AI Is Exposing Years of Data Debt

Enterprise AI is revealing a broader form of data debt that includes semantics, ownership, lineage, access, unstructured knowledge, and reusable data services. Using production AI projects can probe into the data estate and turning recurring gaps into a practical modernization backlog.

Aug 18, 2026 · Blaize Stewart · #ai-adoption #data-strategy #data-governance #enterprise-architecture #ai-readiness #data-debt

Enterprises have a problem that has been building for decades.

Most organizations have evolved from nascent technologies that were introduced years ago and then added to over time. New technologies were layered on top of old technologies, which were layered on top of even older technologies, often without organizations ever having the opportunity, or the budget, to go back and actually update what was already there. Eventually, organizations end up with layers upon layers of technology interfacing and interacting through different methods, protocols, applications, databases, and architectures. The result is an ecosystem that can become brittle, difficult to migrate, difficult to manage, or simply very expensive to maintain.

Pretty much every mature organization has some effort underway to mitigate this in some capacity. The problem is that it always becomes a competition between three things:

Keeping the lights on
Net-new innovation
Updating old and legacy systems

Within those three streams, organizations almost always prioritize operations first, net-new innovation second, and dealing with the old stuff last. And that creates what we generally call technical debt.

Technical debt is the accumulation of old patterns, technologies, architectures, and decisions that remain in an environment long after better alternatives exist. Some of those things may have been cutting edge when they were introduced. In other cases, they were simply bad architecture or bad decisions that were never optimal in the first place, but organizations went with them anyway. Either way, that is where most organizations with any level of maturity, even just a few years, eventually find themselves. One of the places where this becomes particularly problematic is data. The data may be difficult to access. Its quality may be low. It may contain missing values. It may be stale, duplicated, inconsistent, or simply incorrect. It may be structured in some places and completely unstructured in others. And all of that creates a disparate data ecosystem where it becomes very difficult to know what data is good, what data isn't, and sometimes even what the data actually means.

I personally remember working on systems with old databases where fields labeled as one thing were being used for completely different purposes. I've seen phone numbers stored in all sorts of inconsistent ways. Applications then had to program around those inconsistencies just to make the data usable. Organizations end up writing code because decades of accumulated inconsistency have to be compensated for somewhere, even when the business itself has no real need for that code. And cleaning all of that up becomes an enormous undertaking. Organizations have been able to live with this problem for years because applications could be programmed around it. Developers knew that System A stored something one way and System B stored it another way, so they wrote integration logic to compensate for the differences.

AI Exposes the Data Debt

Then AI arrived. And AI is beginning to expose just how much data debt organizations have accumulated.

AI is actually pretty decent at dealing with disparity in data. It can contextually recognize when something doesn't match an expected format and, assuming it understands the context, transform or normalize it. If an application requires a standard phone-number format while the underlying data stores phone numbers in ten different ways, AI can potentially recognize those differences and normalize the values into something usable. That capability is enormously useful. In doing so, AI is also exposing just how much technical debt exists underneath enterprise data environments. McKinsey's June 2026 research found that more than two-thirds of high-performing companies identify data as the primary obstacle to enabling AI. Their conclusion is that organizations need governed, traceable, and reusable data foundations spanning both structured and unstructured information. Accenture reached a similar conclusion in research involving executives at 2,000 companies across 15 countries and nine industries. Sixty-four percent of organizations had moved beyond pilots and into AI production across multiple functions or coordinated enterprise efforts. Yet only 7 percent qualified as what Accenture calls "data reinventors," organizations with advanced, AI-ready data capabilities. That is an enormous gap. Organizations are moving AI into production faster than they are making their data genuinely ready for AI.

The Semantic Problem

There is another problem here that is somewhat different from traditional data quality. A database can be technically clean and still be semantically ambiguous. The issue may have nothing to do with whether the data is formatted correctly. It may simply be difficult to understand what the data means. That matters because part of what makes AI valuable is its ability to retrieve data and understand it well enough to reason about it. If the underlying meaning is ambiguous, then AI has a much harder problem. Something called "revenue" in one system may mean something significantly different from "revenue" in another system. "Customer" may have one definition in sales, another in finance, and another in operations. A policy status, transaction date, account classification, or product category may have subtly different meanings depending on the application that created it.

Every individual database can be perfectly valid according to its own schema while the enterprise as a whole remains semantically inconsistent. Strategy's 2026 survey of senior data and technology leaders at large enterprises found that 99 percent struggle to define consistent business metrics across tools and departments. That is the problem. AI needs to compare apples to apples. Organizations, however, have spent decades creating apples, oranges, pears, and bananas and occasionally labeling all four of them "apple."

