AI document review means giving an AI a stack of case documents and getting back three things: a structured summary, a chronology with dates, and a list of what’s missing. It takes minutes. The work it replaces takes a paralegal most of an afternoon.
There are three tiers, and picking the wrong one is how firms waste money. A general assistant like Claude covers small volumes, a few hundred pages per matter. Purpose-built legal platforms cover discovery-scale review, tens of thousands of documents with privilege screening. And when your document set is the same on every single file, you build your own.
I run the third tier at my lemon law firm. More on that below.
The personal injury case: medical records and chronologies
If you do PI work, this is the highest-value AI task available to you today, and it is not close.
A medical record review is mechanical. Read hundreds of pages of provider records, pull out every visit with a date and a provider name, total the billed amounts, note where treatment stopped and restarted, and flag every mention of a prior injury before opposing counsel finds it first. Then assemble all of that into the chronology that goes in the demand package.
AI is good at every one of those steps. PareIT and EvenUp built entire companies on it, and both will take a record set and return a demand-ready chronology. You can also do a smaller version yourself with Claude and a Project holding your firm’s format.
What AI does well here: the treatment timeline with dates and providers, gaps in treatment, billed totals, and references to pre-existing conditions buried on page 340 of an intake form.
What still needs a human: causation. Which records matter to your theory. Anything a jury will eventually see. An AI does not know that the March gap in treatment happened because your client’s mother died, and it will read that gap as a defense argument. You know better, and you write that part.
What I built at my own firm
Every lemon law case at my firm needs the same three documents: repair orders, a purchase contract, and current registration. Every case. No variation.
For years that meant a human opening each incoming file, checking it against a mental checklist, and sending an email when something was missing. Pure overhead. Nothing about that job required judgment, and it happened hundreds of times a month.
So I built a system with Claude Code that reads what clients upload, checks it against what the case actually requires, and only escalates when something is genuinely wrong. Blurry photo of a repair order? Flagged. Registration expired? Flagged. Wrong vehicle on the purchase contract? Flagged. Everything else moves forward without a person touching it. The paralegal now handles exceptions instead of routine.
The rule I would give any firm owner: if your document set is the same on every matter, stop shopping and build. Off-the-shelf tools are built for firms with unpredictable document sets. Yours is predictable, and that predictability is exactly what makes a custom system cheap to build and reliable to run.
A prompt you can run today
Drop this into Claude with the documents attached. It works on a lemon law file, a PI file, or an employment file with the document types swapped.
You are a senior paralegal reviewing an incoming case file.
Documents attached: [purchase contract, repair orders, registration,
correspondence, medical records, or whatever you have]
Produce four sections, in this order:
1. CHRONOLOGY. A table with columns: Date, Event, Source Document,
Page. One row per dated event. Sort oldest to newest. If a date
is ambiguous or illegible, write "UNCLEAR" rather than guessing.
2. KEY FACTS. Extract and list: parties, vehicle or incident details,
dollar amounts with their source, and any deadlines or statutory
dates visible in the documents.
3. MISSING ITEMS. Compare what is here against this required list:
[your firm's checklist]. List what is absent and what is present
but unusable (illegible, unsigned, expired, wrong party).
4. RED FLAGS. Anything that weakens the case or contradicts another
document. Quote the language and cite the page.
Rules: do not infer facts that are not in the documents. Every number
and date must cite the document and page it came from. If you cannot
find something, say so instead of estimating.That last paragraph is the part most people leave out, and it is the part that makes the output trustworthy.
The e-discovery tier
When document volumes hit the tens of thousands, you leave the assistant tier entirely. Kira and eBrevia handle contract extraction across large agreement sets. Everlaw and Relativity aiR run document review and privilege screening at production scale, with the audit trail you need when the other side challenges your process.
These platforms are not smarter than Claude. They are built for defensibility: custodian tracking, review workflows, sampling, and the ability to prove how a document got coded. That infrastructure is what you are buying, and it is worth real money when you need it.
Most firms do not need it. If you are reviewing a 300-page production, a Claude Project does the job for $20 a month.
| Approach | Volume it fits | Cost tier | Best for |
|---|---|---|---|
| Claude or ChatGPT Project | Up to a few hundred pages per matter | $20 to $30/month | Solos and small firms, case file summaries, chronologies |
| PI record review services (PareIT, EvenUp) | One record set per case, any size | Per file, roughly $100 to $500 | Medical chronologies and demand packages |
| E-discovery platforms (Everlaw, Relativity aiR, Kira) | Tens of thousands of documents | Thousands per matter | Litigation production, privilege screening, defensibility |
| Custom build (Claude Code) | Repeating document sets, unlimited volume | Subscription plus build time | Firms where every file needs the same documents |
Confidentiality: the mistake to avoid
Putting client documents into a consumer free tier is the mistake. That is the version of these tools with the weakest terms, and it is the one most lawyers reach for first.
Four rules that cover it:
- Use business or team plans with no-training terms. Anthropic and OpenAI both offer them, and both are cheap compared to a bar complaint.
- Redact identifiers the analysis does not need. A chronology does not require your client’s social security number or full address.
- Keep the most sensitive material local. If the file involves a sealed matter or a minor, run it through a local model or do it by hand.
- Check your client engagement letter and any protective order in the matter before uploading anything to a third-party service.
Verify before you rely
Every number and every date the AI extracts gets spot-checked against the source document before anything downstream depends on it. Not a full re-read. A sample: pull five entries from the chronology, open the underlying page, confirm they match.
If the sample is clean, the extraction is probably clean. If two of five are wrong, throw out the run and fix the prompt. This is the same quality check you would run on a new paralegal’s first assignment, and it takes about ten minutes.
The firms that get burned skip this step because the output looks polished. Polish is not accuracy.
Do this on one file this week
Take your next incoming case file. Run the prompt above against it. Then have whoever normally reviews that file do it the old way, and compare the two.
You are looking for two numbers: how long the manual review took, and how many items the AI missed or invented. If the AI is fast and clean, you have found a process to change. If it missed things, tighten the checklist in the prompt and run it again on the next file.
Once the output is reliable enough to trust, the next question is what to do with all that extracted structure. For most firms the answer is the demand package, which is where AI demand letter tools pick up. And if you want to interrogate a case file conversationally instead of getting a one-shot summary, NotebookLM is free and takes five minutes to set up.