Title: The 42-Lead Mistake: What Happens When You Let AI Work Unsupervised

A research job that should have covered 42 leads

In October 2026, a research job meant to cover 42 leads processed one. A single setting was wrong. Nothing stopped it. Nothing flagged it. The job finished, reported success, and moved on.

A person caught it only because they checked the numbers afterward.

That is the part worth sitting with if you are a law firm owner thinking about where AI fits into your operations. The tool did not fail loudly. It failed quietly, and quiet failures are easy to miss until they have already cost you something.

The judge with the same last name

A second example from the same month makes the point better than any warning could. A name-matching rule was supposed to suggest contacts at law firms. It matched a surname and suggested a sitting judge as the contact to call.

Nobody dialed that number. A person reviewed the suggestion before any call was made and caught it immediately. But the rule itself did not know a judge from a paralegal. It only knew that two names matched.

Why this matters for your firm, not just for AI vendors

If you run a law firm, the responsibility for the work your firm puts out stays with your firm, whoever or whatever produced the first draft. Using a tool to draft, research, or triage does not change who answers for the output. Your own professional obligations are worth confirming with your bar association or ethics counsel before you put a tool to work on client matters.

That is a reason to build AI tools the way you would train a new hire: with a review step, a log, and a person whose job it is to check the numbers.

What a safe setup actually looks like

In the first case, a single setting was wrong. In the second, a shared surname produced the match. In both, the safeguard that worked was the person checking the output before it went anywhere near a client, a contact list, or a call.

For a law firm, that means a few concrete habits:

A human reviews output before it leaves the building, whether that is a drafted email, a research summary, or a list of contacts to call.

Automated changes get logged, with a timestamp, so you can trace what happened and when.

Nobody runs a workflow against real client data without first testing it on a small batch and checking the results by hand.

The confidentiality question nobody skips

If you are considering any AI tool that touches client information, ask where that data lives and who can see it. A workflow that pulls public information to research a prospect is a different animal than one that touches privileged client files. Treat them differently. Keep access controls tight, and know exactly what the tool can see before you let it run.

The actual time savings, if you supervise them

None of this means AI integrations are not worth the effort. A workflow that drafts a weekly summary, flags calls that went unanswered, or organizes intake information can take the repetitive first pass off your desk. In our view, the value shows up when the system does that first pass and a person does the judgment call at the end.

The firms that get this right are the ones who built a habit of checking before anything goes out the door.

What to ask before you turn one loose

Before you hand any part of your operations to an automated workflow, ask three questions. Who reviews the output before it reaches a client or a contact. What does the log show if something goes wrong. And what happens to the one bad name, the one skipped lead, the one setting that was wrong, before it costs you something.

If you cannot answer those three questions about a tool you are using today, that is where to start.