Unstructured Data Compounds the Problem

Then there is unstructured data, including contracts, policies, tickets, emails, transcripts, and so on. The knowledge contained in those systems can be even more disparate than the information stored in traditional databases because there isn't necessarily a schema at all. I've worked with systems where this is the case, and what I've found is that semantic context matters. Getting a handle on that context can become enormously valuable. First, however, the organization has to know what information exists, where it exists, and what it means. A 2025 Komprise survey of 200 IT directors and executives at U.S. enterprises found that 54 percent identified finding and moving the right unstructured data for AI ingestion as their leading preparation challenge, while 40 percent cited a lack of visibility across storage environments. That research was vendor-sponsored, so the numbers should be understood in that context, but the underlying problem is one most large organizations will immediately recognize. IBM's 2025 CDO research reinforces the same basic point from another direction. In a study involving more than 1,700 chief data officers across 19 industries and 27 geographies, 72 percent of CEOs viewed proprietary data as important to generative AI value. Yet accessibility, completeness, integrity, accuracy, and consistency remain major barriers. Proprietary enterprise data is potentially one of the most valuable things an organization can bring to AI, while simultaneously being one of the things organizations have the greatest difficulty actually using.

Access and Governance Become AI Problems

In other cases, access controls can become another blocker. Some older systems simply don't have the necessary nuance. Access is effectively all or nothing. They weren't designed for an ecosystem in which agents, models, APIs, users, and applications might need different levels of access to different pieces of information. They don't have fine-grained access controls across the data. That becomes particularly problematic with AI because an AI system may technically be capable of finding and understanding information that it should not necessarily be allowed to retrieve for a particular user or application. AI-ready data therefore has to be clean, understandable, discoverable, governed, and appropriately accessible. Deloitte's 2026 State of AI research reflects this broader problem. Organizations report feeling less prepared in infrastructure, data, risk, and talent than they do in strategy, and Deloitte argues that legacy data and infrastructure architectures are increasingly inadequate for more real-time and autonomous AI systems. In other words, organizations may have an AI strategy while the architecture underneath it simply isn't ready.

AI Can Help Address the Problem

Interestingly, AI itself can help.

The first is data discovery. Organizations can take samples from a myriad of data systems and allow AI to explore those samples. With input from users and subject-matter experts, an organization can begin creating an understanding of what that data actually represents. This isn't necessarily a formal schema. It is more of a semantic framework. That begins to create meaning around the data, and once that understanding exists, the organization has a basic framework for doing something about it. AI therefore becomes useful as part of discovering the extent of the problem and as part of addressing it.

From Normalized Data to a Semantic Layer

Second, once organizations have begun creating that semantic framework, they can start sorting and normalizing the data. One approach is essentially to leave the source systems alone and create something like a medallion architecture in a data lake or lakehouse. Organizations have the raw tier, the bronze layer, which is basically the data as it came from the source. Then there is silver, where the data has been cleaned, normalized, and structured. Traditionally, gold is where data is transformed into forms specifically designed for consumption. That architecture still makes sense. With AI, though, sometimes the goal isn't necessarily getting everything to a gold dataset. It may be more useful to get to a gold schema or semantic layer. The silver tier represents a normalized view of the underlying data. Rather than creating another transformed gold copy for every possible permutation or application, an organization can create a semantic layer that defines what the data actually means. Then applications, MCP servers, agents, and other AI systems can request information through that semantic understanding. The AI system understands the data well enough to generate the appropriate structured queries, retrieve the appropriate information, and interpret it consistently across otherwise disparate systems. Traditional gold datasets can still exist where they serve a useful purpose. They can even have their own semantic context. Producing another physical representation of the data for every possible use case simply isn't necessarily as important as it used to be. The semantic layer becomes the thing that creates consistency.

Mitigation Is Not Elimination

There is an important caveat to all of this. These are mitigating approaches. They create an ecosystem in which AI can operate without requiring the organization to completely reorganize, rewrite, migrate, or replace every legacy system first, and that is extremely useful. It doesn't necessarily eliminate the technical debt. In fact, if an organization is not careful, it can create future technical debt because it is adding yet another layer on top of the systems that already exist.

The bigger push to actually deal with the underlying technical debt may still be the ultimate solution. At some point, some systems really do need to be modernized, replaced, consolidated, or retired. Organizations don't operate in a world where everything can stop for three years while the entire enterprise architecture gets rebuilt correctly. Operations still have to run. New capabilities still have to ship. The business still has to innovate.

Turning Exposure Into a Roadmap

So sometimes the practical answer is mitigation. The organizations create the semantic framework, normalize what needs to be normalized, establish governance and appropriate access controls, and build an architecture that allows AI to use the information that already exists. Then what AI exposes can become a roadmap for deciding which underlying problems actually need to be fixed, because that may ultimately be one of AI's unexpected contributions to enterprise architecture.

AI isn't creating the data problem, but it's exposing the one that has been accumulating for years. And maybe it is exactly what is needed to help clean it up.