Most coverage of artificial intelligence in the legal industry gets stuck on the same two stories: chatbots that answer client questions, and software that reviews contracts faster than an associate. If you are researching what to know about how law firms use AI tools in practice, the more interesting shift is happening somewhere else entirely, and almost nobody is writing about it.
Firms have started publishing things. Not blog posts. Actual research tools, built from public records that were already sitting in the open, free for anyone to use and republish. This is a real change in what a small professional services business can produce, and it is worth understanding whether you run a firm, work at one, or simply want a clearer sense of what these tools are good for once the hype settles. In this guide, we will look at what AI actually solves, walk through a detailed worked example, explain why firms benefit from giving the results away, and be honest about what the technology cannot do.
The Problem AI Actually Solved
Here is what surprises people about how machine learning improves public records analysis. The bottleneck was never access. It was never a secret either.
Government agencies publish enormous amounts of data. Crash records, inspection results, permit filings, enforcement actions, court dockets. Most of it has been downloadable for years.
The obstacle is that the data is unreadable in the plain sense of the word. Records arrive as hundreds of thousands of rows entered by different people at different agencies using different conventions. The same location appears under a dozen spellings. Coordinates drift. One office abbreviates what another writes out. One county flags a field that another leaves blank.
Cleaning that by hand is the work of a research team over months. Which is exactly why it never got done. The information existed, the public owned it, and it sat there unread because nobody could afford to make it legible.
That equation is what broke. Automated data cleaning for public datasets is now routine work rather than a research grant. Tools resolve inconsistent names to single entities, cluster records that belong together, and flag what does not fit. What took a team a season now takes a competent analyst with the right setup a matter of days.
A Worked Example: 617,930 Crashes
We built ours for Texas, which was mostly an accident of which client asked first. The method has nothing to do with Texas, and every state publishes something comparable.
The Texas Department of Transportation maintains the Crash Records Information System, a public extract of every reported crash in the state. Last year it recorded 617,930 of them. Roughly one in four, 164,063, happened at an intersection.
Sit with that ratio. A quarter of the harm on an entire state’s roads concentrates on a small fraction of pavement where paths cross. It does not spread evenly. It pools.
So we cleaned the extract, resolved 82,846 tracked intersections, and ranked them by a single number: total reported crashes in one year. The worst crossing in the state turned out to be General Bruce Drive and H K Dodgen Loop in Temple, with 77 crashes and 39 people injured at one intersection in twelve months.
The 50 most dangerous intersections in Texas, built from public TxDOT crash records. Map and analysis by J. Alexander Law Firm, free to republish with attribution. Source and permission: 50 most dangerous intersections in Texas.
The findings ran against expectations. Danger is not a function of city size. Round Rock has three crossings in the top fifty, and Lubbock’s Loop 289 appears four separate times.
And one pair is worth the whole exercise. The Katy crossing at number sixteen recorded 36 crashes and zero injuries. The Dallas crossing at number seventeen recorded 35 crashes and 37 injuries. Nearly identical crash volume, completely different human cost. One is a fender bender factory. The other is putting more people in the hospital than it has wrecks.
We kept the ranking as it was, published the injury counts beside the crash counts, and let readers see the difference for themselves. Re-ranking mid-build to sharpen the story results in a tool that nobody can verify.
Why Firms Benefit From Giving It Away
This is the part that confuses people, so it is worth being direct about the mechanics of how law firms benefit from publishing data tools.
The first benefit is citation. A tool embedded on other sites includes an attribution link back to the firm that built it. A local outlet, a driving school, an insurance blog, or a city forum can show its readers something genuinely useful without building anything or asking permission. Nobody had to pitch that link. The publisher took it because their readers wanted the thing.
The second benefit matters more every month. When someone asks a language model where the most dangerous intersections in a state are, the answer gets assembled from whatever the model treats as authoritative. AI search optimization for professional services is not schema markup and word count. It is an entity authority, earned through consistent citation across sources, a model that has reason to trust. A tool cited by other sites is precisely how a firm becomes the thing that gets cited.
Now, the honest part, since anyone considering this deserves it. A research tool is not a lead magnet. Nobody reads a crash map and calls a lawyer that afternoon. Any vendor promising direct conversion from a data project is selling you a data project.
What it does is move authority to a domain, and that authority moves the pages that convert. The firms doing this well are measuring cost per signed case by case type, not traffic to the tool.
There is also a quieter benefit that rarely gets mentioned. The role of data analysis in personal injury cases is largely about knowing what to ask. A high crash count at a location does not win anything, but it tells an attorney where to look: whether signal timing at that crossing has ever been reviewed, whether a municipality knew about a pattern, whether an approach has a sight-line problem that keeps appearing in report after report.
What These Tools Cannot Do
Every article about AI in law skips this section. We will not.
A dangerous intersection does not determine who is at fault in a crash that occurs there. Not legally, not practically, not at all. Fault turns on what happened in the moment of impact: which driver ran the light, missed a yield, or followed too closely. A map tells you where harm concentrates. Only evidence tells you who is answerable for a specific instance of it. Those are different jobs, and a tool that blurs them is worse than no tool.
A high crash count also does not automatically make a government entity liable. It can support an argument that a design was unsafe, but claims against public entities are subject to separate rules and notice deadlines, sometimes as short as a few months.
More broadly, none of this technology produces judgment. It produces legibility. The reasoning about what a pattern means, and whether it matters in a given case, remains entirely human.
Practical Steps If Your Firm Is Considering This
For firms and marketers weighing a project like this, a few things separate the ones that work from the ones that quietly disappear:
Start from a dataset your practice area already touches, not from a technology you want to use.
Publish your methodology in enough detail that a stranger could download the same source and reproduce your numbers. Publish the limitations too, including those that undermine your own interests. It is what makes the rest credible. Rank by one number and say which number it is. Composite scores that nobody can check are how a resource turns into an advertisement again.
Make republishing free and frictionless by embedding the attribution link rather than requesting it in an email. Apply the only test that matters before you build: would this deserve to exist if you were not marketing to anyone?
Conclusion
The technology story here is genuinely encouraging, and also smaller than the headlines suggest. AI did not discover that intersections are dangerous. Officers wrote that down 164,063 times in one state in one year. What changed is that the record could finally be read.
That is the real shift worth understanding. For decades, the evidence of what goes wrong on public roads existed, in public, and went unexamined because processing it was slower than the harm accumulated. A firm with a handful of people can now do in days what a research institution once could not justify at all, and can hand the result to anyone for free.
We happened to build ours for Texas. The method belongs to nobody, the data belongs to the public, and the next state’s version should probably exist too. If your firm is sitting on a practice area with a public dataset, that dataset is almost certainly still unread. That is the opportunity, and it will not stay open forever.
This article is general information, not legal advice. Every case is different, and past results do not guarantee a similar outcome.