The AI Intake Agent for a Law Firm That Can't Quote a Number (Simply Explained)

The One Question That Could Get a Law Firm Fined

A personal-injury law firm near Los Angeles came to me with a problem that sounds simple. They wanted an AI assistant to answer their phones. Calls were up, after-hours calls went to voicemail, and they were losing cases to firms that actually picked up.

Easy enough. I build these kinds of systems all the time.

Then they told me the catch. By law, this AI cannot tell a caller what their case is worth. It cannot promise an outcome. Not "shouldn't." Cannot.

Here's the nightmare. A caller, often in pain and scared, asks the most natural question in the world: "How much is my case worth?"

A normal AI, trained to be helpful, makes something up. It says, "Cases like yours usually settle around $40,000." It means well. It just broke the law.

In California, that made-up number counts as misleading advertising. The firm didn't say it. A lawyer didn't say it. But the firm's AI, on the firm's phone line, said it. That's real legal trouble, and "the robot said it" is not a defense any insurance company wants to hear.

So the job isn't "build a chatbot." The job is: make "never say a dollar amount" actually hold, even when the AI occasionally slips.

That slip is the fear underneath every AI project people bring me. Not "will it sound good." It will. The fear is that one day it says something off-script and creates a mess you can't take back.

That fear is correct. It's also solvable. Here's how I solved it.

Why "Just Tell It Not To" Doesn't Work

The obvious fix is to give the AI clear instructions. "Never quote a dollar amount. Never estimate case value." Done, right?

No. And understanding why is the whole point.

Think of AI like a brilliant new employee who follows your instructions almost all the time. Almost. Most of the time the rule wins. But "most of the time" is not good enough when one slip can get a firm fined.

Put it under pressure and the cracks show. A caller who asks three times. A sad story that pulls the AI toward wanting to comfort them. A reworded question: "I'm not asking you to promise, just ballpark, what do these usually go for?"

Under that pressure, the AI sometimes sides with the upset human in front of it over the rule you gave it. I've watched it happen in testing. Not often. Often enough.

So I never trust one instruction to hold the line. I build the safety into the system around it. I assume the AI will slip, and I build so it doesn't matter when it does.

Three Locks on the Same Door

Instead of one rule, I built three independent safeguards. Think of it like a bank with a guard at the door, a vault, and a camera. Each does a different job. For one bad thing to get through, all three have to fail at once.

Lock one: catch the question before the AI even sees it.

Before the AI reads anything, a simple filter scans the caller's message. If the caller is asking "what's my case worth" in any form, the AI never gets a turn to answer at all.

This firm's callers speak English, Spanish, Farsi, and Armenian. So the filter covers all four. Native speakers helped me build the word lists so it wasn't just sloppy machine translation.

This part isn't fancy AI. It's a plain checklist of words and patterns. That's exactly why it's reliable. It either matches or it doesn't. There's no "the AI was feeling generous today."

When it catches a money question, the caller hears a fixed line the firm's lawyers wrote and approved:

"I can't put a number on your case, but I can connect you with an attorney who can review the details and give you real answers."

The AI didn't write that line, so the AI can't mess it up.

I made one tradeoff on purpose. It's fine if the filter is a little too cautious and blocks a harmless question that happened to include the word "worth." Slightly awkward. Survivable. What's not fine is letting one real money question slip through. So I tuned it to over-block, every time. I'd rather deflect ten harmless questions than let one made-up number through.

Lock two: a hard rule inside the AI itself.

Some questions slip past the first filter. Reworded ones. "I just want to understand what to expect financially." No trigger word, but the meaning is obvious.

So the AI also carries one absolute rule: never state, estimate, hint at, or imply a dollar amount or outcome. Not even a range. And, just as important, I give it somewhere safe to go instead: acknowledge the question, show you care, and hand off to an attorney. An AI with no approved alternative will improvise to fill the silence. An AI with a clear exit takes it.

I'll be honest. This is the weakest of the three locks, because it still relies on the AI behaving. That's exactly why it sits in the middle, with a reliable filter in front and a reliable filter behind it.

Lock three: catch it on the way out.

The last safeguard checks the AI's answer after it's written but before the caller hears a single word. It scans for dollar signs, numbers that look like money, words like "thousand" or "million." If it finds one, it strips it out or falls back to the safe approved line. The caller never hears the number.

And every single catch gets logged. Time, what the AI tried to say, what got removed.

That log is the whole reason this system is safe to turn on. When the firm's insurance company asks "how do you know it never quoted a number," the answer isn't a shrug. It's a record. Every near-miss is documented, including the ones the caller never heard because the filter caught them. Those logs also get reviewed by a real person, who tightens the filters over time.

Why Three and Not One

Here's the math that makes this work. No single lock is good enough alone. But they fail in different ways and for different reasons. So they almost never fail at the same time.

If each one slipped once in fifty tries on its own, all three slipping on the same call would be about one in 125,000. And the times two of them fail, the third still catches it.

Here's the reframe for any business owner worried about this. The question isn't "will the AI ever get something wrong." Any honest builder will tell you it can. The real question is: what happens when it does, and can you prove your safeguards held?

That's the difference between a demo and a system you can actually run a business on. A demo behaves on stage. A real system has documented backups and a paper trail.

This same approach works anywhere an off-script sentence is a liability. Healthcare lines that can't give medical advice. Financial bots that can't recommend an investment. The industry changes. The discipline doesn't.

Most AI consultants can talk about this. Far fewer can build it and hand you the audit trail your lawyers will actually accept. I build it, I try to break it myself, and I hand you the proof.

If you've been sitting on a customer-facing AI project you're afraid to turn on because of exactly this risk, hear me clearly. That fear is the right instinct. The mistake is letting it stop you when the controls exist.

Want to explore what AI could do for your business?

Book a free 30-minute strategy call. No pitch deck, no sales team, just a real conversation about your operations and where AI fits.

Book a Discovery Call