AI Moves Fast on Customer Data, Including Bad Data

In a recent MarTech by Semrush interview, Subu Desaraju, head of commercial and operations at iceDQ, makes the case that AI does not fix bad marketing data. It inherits it, acts on it quickly and confidently, and scales the consequences before anyone notices.

"While everyone's rushing into the AI game, I really fear for the output of that process without the right foundations in terms of reliable data," he said.

Regulated industries are ahead on data discipline because the penalties for getting it wrong are clear and steep. Financial services pay close attention because regulators like FINRA step in when trade and position data do not match. Consumer and retail marketing has no equivalent forcing function. When a campaign underperforms, the creative gets blamed. The data rarely does. Gartner estimates poor data quality costs organisations an average of $15 million per year.

Most organisations check data quality at the consumption layer, the point where problems show up on dashboards after moving through every system in the pipeline accumulating faults. Desaraju maps the data pipeline as four stages: the lake where data enters, the distribution system where it is filtered and cleaned, the pipelines where it moves between platforms, and the tap where most teams currently do their quality checks. "If we want to consume clean, consistent, accurate data, we have to look at the value chain of the data and ensure that there are checks and controls in place at each point."

He offers two frameworks:

Trace a campaign backward. Starting from what the customer received, work back through five questions covering the execution report, the audience file, the segment definition, every data transformation, and the original source systems. For each step, fill in three columns: what check exists today, who owns that check, and what happens when it fails. Most teams find the ownership column is mostly empty and the failure column is completely empty. Those blanks are the plan.

Build across people, process, and tools. On process, three stages must be addressed in order. Test before you deploy. Monitor while it runs, with continuous checks on volume, timing, and completeness so problems surface in the pipeline not on a dashboard. Observe the output. If the first two stages are working, roughly 99% of defects never reach the consumption layer. On tools: stop asking which tool you need. Ask what business outcome you are responsible for, where the data value chain breaks down, and whether a specific tool closes those gaps. Desaraju advises: "The more fragmented the ecosystem becomes, the higher the complexity and lower the intelligence."