What actually works, and the rule that decides whether it helps you.
Sixty-nine percent of legal professionals use these tools and nine percent of firms have a written policy. The gap is where the trouble is. Here is what earns its place in a plaintiff practice and what does not.
Escalated Two handwritten pages it will not guess at. Flagged for a person rather than filled in.
What it produces
- Reading records at volume
- The clearest win. Dates, providers, gaps and billing pulled out of a thousand pages and cited back to the page, so a person checks rather than types.
- Answering an inquiry at 2am
- Qualification and routing need no judgment and no sleep. This is the least controversial thing on the list and the most neglected.
- Assembling what you send out
- Demands and records summaries built from your own prior work. The gathering is automatable; the argument is not.
- Finding what is absent
- The provider in the billing index with no records attached. Absences are what people miss, because there is nothing on the page to notice.
Limits
What it does not do.
Stated up front, because you find these out in week two anyway and it is better if you hear them from us.
- Legal research you do not verify. This is what the roughly 1,490 tracked sanction decisions have in common, and none of them involved a careless firm — it involved a plausible sentence nobody traced to a source.
- Anything you cannot check cheaply. A tool that is right 97% of the time and gives you no way to find the wrong 3% has moved the work rather than removed it.
- Client-facing advice. Not because it cannot draft it, but because the failure mode is a confident wrong answer and the exposure is yours.
- A firm-wide program. One workflow, shipped and adopted, changes what people believe is possible. Three, nearly done, changes nothing.
The question everybody asks
Is it safe to put client records through an AI tool?
It depends on an arrangement you can check, not on the tool. Three things get conflated: whether your data trains a model, how long it is retained after processing, and where it physically runs. Confirming one is not confirming the others. The strongest answer is that the software runs in your own cloud on your own keys, so the file never sits in a vendor’s system at all.
What it costs
- Intake Rebuild$6,500 Two to three weeks
Your site gets inquiries. Some of them never get answered in time.
- Search Foundation$9,500 Three to four weeks
You buy all your cases. You want a source you keep.
We do not promise rankings. Nobody honestly can. You are buying the setup and owning it, not a spot on page one.
- Case Pipeline Most start here$15,000 – $25,000 Four to six weeks
You spend five figures a month on cases you do not own.
- Records and document automation$60,000 – $90,000 Six weeks
Signing cases is not your problem. Working them is.
Before Every one of these starts with a free 30-minute call. The biggest one starts with a paid week first, the Workflow Audit, at $7,500. You get a working prototype and a fixed quote out of it, and you can stop there.
After Keeping it working costs $1,000 – $1,500 / mo. You can cancel any time, and nothing switches off if you do.
Always Fixed price. You own the code. No monthly fee to keep using it.
Questions a partner asks
What is AI actually good at in a personal injury practice?
Reading records at volume, answering and qualifying inquiries out of hours, assembling demand packages from material the firm already has, and finding what is missing from a production. All four are gathering rather than judgment.
What are the risks?
Confident wrong answers, and output nobody traced back to a source. Roughly 1,490 court decisions now involve AI-hallucinated material. What they share is not carelessness — it is that a plausible sentence reached a filing without anybody checking whether the thing behind it existed.
Do we need an AI policy first?
If people at the firm are already using these tools, the firm has an AI practice whether or not it has a policy. 43% have none and no plans. A short approved list with a real process beats a ban nobody follows, because a ban moves usage onto personal accounts where you cannot supervise it.
Should we buy a product or build something?
Buy if a product fits your process without you reshaping the firm around it, which is more often than vendors admit. Build when the value is in connecting systems you already run — the layer no product company can sell you, because it differs in every firm.
How do we evaluate a legal AI vendor?
Ask whether every statement traces to a source, whose model is underneath and what their data terms are, how long data is retained, and what happens if you stop paying. A vendor who answers those in writing is workable even where some answers are imperfect.
Will this replace paralegals?
At the firms we have seen discuss it, the constraint is capacity rather than payroll — three quarters run at or beyond it, so the work that was queued gets done instead of declined. But a firm should be honest about which it is doing, because the wrong conversation handled badly poisons adoption.
Not ready to book a call
Send us five pages of a record set. We will send back what we found in it.
Five pages is enough, redacted however you like. You get a short video back within two business days showing what a chronology would surface from it: the dates, the gaps, the things worth knowing before the other side finds them.
- Send five pages of a real file. Redact whatever you like first.
- We run them and record what comes out, including what it misses.
- You get the video within two business days. If there is nothing worth showing, we say so.
Law firm intake software AI demand letter software Medical chronology software