in one workflow.
Approving a new AI data processor is a 3 to 6 month exercise at a mid-size firm. When an answer is wrong, the escalation path is not obvious: content, model, and orchestration are not the same vendor.
Marcella runs inside the Microsoft tenant your firm already contracts for. No new data processor to approve, no new custodian holding your prompts. Citation treatment engine crosses every publisher. Six-item Daubert attestation on every citation. Why the custodian question is the expensive one.
Clio + vLex
$1B acquisition · Nov 2025
Clio users search vLex's 110-country database. They don't get Prosser, Wright & Miller, or Collier without separate subscriptions. Marcella crosses every silo from the public record.
Westlaw / Lexis
Traditional platforms
Each is a silo. You pay per silo, per year. Marcella's citation treatment engine crosses every silo by mapping citation relationships from 10M+ public opinions. Full breakdown.
Public AI (ChatGPT, Claude, Gemini)
Used directly for legal research
They generate text, not retrieve from a legal corpus. That is the Mata problem: citations that look real and were never decided. Marcella retrieves. Six-item Daubert attestation. The architecture prevents fabricated citations.
One page per competitor.
Sourced pricing, an honest account of where each one still wins, and what Marcella does not do.
The number on the invoice is not the number.
A seat license is one line. The cost of putting a third-party AI platform into a law firm has five, and the last one has no price on it until it lands.
| Cost line | What it is |
|---|---|
| 1. Seat license | The published or quoted per-user rate. The only line most firms budget for. |
| 2. Model spend | Agentic workflows consume tokens per run, not per seat. Long-horizon document work is metered separately from the license on most architectures. |
| 3. Data processor review | A new AI vendor in the data path is a new processor. Approval runs 3 to 6 months at a mid-size firm, and the hours are partner and general counsel hours. |
| 4. Integration and support | Content, model, and orchestration layers can be different companies. When an output is wrong at 6pm before a filing, the firm has to know which one to call. |
| 5. Custody of the record | Prompts and outputs are electronically stored information. If a third party holds them, the firm is not the custodian of its own work product. No line item. See below. |
The custody problem
The common reading of United States v. Heppner is that consumer AI destroys privilege and enterprise AI is the safe harbor. That reading is correct as far as it goes. Judge Rakoff's opinion turned on consumer terms permitting training and disclosure, and on the absence of attorney direction. Change either fact and the analysis changes with it.
The part that gets missed is that a confidentiality term is a contract between a firm and a vendor. A preservation order is not. It runs against whoever holds the data.
In the OpenAI copyright litigation, a magistrate judge ordered OpenAI to preserve all output log data including conversations users had already deleted, and a district judge later compelled production of 20 million anonymized ChatGPT logs. Every one of those users was a non-party. They had terms of service. The terms did not reach the order, because the order was not addressed to them.
That is the exposure, stated precisely. It is not that a frontier model is discoverable. It is that when a third party holds your prompts and outputs, your firm is not the custodian of its own work product. You cannot assert a privilege over a record you do not hold, you cannot guarantee a litigation hold you do not control, and a subpoena or preservation order aimed at the vendor never reaches your desk before it is executed.
Marcella's answer is structural rather than contractual. The retrieval and the matter memory sit inside the Microsoft 365 tenant the firm already contracts for, and on BOYA tiers inside the firm's own Azure tenant. No additional custodian is created, because no additional party holds the record. The firm asserts its own privilege over its own data, on infrastructure it already controls.
This is a design argument, not a legal opinion, and no court has ruled on the architecture either way. What is on the record is narrow and worth reading directly: the preservation and production orders, the Heppner privilege ruling, and the sanctions decisions are documented with primary sources at heppner-problem.com.