Data Quality & Governance

The Hopmann AI Foundation, Part 3

In the first post of this series, we introduced the five layers of the Hopmann AI Foundation: Measurement Strategy, Data Quality & Governance, Semantic Layer, Context Layer and Agentic Marketing Intelligence. The previous post focused on Measurement Strategy, asking what should be measured in the first place. This time we move up one layer and ask a different question: can you trust the data you are measuring with?

It is the layer that gets talked about least. Data quality has no demo, and nobody gets applause for catching a sync error before it made its way into a board deck. It is still the layer where most AI projects in marketing actually fail.

Why Trust in AI Starts with the Data Set

A conventional report behaves fairly kindly when data is missing. A row stays empty, a total looks too low, somebody on the team trips over it and asks.

AI systems behave differently. A language model generally treats the data it receives as trustworthy. Unless the surrounding system validates the input, it will explain what it sees rather than question whether the underlying data is incomplete or wrong. In the best case it flags that the data may be incomplete, or that it cannot be confident in its conclusion. In the worst case it fills the gap with a plausible explanation and presents it as fact.

If three days of campaign data are missing, the model may not see a tracking error, it may see a drop in performance and offer a convincing reason for it. If a lead sits in the CRM twice, it may conclude that the audience has grown rather than recognising a duplicate. Both answers can sound just as confident as a correct one.

With AI, errors do not necessarily become more visible. They often become quieter.

This matters even more with agentic AI. An agent does not answer one question and stop. It retrieves data, performs calculations, generates recommendations and can trigger downstream actions. If the underlying data is wrong, every step that follows builds on the same faulty assumptions.

The numbers are clearer than you would expect for such an unglamorous topic. In the Precisely Data Integrity Report 2025, 67 percent of companies say they do not trust their own data enough to base AI-driven decisions on it. In Deloitte’s AI Survey 2024, 38 percent of executives report having already made a wrong decision based on incorrect AI output. Gartner names poor data quality as the most common reason AI applications fail.

These numbers are about trust, not about technology. And the biggest damage from unreliable data is not the single wrong figure. It is the point where the team stops believing the dashboard and goes back to deciding on instinct.

What “good data” actually means: the six dimensions

“We need better data quality” is a sentence as true as it is useless. As long as it is unclear which type of quality issue is meant, there is no way to check whether it has been fixed. That is why we work with six dimensions. Each one answers a different question, and each can break independently of the others.

Completeness: are records or values missing?

A connector fails to sync over the weekend, an account loses its authorisation. Neither produces an error message, both produce lower numbers. The campaign keeps running, but without feedback on its effect, and the budget optimises against nothing.

Accuracy: does the data describe reality?

Here nothing is missing, what is there is wrong. A conversion tag that fires twice after a relaunch delivers complete, current, formally flawless data. ROAS then looks twice as good as it is, and the recommendation is to move more budget into the weaker channel.

Uniqueness: are duplicates distorting the analysis?

Duplicate records do more than just inflate numbers, they create audiences that do not exist. If the same customer sits in three systems under three spellings, a model ends up optimising against segments of phantom people. Depending on where the duplicates occur, they can also double-count conversions or skew performance metrics in ways that are hard to trace back to the source.

Timeliness: is the data current enough for the decision?

This dimension is relative, not absolute. For monthly reporting, yesterday’s data is perfectly adequate. For a budget decision taken at midday, it is not. What matters is whether the freshness matches the cadence of the decisions built on it.

Consistency: do the values match across systems?

Marketing counts leads in the ad platform, Sales in the CRM, Finance works with a third number, and each is correct in its own system. For an AI system this creates ambiguity. Unless it has been given clear business rules, it has no reliable way to know which definition should apply, and whoever reads the answer has no way of knowing which one it used.

Validity: does the data follow the agreed rules?

Formats, value ranges, naming conventions, mandatory fields. It sounds like administration, but in practice it determines whether the data can be analysed at all. If campaign names do not follow the convention, you cannot reliably group by market, funnel stage or audience. The underlying data may be correct, but the structure needed for analysis is missing.

The six dimensions are not a checklist to complete once. They provide a shared language for describing where a quality problem sits. That is the difference between saying “our data is bad” and “we have a uniqueness problem in the customer domain that inflates our audience sizes by roughly fifteen percent”. Only the second statement can be prioritised, measured and fixed.

Why data quality is not a project with an end date

The most common misconception on this topic is that data quality is a state you establish once. An audit, a clean-up phase, a tick in a box.

In reality it is a continuous process, because data constantly changes. Platforms update their APIs, tracking implementations evolve, new channels introduce their own naming conventions, and business processes change over time. People leave teams, taking with them the knowledge of why certain fields exist or how particular transformations work. None of these changes looks dramatic on its own, but together they slowly erode trust in the data. An audit tells you how things look today. Only automated checks and thresholds tell you whether that is still true tomorrow.

The payoff

The payoff of this layer is trust, although that makes it sound softer than it really is. When the data is reliable, discussions about numbers become shorter, because everyone works from the same foundation. Analysts spend less time explaining discrepancies between systems and more time on the questions they were hired for. Problems are identified before they influence a budget decision, not after.

These benefits exist whether or not you ever build an AI agent. Clean data improves reporting, attribution, forecasting and decision-making across the organisation. AI simply raises the stakes: once a language model starts explaining trends, recommending actions or making decisions on behalf of users, poor data quality becomes flawed reasoning rather than just incorrect numbers. That is why this layer is not simply groundwork to get through before reaching the interesting part. It delivers value from the beginning while also providing the foundation for every layer that follows.

What comes next

With this layer in place, the data itself is dependable. One question still remains: what do the numbers actually mean? Clean data does not automatically mean everyone in the organisation interprets it in the same way. A “conversion” might refer to four slightly different calculations across four different tools, and every one of them can be technically correct. That is the focus of the next post in this series: the Semantic Layer. It defines metrics and calculation logic in one place so that people, dashboards and AI systems all work from the same definitions.

Anyone who wants agentic AI in marketing to succeed does not start with the agent. They start with the data the agent depends on.

Want to find out where your organisation stands on data quality? Schedule a free 30-minute initial consultation with our team.

Tobias Lanzl is Manager Data Analytics at Hopmann Marketing Analytics, specialising in AI-powered analytics architectures at the intersection of semantic models, context layers and large language models. As a former research associate and university lecturer he brings the ability to make complex AI and data architectures accessible in a way that works in practice.

Follow Hopmann Marketing Analytics on LinkedIn for more insights on marketing analytics and AI.