I Built an AI Compliance Review Tool That Audits Ads in 60s (Simply Explained)

A radio show stuck in review for days

I worked with a financial advisory firm that runs a weekly radio show. Before any segment could air or get turned into a clip, it had to pass a legal review. That's required for a firm in their industry. No way around it.

The problem wasn't the review. It was how long it took. Segments sat for days. By the time they got cleared, the market talk inside them was old news. The hosts were frustrated their best, timely takes got delayed or watered down. The review team was buried in a queue they could never clear.

When I looked closer, I found the real issue. The reviewers treated everything the same way. A casual, live, off-the-cuff radio chat got judged by the exact same harsh checklist as a polished, scripted commercial. Under the rules, those are two completely different things. But they were being held to the same standard.

That's not being careful. That's a reflex. When a reviewer isn't sure, they reach for the safe move: apply the strictest rule and nobody can accuse you of being too soft. Understandable. Also wrong, because it slows everything down and forces changes the actual rules never required.

Why "treat everything the strictest way" backfires

The rules in this industry sort communications into different buckets. A written ad sent out to the public carries the heaviest set of requirements. A host talking live and unscripted on air falls under a lighter, narrower set. They are not the same.

Then there's the in-between case: a segment that starts with a scripted promo, then drifts into live conversation. That one needs a mix of the rules, not the maximum of both piled on top.

Most overloaded review teams collapse all of this into one. They grab the strictest checklist and run it against everything, because that feels safest when they're unsure.

Here's the thing: being too cautious feels free. It isn't. When you force a live chat to carry full commercial-ad disclosures, you make the host read legal language that doesn't fit the format. You delay timely content until it's worthless. And listeners can tell when something's been sanded down by lawyers.

Over-caution is a process failure, same as being too loose. One gets you in trouble with regulators. The other quietly bleeds speed, money, and audience trust.

What I built instead

I built a tool that does the one thing the humans were skipping under pressure: figure out what kind of content it is first, then apply only the rules that actually govern it.

It works like an assembly line.

First, it takes in the content. If there's a script, it reads the text. If all that exists is the recording, the tool listens to the audio and writes out what was said. That matters, because most live segments were never scripted in the first place.

Next comes a quick scan for words that are flat-out banned no matter the context. Things like "guaranteed returns" or "risk-free." This is simple, exact pattern-matching. If those words show up, they get flagged with certainty.

Then the smart part kicks in. An AI that reads and understands language steps in, and its first job is not to hunt for problems. Its first job is to decide what kind of content this is. Scripted ad? Live appearance? A mix? Once it decides, it only checks the rules for that category. The rules that don't apply never get touched.

Finally, it produces a report you can act on:

• A score from 0 to 10, so a reviewer knows at a glance if it's clean or a problem.
• A list of issues, each with the exact rule it breaks and the exact words from the transcript that triggered it. No vague "this might be a problem."
• A list of any required disclosures that are missing.
• A cleaned-up rewrite, in the host's own voice, that fixes the issues without flattening the personality that makes the show worth listening to.

The whole thing runs in about 60 seconds. The old process took days.

Classifying first is the whole trick

If I had to point to the one decision that makes this work, it's this: the tool doesn't start by looking for violations. It starts by deciding what it's looking at.

That's how an experienced reviewer actually thinks. Hand a senior compliance officer a transcript and they don't immediately run a checklist. They first ask, "What kind of communication is this?" Because the answer decides which rulebook even applies.

A junior reviewer skips that question and applies the harshest checklist to everything. That was exactly the problem in the radio show queue.

Sorting the content first prevents mistakes in both directions. It stops the tool from flagging a casual chat for disclosures that only formal ads need. And it stops the tool from going easy on a scripted ad because it wrongly assumed it was casual talk.

Can AI do this without creating legal risk?

This is the first question every compliance lead asks me, and they should. The answer is yes, but only because of how the system is fenced in.

The tool never approves anything on its own. It never publishes. It produces a recommendation, and a human signs off. That's the core rule in every system I build: the AI proposes, a person decides.

What changes is the reviewer's job. Instead of reading a raw transcript cold and reaching for the safe default, they get a sorted, scored, fully cited document. They're faster because the busywork is done, and they make better calls because they can see exactly which rule applies and the precise evidence behind every flag.

Every run also gets written into a permanent record that can't be edited after the fact. The input, the decision, the rules considered, the evidence, the final verdict. All of it saved.

That's what actually lowers legal risk. Not the speed. The paper trail. When a regulator asks why something was approved, the firm can produce a complete record of how that call was made.

I'll be honest about the limit, because it's the whole reason the design holds together. The AI can be wrong. It can misjudge a tricky edge case. That's exactly why a human stays in charge and every decision stays documented.

What it doesn't do

The tool doesn't replace human judgment on brand-new or borderline cases. When something genuinely novel shows up, that's a human call. The tool flags the uncertainty, it doesn't pretend to settle it.

It doesn't know about a rule change nobody told it about. The rulebook is only as current as what you load in.

And the transcription isn't perfect. If hosts talk over each other or the recording is poor, accuracy drops. On those runs, the tool says so, and a person checks the transcript before trusting anything built on it.

Here's why I call all of that a feature, not an apology. The tool handles the mechanical 90% reliably and fast. That leaves the 10% that needs real judgment to the person who's actually accountable for it. AI does the volume. The human owns the calls that matter.

Strip away the radio details and this same approach works anywhere regulated marketing gets reviewed: healthcare claims, supplement copy, insurance, legal ads. Sort the content first, apply only the rules that fit, flag the banned stuff with certainty, cite the exact evidence, log everything, and let a human make the final call.

If your team is drowning in review, or quietly applying the harshest rule to everything because it's the only safe move you've got, that's exactly the kind of tool I build. Not a demo. A working system that fits your actual rules and your actual workflow.

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