A Billion Dollars and a New Model Later: 7 Questions Every Plaintiff Firm Should Ask a Legal AI Vendor
In one week this September, the largest plaintiff firm in the country committed a billion dollars to its own AI platform, and OpenAI shipped a model built for Am Law 200 firms and legal tech vendors. Neither announcement tells a twelve-attorney practice anything useful about what to buy, so here are the questions that do.
What should a plaintiff firm look for in legal AI?
The legal AI market in 2026 sells four different things under one label, and the differences matter more than any feature list. This is what to test before you sign, using the work your firm actually does.
Two Announcements, Neither One About You
On September 17, OpenAI released Astra for Law, a version of its GPT-6 Astra model paired with a legal search index. According to OpenAI, that index covers U.S. caselaw, statutes, regulations, court rules and administrative decisions across more than 230 million URLs. Access is initially limited to selected firms, and the company named Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell Lipton among its development partners. Those are large firms serving a very different segment of the market than the typical plaintiff practice.
Three days earlier, Morgan & Morgan announced it would invest $1 billion in AI and technology over the next decade. The firm has already spent $300 million building MX2, its platform for extracting medical information and generating case documents, and it now has nearly 5,000 monthly users.
So the two loudest AI stories of the month describe a model built for Am Law 200 firms and a platform built inside a plaintiff firm, with a broader commercial rollout to other firms planned by invitation for late 2027. Neither answers the question a six-, twelve-, or twenty-attorney plaintiff practice needs answered: what should we actually test before buying?
What Is an AI Litigation Support Platform?
The phrase covers at least four distinct products, and vendors rarely announce which one they are.
General-purpose models answer questions and draft text well. They have no persistent knowledge of your case, they don't come with a business associate agreement unless you've arranged one, and they have no obligation to tell you when they're unsure. Legal research tools search authority and are genuinely good at it, but research is one stage of a case, not the case.
Point solutions do one job. A chronology tool, a demand drafting tool, an intake tool. They can be excellent at that job, and firms often end up with four of them that don't talk to each other. An AI litigation support platform, properly understood, connects those jobs against one set of case materials, so that what the records analysis found is available to the demand draft without anyone re-uploading anything.
None of these categories is universally better. The question is which one matches the shape of your work, and for plaintiff litigation the shape is unusual: enormous document volume, medical complexity, and a factual record that has to stay consistent from intake through trial.
What Does Legal AI Do With Medical Records?
This is the test that separates products fastest, because it's the work plaintiff firms do most and the work general tools handle worst.
Uploading 1,400 pages of records from five providers and asking for a summary produces a summary. What a case needs is a treatment timeline that survives cross-examination: dates, providers, diagnoses, gaps in care, and the inconsistencies between what one chart says and what another says three months later. A medical malpractice case lives or dies on whether the deviation from the standard of care sits visibly in that timeline.
So the demo question isn't "can you summarize records." It's: show me what you do with a duplicate record set where the second copy has different handwritten notes. Show me a treatment gap. Show me what happens when a provider's records arrive out of order. Bring a real file, redacted if you need to. Vendor sample cases are chosen because they demo well.
George Palaidis, a Florida solo handling auto, bicycle and pedestrian cases, reports cutting medical record review time by 75% on files running 100 to 200 pages. The detail worth noting isn't the percentage: he also found the analysis caught items on a crash report that he had missed on his own first read. That's the test. Not whether it's faster than you, but whether it sees what you'd have seen on a good day.
Where Did That Answer Come From?
An answer you can't trace is an answer you can't use. Every output should link back to the page it came from, and you should be able to click it in the demo, not be told the feature exists.
In 2025, a federal judge in Wyoming fined three lawyers after a filing in a Walmart case cited nonexistent authorities that had been generated by AI. It's worth knowing not as a story about one firm, but because it happened at a well-resourced practice with experienced litigators. Verification isn't a discipline problem that careful lawyers avoid. It has to be part of the workflow.
ABA Model Rule 1.1, Comment 8 already frames technology competence as part of the duty of competence, which means the obligation to check the work doesn't transfer to the vendor. Ask how the tool makes checking fast, because a citation feature that requires opening a separate PDF and scrolling is a feature nobody uses twice.
Data handling belongs in the same conversation: whether your files train the model, how protected health information is stored, and what the vendor can see. We've covered that ground separately in what to look for in secure legal AI, so the short version here is that "we don't train on your data" should be a written contractual term, not a sentence on a slide.
