What Happens When You Put AI on Top of Bad Data

Most organizations are not behind on AI. They are ahead of their own data.

The demo problem

Mihai Bejgu has been in marketing operations for close to two decades. He's sat through a lot of vendor demos. He notices the same thing every time: the data in the demo is perfect. Every field populated, every lead source clean, every routing rule firing correctly.

The technology works beautifully. Because someone spent hours making the data beautiful first.

Then you go back to your own CRM.

"If you're just putting an AI layer over bad data, it will just automate chaos."

That is the gap between the demo and the deployment. Not the software. The foundation underneath it.

The lead source problem

Mihai's go-to example is simple and ruthless: lead source.

Most organizations have it as a free text field. Someone types "LinkedIn." Someone else types "linkedin." A third person writes "LinkedIn Campaign." A fourth writes "LI." You now have four versions of the same source, and none of them aggregate cleanly.

His fix sounds unglamorous: pick list values. Twenty sources, maximum. Everyone aligned on exactly what each one means. Nothing outside the list gets in without being investigated.

Before you automate anything, says Mihai, you need people working from the same map.

The orchestra metaphor

Here is where the conversation gets interesting. Because Mihai is not against AI. He's against the lazy version of it.

Most teams are using AI as a single instrument. One prompt, one task, one agent trying to do everything at once. The results are inconsistent and hard to trust.

What Mihai describes instead is an orchestra. Multiple agents, each trained for a specific task. One monitors the CRM for data gaps. A second enriches and corrects bad records. A third updates scoring and routing. A fourth analyzes engagement patterns and recommends next best actions.

Each agent does one thing well. A QA agent oversees the others. And a human sits above all of it.

"Instead of trying to solve all of the things with one prompt... it's best to divide and train your AI agents to do small tasks and make them good at it."

This is not how most teams are working today. But Mihai says the companies that have gotten there are already seeing results most teams are still pitching in slide decks.

The .con problem

The moment in this conversation that stuck with me most is the one that sounds smallest.

You import an event list. A thousand people. Someone's email address has .con instead of .com . One character. It gets imported. It bounces. It sits in your database as a dead record. Nobody catches it until months later, if ever.

A human doing a spot check won't find it. A properly trained agent will scan every row, every cell, every time.

"It will be much easier for an agent to do that instead of having a person that cannot humanly do it."

This is what Mihai means when he talks about what AI is actually good for. Not replacing strategy. Replacing the relentless, boring, row-by-row work that humans skip because they have to.

The human layer

After all the architecture talk, Mihai comes back to something that surprised me.

No matter how many agents you build. No matter how well-orchestrated the system. You never remove human oversight.

Not because AI can't be trusted. Because the stakes are too high when it gets something wrong and nobody was watching.

The goal is not a fully automated system. It is a system where humans spend their time on things that actually require human judgment, and agents handle everything that doesn't.

"No matter how many agents you have, you still need to have some human supervision over your whole architecture."

The bottom line

Mihai's parting advice is the same thing he says at every stage of this conversation: stop before you start.

Before the AI project kicks off. Before the vendor demo. Before the procurement process. Stop and find your data gaps. Understand what you actually want to achieve. Build the foundation first, then decide what to automate.

It is not a slow strategy. It is the only one that works.

If you work in marketing operations and this landed for you, Mihai is worth following. Find him on LinkedIn or through Revenue Pulse. He has been in the room where these problems live for a long time, and he is not selling easy answers.