The evidence has a chain of custody. The analysis of it does not
Documents pushed from eDiscovery into an AI layer keep their chain of custody. The analysis that comes back has only a vendor log, and litigation is where the other side is paid to find that gap.
Thomson Reuters and Everlaw announced plans at ILTACON to bring Everlaw documents into CoCounsel Legal in bulk. Emily Colbert, who heads CoCounsel Litigation, wrote it up on the Thomson Reuters blog under the title "Legal AI is moving closer to the evidence," and she's right that this is the direction the market has to go. A litigator's question is concrete: whether this deposition transcript squares with that email, and whether the two of them together support the position going in front of a judge. An AI that can't see the record is answering a different question than the one the lawyer asked.
The integration is the right move. What's at risk of getting lost in what Colbert calls the handoff is the thing that makes the eDiscovery side of it work.
Everlaw, Relativity and the rest of the eDiscovery market run on a discipline the rest of legal tech has never had to develop. Every document has a Bates number. Files are hashed at collection. There is a production log that says what went to the other side and when, a privilege log that says what was withheld and why, and a chain of custody that can put a custodian on the stand to explain how a file got from a laptop to a review platform without changing. That discipline exists for one reason. Discovery is the one part of legal work where an adversary has a procedural right to ask how you got what you got, and two decades of that pressure built the record.
Everlaw's own announcement describes the integration as forthcoming and says Everlaw continues to serve as "the governed system of record" for the evidence. That is the context Colbert's post says should travel with the documents into CoCounsel. Fine. The Bates numbers will travel. The hashes will travel. The question for legal and product teams is what travels back.
A team in Everlaw tags 400 documents as hot. They push them into CoCounsel Legal. CoCounsel reads them against Westlaw and Practical Law, drafts the fact section of a summary judgment brief, and the lawyer edits it and files it. Eight months later opposing counsel moves to compel, or a judge asks at a hearing, or the client's new general counsel asks after a bad result: which documents did the analysis actually rest on? Were all 400 read, or did retrieval surface 60 of them and the argument never touched the rest? What did the model pull from Westlaw, and what did the lawyer change before filing?
The documents can answer for themselves. They have a chain of custody. The analysis that came out of the AI environment has whatever the AI vendor logged, in a format the vendor controls, retained for as long as the vendor's retention policy says, and produced on the vendor's terms. I wrote in July that logs were built to help engineers debug a system and were never designed to hold up when a decision is challenged in a courtroom. Nothing in this announcement suggests that has changed, and I don't say that as a criticism of Thomson Reuters. Neither company's post says anything about a record of what CoCounsel does with the documents once they arrive. The evidence arrives in the AI layer with its custody intact, and the brief leaves with a vendor log.
Colbert's post comes closer than anyone to naming that gap. "The lawyer is still responsible for the argument. Still responsible for the citation. Still responsible for understanding whether the evidence supports the conclusion." Every word of that is correct, and it is also the reason the gap matters. Under Rule 11, the signature on the brief certifies that the factual contentions have evidentiary support or, if identified as such, will likely have it after further investigation or discovery. That certification was written for a world where a reasonable inquiry meant a human read the documents and a court could ask that human what they read. When the reading happened inside a model and the human reviewed the output, the inquiry is real and the account of it is thin. The lawyer can say "I checked it." The lawyer cannot yet show a court what the check consisted of, in a form the other side would have to accept.
The moment evidence starts flowing into AI environments in bulk, the AI environment becomes part of the discovery record, whether or not anyone designed it that way. Which documents were loaded, which were used, which prompts ran and what came back are going to be requested. Some of it will be privileged work product. Some of it won't be, and the fight over which is which will go better for the side that can show what went in, what the model used and what the lawyer changed, than for the side holding a vendor log.
Thomson Reuters calls its standard Fiduciary-Grade AI and defines it as AI built so professionals can "review, verify, and ultimately stand behind the work it helps produce." You can stand behind work all day. In front of an adversary you need something to hand over, and it has to come from somewhere they can't argue was edited after the fact. The eDiscovery vendors solved that problem for documents a long time ago, under pressure, because they had to. The AI layer is about to get the same pressure. Litigation is the one domain where the other side is paid to find the gap.
The first motion to compel a CoCounsel prompt history will do for the AI layer what the 2006 amendments did for email. Somebody is going to be the test case.