One Task, or the Whole Case?
Here's the question that reorders most buying decisions: after the tool finishes, how much work is left before the case moves?
A tool that produces a chronology has completed a task. The paralegal still pulls the specials, still drafts the demand, still cross-references the discovery responses against what the records showed. A system that carries the same case materials through those stages removes handoffs rather than steps, and handoffs are where plaintiff files actually stall.
This is the practical meaning of agentic AI for plaintiff lawyers, and it's less exotic than the marketing suggests. Anytime AI's conversational interface, Talk to Teddy, works against the case file rather than a pasted excerpt, which is why it can answer a question about one provider's notes and then use that answer in a draft. We've written a fuller explanation of agentic AI in plaintiff practice if you want the mechanics.
Lorraine Gingery runs Lorraine Law, PC on her own, and her demands used to take four hours or more each. With record search and drafting working against the same case file, that dropped to under 30 minutes, and she can query a 456-page record set for every mention of a body part or procedure instead of paging through it. She grew her docket without hiring anyone, which is the version of this that matters for a solo: the constraint that lifted wasn't drafting speed, it was the ceiling on how many cases one person can carry.
Scope has a cost, though, and the honest version of this argument includes it. A connected system takes longer to configure than a single-purpose tool, and it earns its keep only if your firm actually uses more than one part of it.
What Legal AI Still Can't Do
A vendor who won't answer this question is telling you something.
Anytime AI doesn't currently handle image exhibits, so document compilation is in scope and exhibit work involving images isn't. No AI decides whether a deviation from the standard of care is actionable in your jurisdiction, and none of them knows what a particular adjuster has settled comparable files for. Output still requires attorney review before it goes anywhere, and any vendor promising otherwise is describing a liability, not a feature.
The realistic claim is narrower than the category's marketing. It takes over most of the reading and the first draft. The judgment and the client relationship stay with the lawyer.
Questions to Ask on the Demo Call
Bring these seven, in this order, and use your own file:
Run this record set. What did you find that I'd have caught on page 400?
Show me a citation, clicked through to the source page, live.
Is "your data is never used for training" in the contract, or on the website?
What happens to this analysis when I move to the demand? Does anything get re-uploaded?
Does it connect to the case management system we already run, and who does that work?
What does this not do?
Who at your company do I call when an output is wrong?
If a vendor handles all seven comfortably, the product is probably real. If the demo only runs on their sample case, you've learned something too.
Final Thoughts
The announcements this month don't change what a plaintiff firm should evaluate, and they're a poor guide to it, because the money is flowing toward practices that don't look like yours. What changes a firm's outcomes is duller: whether the tool holds the whole case, whether you can check its work in one click, and whether it says plainly what it can't do.
Don't evaluate this on someone else's case. Pick the two files on your desk you least want to read, and bring those.
FAQs
What's the best legal AI for plaintiff firms?
There's no single answer, because firms differ on case volume, practice area, and how much of the workflow they want covered. The useful comparison is which product handles your medical record depth and connects to the next stage of the file, tested on your own materials rather than a vendor demo set.
What should I ask a legal AI vendor before buying?
Ask them to run the demo on your own case materials rather than their sample set. Test medical record analysis, whether every output links back to a source page, what the contract says about training on your data, and what happens to the analysis when you move to the next stage of the file.
What are the risks of using ChatGPT for legal work?
General-purpose models have no case-specific grounding, can produce citations to authority that doesn't exist, and don't come with the data-handling terms plaintiff work requires for protected health information. They're reasonable for brainstorming and poor for anything filed.
Does AI replace paralegals at plaintiff firms?
Most of the gain is in reading time and first drafts. Judgment and client contact stay with the lawyer, and firms that see real gains typically move paralegal time toward case strategy and provider follow-up rather than reducing headcount.
How do I know if a legal AI tool is HIPAA compliant?
Ask for the contractual language on data handling and model training, not a compliance badge. A vendor should be able to tell you where protected health information is stored, who can access it, and whether a business associate agreement is available.
Can AI build a medical chronology I can use at deposition?
It can produce a dated treatment timeline with source links, which is what makes it checkable. Every entry still needs attorney verification against the record before it's relied on in testimony.
Get Started
Ready to go deeper — and safer?
See how Anytime AI gives plaintiff firms the strategic edge
and the security their clients deserve.