What I'd Ask Before Buying a Legal AI Tool

This week I put LegalTech Hub's generative AI map from early 2025 next to the September 2026 version. Their count went from 505 generative AI solutions for legal in March 2025 to 1,416 in September 2026, from 1,117 providers. You can easily see why law firms are absolutely overwhelmed by vendors right now.

Image: Litem, using figures from LegalTech Hub's GenAI Growth Trajectory in Legal Tech 2025–2026.

I can't tell you which of those tools your firm should buy. You know your firm, your clients and your attorneys better than any vendor does. What I can tell you is what's inside these tools, because we build one, and knowing that changes how the map reads. I've also put the questions I'd ask into a one-page checklist at the end.

Most of the Map Runs on the Same Models

Most of the logos on that map are built on the same handful of foundation models. That isn't a criticism, since we build on them too. It does mean the model usually isn't what separates one tool from the next, so the more useful question about any part of your stack is what the tool does around the model.

The Real Work Is What the Model Is Given

A model only knows what it's handed. Most of our engineering effort goes into pinpointing the right documents, facts, testimony, and other information in a case and engineering solutions around the strengths and the limitations of models, and I'd guess that's true at any serious legal AI company. Two tools on the same model can give different answers to the same question because one found the right transcript page and the other didn't. You can't see that layer in a demo, but it's most of what you're buying.

Demos Are Small, and Matters Aren't

A demo runs on a tidy sample. A real matter is years of productions, pleadings, exhibits and testimony, and a tool that looks sharp on a few hundred pages can start missing things when the matter is a hundred times that size. We built Litem for the set your team actually works from after first-pass review, however large that gets, because a case can turn on one document and nobody knows in advance which one it will be.

A Citation Is a Design Choice

Whether an AI system's answer can show you the exact page it came from is decided when the product is built, not by the model. Every answer in Litem points back to the document and page, and testimony points to page and line, because it's the attorney's name on the filing and they need to be able to check it. When an AI tool can't show its sources, it's usually because it wasn't built to.

The Model Should Follow the Work

New models come out every few months. Some are better at legal reasoning and some are far cheaper to run. A vendor should be choosing models to serve your work, a frontier model where the reasoning matters and a cost-effective one where it doesn't, and changing that mix as the field moves. A tool tied to a single model is betting your workflow on that model staying the best, and in this market that's a hard bet to win for long.

When the Help Arrives Matters

A lot of AI in legal works after the fact, summarizing the transcript or reviewing the production. Some work only pays off in the moment. In a deposition, a contradiction found the next day may be somewhat useful, and one found while the witness is still under oath is a question you can ask in the moment. We're able to flag those during the deposition, and that's a different kind of engineering from summarizing afterward. So when you look at where a tool fits in your stack, it's worth knowing whether its value lands before the work, during it or after it.

What This Means for the Map

When I look at that map, I see a lot of tools on the same few models, and the differences that matter sit underneath. They come down to what each tool can find in your documents, whether it shows its sources, how it chooses its models, and when in the work it helps. Those are the things I'd want to know about every tool in a stack, including ours.

Download the 12-question checklist to use before your next legal AI demo (PDF)

The map will be bigger again next year, and the models underneath it will have changed more than once. The tools worth keeping are the ones that get better when they do.

Mitchell Kossoris is the Co-Founder and CEO of Litem.