AI Is About to Expose 20 Years of Analytics Technical Debt.

What analytics technical debt looks like

For a long time, companies have been able to live with messy analytics. Not happily, necessarily, but they could live with it. Finance knew which spreadsheet contained the "real" forecast. Sales knew which Salesforce field was unreliable. Operations knew that one report was always a day behind. Marketing knew that its customer count did not match Finance's customer count. Everyone learned which numbers to trust, which numbers to ignore, and who to call when something looked wrong. That is what analytics technical debt looks like in the real world. It is rarely one catastrophic problem. It is usually hundreds of small compromises accumulated over years: a field that means two different things depending on the business unit, a pipeline that was supposed to be temporary but became permanent, a spreadsheet that started as a workaround and quietly became part of the company's operating process, or a dashboard built on top of another dashboard because nobody wanted to touch the logic underneath it. Companies have been carrying this kind of debt for years. AI is about to make it much harder to hide.

Tribal knowledge filled the gaps

The reason is simple. Traditional analytics depends heavily on human context, and AI does not automatically have that context. A CFO can look at a report and know that one division books revenue differently. A sales leader can see an opportunity value and know that the field is usually overstated. An analyst can look at a broken join and immediately recognize which customer table is causing the problem. That knowledge often does not exist anywhere in the system. It exists in people. For years, we have compensated for weak analytics architecture with tribal knowledge. Someone knows the workaround. Someone knows the right query. Someone remembers why the number is different. Someone knows that the report called "Revenue Final" is actually less reliable than the report called "Revenue Final v3." Humans are surprisingly good at operating inside systems like this. AI is much less forgiving.

AI does not have that context

When we ask an AI system a question like, "Why did revenue drop last month?" it does not know that Finance and Sales use different definitions of revenue unless we have explicitly given it that understanding. It does not know that the CRM contains duplicate customer records unless the architecture exposes that problem. It does not know that one data source stopped updating three days ago unless freshness is being monitored. It does not know that a particular metric changed definition in January unless that definition is versioned and governed. This is where I think a lot of companies are going to run into trouble. They will believe they have an AI problem. What they actually have is an analytics technical debt problem that AI has finally made impossible to ignore - the same pattern AWS has described: everyone adopted AI, yet tech debt is still growing.

Another dashboard is not the answer

For the last decade, the answer to analytics problems was often another dashboard. If a team did not trust the existing dashboard, someone built a new one. If the metric definitions were inconsistent, someone created a different report. If the data pipeline was too slow, somebody exported the data to Excel. If the ERP could not answer the question, someone wrote a custom SQL query. None of these solutions were necessarily irrational. Most were reasonable responses to immediate business pressure. The problem is what happens when they accumulate. Eventually the company has several versions of revenue, several versions of customer, several versions of pipeline, multiple data pipelines performing similar transformations, and dozens of reports that all claim to describe the same business. People learn to navigate that complexity. AI will expose it.

Confidence without authority

A language model can make this especially dangerous because it is very good at presenting uncertainty as confidence. If there are three plausible definitions of a metric and nobody has told the system which one is authoritative, the model may simply choose one. And it will probably explain the answer beautifully. That is worse than a dashboard being wrong. A bad dashboard looks like a bad dashboard. A conversational AI system can give you the wrong answer in exactly the language you wanted to hear. This is why I think companies need to rethink the idea that they can simply "put AI on top of the data." Which data? That question sounds trivial until you actually try to answer it. The CRM has customer data. The ERP has customer data. The support platform has customer data. Marketing has customer data. The data warehouse probably has customer data. There may also be several spreadsheets with customer data. Which one represents the customer? And if the answer is "it depends," then that dependency needs to become part of the architecture. I wrote more about that boundary in Enterprise AI Is New. The Rules for Running Production Systems Aren't.

Definitions the company must own

The same is true for business logic. What is recurring revenue? What counts as churn? When does an opportunity become pipeline? What is gross margin? Which orders count toward bookings? How do you treat refunds, credits, renewals, and partial shipments? Those are not questions the AI should be inventing answers to. They are business definitions, and the company needs to own them. This is where analytics technical debt starts to become more than an inconvenience. It becomes a constraint on how much autonomy you can safely give AI. If the data is poorly connected, the AI has incomplete context. If the metric definitions are inconsistent, the AI has ambiguous meaning. If data lineage is weak, the AI cannot explain where an answer came from. If freshness is unknown, the AI cannot tell whether it is reasoning from yesterday's business or last week's. If permissions are broad, the system may be able to act on information it does not fully understand. That is not an AI failure. That is an architecture failure - the next wave of technical debt is architectural, and AI is accelerating it.

Pay down the debt that matters

The good news is that I do not think companies need to solve every data problem before they can start using AI. That would be unrealistic. Most companies will never reach some mythical state where every dataset is perfectly modeled, every metric has universal agreement, and every legacy system has been retired. The goal is not perfection. The goal is to make the important parts of the business explicit. Start with the workflows where AI can create real value. Identify the systems those workflows depend on. Establish which sources are authoritative. Define the important business terms. Make freshness visible. Preserve lineage. Put boundaries around what the system is allowed to do. In other words, pay down the technical debt that matters to the decision you are trying to automate.

That approach is very different from trying to "AI-enable the enterprise." I am not even sure that phrase means anything useful. The better question is: what decision are we trying to improve, what data does that decision require, and what needs to be true about that data before we trust AI to participate? That is a much more manageable problem. It is also where I think the next generation of analytics platforms will look very different from the last one. Traditional business intelligence was primarily designed to help humans see the data. Enterprise AI needs to understand the data. That requires more than charts and dashboards. It needs context, relationships, definitions, history, an understanding of what is missing, and increasingly, rules around what happens after an insight is discovered.

An operating layer for analytics

That is one of the ideas behind Sherpa AI. We are not trying to solve the problem by putting a chat interface on top of a pile of disconnected enterprise systems. The goal is to create an operating layer for analytics where the business data is connected, the meaning is governed, and AI has enough context to reason about what is actually happening. Because AI is not going to eliminate analytics technical debt. It is going to expose it. And in many companies, I think that exposure is going to be uncomfortable. For the production-side rules that keep that exposure from becoming irreversible action, see Enterprise AI Is New. The Rules for Running Production Systems Aren't.

Final thoughts

The organizations that benefit most from AI will not necessarily be the ones with the most models, the most agents, or the biggest AI budgets. They will be the ones that finally make their data understandable enough for machines to work with. For twenty years, humans have been filling in the gaps. AI is going to force us to start fixing them.