Stop Using AI as a Search Engine

Almost every lawyer I talk to is using AI as a fancier Google. They ask it a question. They get a paragraph. They decide the paragraph is either impressive or wrong, and that’s the end of the experiment.

That’s the least valuable thing these tools do.

I run a personal injury practice with a live docket — real deadlines, real discovery, real depositions on the calendar. That’s the environment I’ve built my system in, and it’s a decent stress test: if something works when there are forty-odd active cases and a hearing on Thursday, it’ll work anywhere. Here’s the shift that actually changed my practice.

Search asks a question. Work gives a job.

A search-engine prompt sounds like this:

What are the elements of a negligent security claim in Georgia?

A work prompt sounds like this:

Here are 340 pages of incident reports produced by the defendant. Build me a table of every prior incident at this property in the last five years — date, type, location on the premises, whether police were called, and the Bates number. Flag anything within 200 feet of where my client was attacked.

The first one gives you a paragraph you could have found yourself. The second one gives you two days of your life back, and it produces something you can actually use in a deposition.

The difference isn’t the model. It’s whether you gave it material and a deliverable.

The three ingredients

Every prompt that’s ever been worth writing has the same three parts.

Context. The actual documents. The actual record. Not a summary of them — the things themselves. A model reasoning about your case from your two-sentence description is guessing. A model reading the deposition transcript is working.

A defined output. “Summarize this” is not a deliverable. “A chronological table with columns X, Y, Z, one row per entry, citing the page” is a deliverable. If you can’t describe the shape of what you want, you’re going to get something shaped like nothing.

A standard. Tell it what good looks like and what failure looks like. Every fact must cite a page number. If a date is ambiguous, flag it rather than guessing. Models will hedge and smooth things over unless you tell them not to.

Miss any one of these and you get the paragraph.

Where the real leverage is

The tasks worth automating are almost never the interesting ones. They’re the ones that are:

• High volume — hundreds of pages, dozens of entries
• Structured — the output has a predictable shape
• Verifiable — you can spot-check it against the source in seconds

Medical chronologies. Document indexes. Deposition outlines built from the record. Discovery response drafts. First passes at a timeline. Gap analysis — what should be in this production that isn’t?

That last one is where it earns its keep. Last quarter a records review flagged a two-month gap in a treatment history that would have taken a chunk out of a demand if it had shown up first in a mediation. It didn’t find that because it’s clever. It found it because it read all four hundred pages and I would have skimmed.

The part nobody wants to hear

You have to check it.

Not “trust but verify” as a slogan — actually check it, against the source, every time it matters. That’s why the citation requirement isn’t optional. A chronology where every line points to a page is a chronology you can audit in ten minutes. One without page numbers is a liability with nice formatting.

This is also the professional obligation. Georgia Rule 1.1 didn’t get an AI exception. The work product is yours no matter what generated the first draft, and courts have been unambiguous about what happens to lawyers who file things they didn’t read.

Verification isn’t the tax you pay for using these tools. It’s the thing that makes the output worth having.

Where to start this week

Pick one task you did last month that was tedious, high-volume, and structured. Not a task you found difficult — a task you found boring.

Give the model the actual documents. Describe the exact table you want. Tell it to cite everything.

Then check it against the source, line by line, once. You’ll learn more about what these tools are for in that one hour than in a year of asking them questions.

---

I write about the systems behind a working litigation practice — the tools, the workflows, and the parts that don’t work. If that’s useful to you, the rest of it is here.