Mary knows the matter. Now your legal AI does too.
Mary reads every document in a matter and surfaces what they establish, what they contradict and what's missing. Claude, ChatGPT and Harvey can now start from that.
Ask a litigator how the firm's AI gets used on a real matter and the answer is usually a practical one. They open a chat window, drag in a handful of key documents, paste in a routine prompt, and get something that looks useful back. The next day, on the same matter, they do it again, because the context window has no memory of yesterday's work.
Most firms have generalist AI now. ChatGPT and Claude arrived through the people already using them at home, the practice management system added its own, and they have earned their place. Generalist AI reasons extremely well: the output is always neat, fast and convincing. What it doesn't have is access to the matter.
It helps to split the work on a matter in two, whether a lawyer or an AI is doing it. There's a reasoning layer, which argues, drafts and summarises. Under it sits the fact layer of the matter: what the documents establish, where they conflict, what's ambiguous and where the gaps are. Firms have bought the reasoning layer. Almost nobody has bought the fact layer, so that work is done by hand, usually by the people who know the matter best, and it doesn't carry over to the next task. The first job of any legal tool isn't to draft. It's to miss nothing. Every AI-assisted workflow in a dispute rests on the facts underneath it, which is why that part is worth building once and building properly.
A fact layer isn't a folder or a search box. It's the matter read in full, with each fact traced to exactly where it came from, the contradictions between documents surfaced for action, and the gaps brought to your attention rather than smoothed over. And it's kept. When the next tranche lands, the new material is assessed against what the team has already examined, so the review carries forward instead of starting again.
It also has to be honest about its own limit. The line between a fact and an inference can never be fully automated, and no confidence score can stand in for a lawyer reading the page. A useful tool includes purposeful friction where the high standards of the law demand it: it shows you the contradiction instead of quietly resolving it, and keeps what's missing in plain view. Negative space decides matters, too. A document three witnesses mention and nobody produced is worth more than a tidy summary.
This is where the Model Context Protocol (MCP) comes in, and it's worth saying that the AI companies built it for exactly this reason. MCP is an open standard, published in late 2024 and now supported across the major AI tools, for connecting them to the systems that hold real information. In plain terms, it lets two pieces of software talk to each other, the way an interpreter lets two people who don't share a language hold a conversation. It's what lets a fact layer sit in the middle of the stack, reading from the systems the documents live in on one side and answering the firm's AI on the other.
Mary is built for that middle, and reads every document in the matter up front. It builds the index and the chronologies, raises the issues, and flags where one document contradicts another or where something should exist but doesn't. Through MCP it connects on both sides: to iManage and Smokeball, where the documents already sit, and to Claude, ChatGPT and Harvey, where the lawyers already work, along with any AI a firm builds for itself. The lawyer asks the question where they always have. The AI tool asks Mary about the matter and gets back what's established, what's contested, what was cited but never produced, and who on the team checked it. Then it drafts from that.
Mary isn't asking a firm to move its lawyers somewhere new, which is usually what kills a rollout. Mary connects to the systems the documents and the lawyers already live in, and handles the loading and the set-up. It empowers the AI tools your lawyers are already using every day.
So keep the AI you've paid for. Let it trust Mary to be the fact layer, and Mary will trust you to be the legal judgement. The reasoning layer will keep improving quickly and firms will keep taking advantage of that, which is a good reason not to rebuild the facts underneath it every time something better arrives. The question worth putting to any AI your firm runs on a live matter is a short one: when a lawyer asks it about the case, how carefully is it reading the facts?
