AI Evidence Analysis vs. Manual Review: What It Looks Like on a Sample Custody File

How long does it take to go through a custody file and pull out every person, date, place and financial reference? Done by hand (reading every document, highlighting names, building a spreadsheet), it is usually measured in hours.

To show what AI entity extraction does instead, we ran it on Martinez v. Martinez, the sample case that ships with TextEvidence. The parties, documents and messages are made up, but the file is built like a typical custody modification matter. You can open the same sample case and check every result below yourself.

The sample file

• A petition to modify custody
• A declaration from the petitioner
• An incident report from an urgent care visit
• 123 text messages across five conversations, including the other parent, a grandparent, a babysitter and the petitioner's attorney

What a manual pass involves

A careful manual review of a file like this usually means:

1. Reading every document and conversation start to finish
2. Noting every person mentioned: parties, children, relatives, witnesses, doctors, counsel
3. Pulling out dates: incidents, filings, hearings, deadlines
4. Flagging money: support amounts, medical bills
5. Noting places: residences, exchange locations, providers, courts
6. Cross-checking names and dates between documents
7. Writing it up in a form someone else can use

We did not time a manual review of this file, so we won't put a number on it. Most attorneys and paralegals can estimate from experience what a pass like that takes on their own cases.

What the extraction found

We ran Names & Places → Scan all messages on the sample case on October 22, 2026. It read all 123 messages and the 4 documents and finished in 98 seconds, with 61 items:

Each item comes with its role or context (for example "custody exchange location"), how important it looks, how often it appears, and the source: which document or message it came from, with the quote. That last part matters: nothing in the list has to be taken on faith, because every entry points back to where it was said.

Where it helps, and where it doesn't

It helps with the inventory. Extraction is good at the tedious part: making sure a name mentioned once in a text from the babysitter, or an exchange location buried in a declaration, ends up on the list.

It doesn't replace judgment. The list still needs a review. Two mentions of the same child under slightly different names can show up as separate entries, and deciding which facts matter to the case is still the attorney's job. Treat the output as a well-organized first draft with citations, not a finished work product.

What the list feeds into

Once the names, places and dates exist, other parts of the case use them:

• The case timeline is built from the extracted dates
• The case calendar can import the dates that look like court dates and deadlines
• The AI assistant uses the list when it answers questions about the people and places in the case

A rough way to think about the value

Take your own numbers. If a manual inventory of a typical file takes you a few hours, and checking an extracted list with sources takes a fraction of that, the difference is the time saved per case. Multiply by the cases you open in a year. We would rather you run that math with your own figures than ours.

---

Want to see it on a real-shaped file? Open the sample case, or start a free trial and run it on your own.