The Data Quality Crisis Killing 85% of AI Projects (And How to Fix It)

Hi, folks. Welcome to today's AI briefing. I'm a slightly ill Tom Barber, and I'm coming to you today from Warsaw in Poland. Um, and today we're gonna talk about the data quality conversation your AI team doesn't really want you to have. So welcome to the AI briefing, and here is where we dig into small topics and soundbites that help cut through the noise as to what is important and what is not in-in the year of twenty twenty-six.

So let's get into it. Uh, KPMG's twenty twenty-five AI quarterly pulse survey found that eighty-five percent of leaders cite data quality as their most significant challenge in AI strategies, which may not come as a huge surprise to you all. Not compute power, not talent, not budget, but data quality. Yet most AI initiatives launch without addressing the foundational data problems that will eventually undermine them. This is a conversation that gets skipped because it's unsexy, it's expensive, and nobody wants to be the person who slows down the AI roadmap.

Of course, looking at the scale of the problem, the Kinsha's State of the Cybersecurity Resilience twenty twenty-five report said twe-- seventy-seven percent of organizations lack the essential data and AI security practices needed to safeguard critical business models, data pipelines, and cloud infrastructure. IBM's twenty twenty-five CEO study found that seventy-two percent of CEOs view their organization's proprietary data as keys to unlocking GenAI value. Probably not a huge surprise. But fifty percent of those acknowledge that the pace of recent investments has left the organization with a significant data challenge. Drexel University and Precisely study found that leaders lack confidence that their enterprise data is ready to make the leap from human to AI-- human-first to AI-only decisioning.

And that translates into executives who know that the data is fuel, but they're not confident that the fuel is clean. But this problem has been around since the dawn of time. It's just we choose to ignore it for reasons I do not understand. So why, of course, AI makes data problems worse and not better. Traditional software fails predictably.

Bad input mer-- input, error message, process stops. AI quite often fails quietly. Bad input, confident sounding, wrong answer, decisions made on flawed foundations. PwC research notes that AI raises questions about data lineage, sourcing, and usage rights that traditional systems have never had to answer. The complexity compounds, of course, with AI systems often using and processing more diverse data sets than traditional systems, making governance exponentially harder.

Poor data quality doesn't just produce wrong answers. It can embed bias, create compliance violations, and erode trust in AI outputs across the organization. So let's-- let me ask a few different questions. Question number one: Do we have a single source of truth for data in our-- for our AI models to consume? Every organization should ask this question, and most organizations have scattered data across systems that don't talk to each other.

AI models trained on conflicting versions of the same data will produce conflicting outputs. The question isn't whether you have a data warehouse, it's whether your AI initiatives are actually using it or pulling from whatever source was convenient at the time. Lance research, uh, identifies siloed repositories as key governance failure, different departments creating conflicting AI policies and data access controls without coordination. Question number two: Who owns data quality in our organization, and do they have the authority to block a deployment? Because data governance often falls into a gap between IT compliance, data science, and business leadership.

Atland again identifies unclear ownership as a core challenge, data stewardship responsibilities scattered across teams without accountability. So if your data governance... Uh, sorry, if your data quality owner can raise concerns but not stop a launch, you don't really have data governance, you just have someone. So the test is when your last AI initiative went live, did anyone sign off specifically on data quality, or was it just assumed? Question number three is when our AI system gives a wrong answer, can we trace, uh, it back to the data that caused it?

This is data lineage and the ability to track data where it came from, and how it was transformed, and how it ended up being in your AI model. Without lineage, debugging AI failures becomes guesswork. Regulatory requirements are increasingly demanding of this traceability. The EU AI Act began phasing in obligations from February twenty twenty-five. If you can't explain why your AI made a decision, you can't defend it to regulators, customers, or your own board.

So the real cost of skipping this work, Forrester predicts it to be at least thirty percent of GenAI projects will be abandoned after proof of concept by the end of twenty twenty-five. Obviously, we're now in twenty twenty-six, so we'll find out. Poor data quality is listed as primary cause alongside escalating effort-- uh, escalating costs of unclear business value. Forrester has warned that impatience with AI ROI could lead to premature cutbacks, potentially hindering long-term benefits. And so the pattern is that organizations skip data governance to move faster, hit problems at scale, abandon the initiative, and then start over with a new pilot that also skips data governance to move faster.

Endless cycle. Tech debt from rushed AI implementations comp-compounds. Forrester highlights increasingly intricate, uh, IT landscapes as a major barrier. So you have to be able to reframe data governance as an accelerant. The conventional view is data governance slows things down, adds bureaucracy, and creates friction.

The reality, of course, is that organizations with sl-- strong data foundations scale AI faster because they don't hit the same walls. KPMG research positions modern data governance as integrating AI and data governance under a single umbrella, enable complete transparency, enforceable policies, and eliminating duplicate data sets. PwC's guidance is to elevate data governance to a board-level priority, making it central to your AI strategy and not an afterthought. And the organizations treating data governance as strategic investment rather than a compliance checkbox are the ones whose pilots actually become product. So what does good look like?

Atland's framework identifies key components. Data security Data lineage, data quality, and compliance. PwC recommends that people, organizations rationalize data sources to reduce silos, establish a single source of truth for high-value data assets, and standardized data definitions and access protocols across business units. The goal isn't perfect data. It's knowing when your data is imperfect and governing AI use accordingly.

Some AI use control tolerate lower data quality, others cannot. The governance framework should distinguish between them. So the conversation, going back to the start of this podcast, that your AI team doesn't want you to have. AI teams are under pressure to show progress, demonstrate innovation, and hit deployment milestones. Raising data quality concerns feels like the person who slows everything down.

But the MIT research on pilot failures found data quality consistently among the root causes of initiatives that stalled. The conversation that feels uncomfortable now is far less uncomfortable than explaining to the board why a ho- high pro-profile AI initiative failed eighteen months in. Executives can make space for this conversation by asking three questions directly and making clear that honest answers are valued to more optimistic timelines. So what can you do about it this week? Ask your AI lead or CTO directly what is the data quality score for the training data in our highest priority AI initiative?

And if there isn't a clear answer, that's your answer. Identify who in your organization owns data quality decisions and whether that person has the authority and resources to enforce standards. Review your most advanced AI pilot and ask, "If this produces a wrong output next month, can we trace it back to the specific data that caused it? " And if you're planning AI investments for the next budget cycle, ensure data governance is a line item, not an assumption buried in AI project budget. So closing thought, the organizations successfully scaling AI aren't the ones with the most sophisticated models.

They're the ones who did the unglamorous work on the data first. Data quality isn't a technical problem to delegate. It's a leadership question about whether your organization is building AI on a solid foundation or on sand. The eighty-five percent of leaders citing data quality as their top challenge know this matters. The question is whether knowing translates into action before the next pilot stalls.

I have been Tom Barber. Hopefully, this has been useful for you. This is the AI Brief-Briefing. Make sure you subscribe to this podcast for more hopefully useful insight. If you would like to know something s-specific, please do get in touch.

I am happy to address whatever is useful for people. For now, thank you for tuning in. I'll be back tomorrow with some more AI Briefing content. Bye for now.