Companies are paying very close attention to technology spend right now. AI tools that were approved fairly casually a year ago are suddenly being asked to prove their value. Teams are putting limits on token usage, consolidating subscriptions, questioning cloud costs, and trying to figure out which experiments are actually worth continuing. None of this is particularly surprising. Mergers and mass restructuring abound, and “we should probably have an AI strategy” is no longer enough justification for an unlimited tab.

But while companies scrutinize what they are spending on the newest layer of technology, there is another expense that has been sitting quietly inside most organizations for years: bad data. There is no line item in the annual budget that says, “Employees manually reconciling numbers nobody trusts: $184,000.” Instead, that expense gets distributed across salaries, meetings, engineering work, software licenses, consulting engagements, delayed decisions, and all the little workarounds people build because the official answer is not quite good enough.

In other words, you probably already have a bad data budget line item. You just don’t call it that.

I started working in business intelligence long before every organization was trying to figure out where to put generative AI. The technology has changed significantly since then, but one pattern has remained remarkably consistent:

A report produces a number. Someone looks at it and says, “That doesn’t seem right.” They download the data into Excel to check it. Someone from another department runs their own calculation because their team defines the metric slightly differently. Eventually, enough people have produced enough versions of the answer that a meeting has to be scheduled to determine which one is correct.

By the time that meeting happens, the company has already paid several people to answer the same question. (If this pains you and you enjoy that sort of thing, feel free to read last week’s article).

Sometimes the report really is wrong. Sometimes the underlying data is perfectly fine and the disagreement comes from two departments using different definitions or calculations. Sometimes a business rule that worked beautifully five years ago has quietly become obsolete as the organization changed around it. Regardless of the cause, the organization is still paying for the confusion, which is why I think “data quality” can be an unfortunately sterile way of talking about the problem. It makes bad data sound like technical housekeeping, something the data team should clean up when they have a free afternoon.

Worse, the cost rarely stays inside the data team. An analyst spends half a day rebuilding a calculation. A director postpones a decision because she isn’t confident in the result. A developer gets pulled away from planned work to figure out why two reports disagree. Leadership decides the current reporting environment must be the problem and buys yet another tool, which eventually gets fed the same definitions, processes, and inconsistencies that made the first system unreliable. Now the company has paid for the original problem and the shiny new container it lives in.

One of the most expensive consequences is also one of the easiest to normalize: duplicate work. People still need answers even when they don’t trust the official source, so they build parallel systems. Finance has a spreadsheet. Operations has its own query. Marketing keeps a tracker because the CRM never quite matches what they expect. An executive has an analyst she trusts, so anything important gets sent to that one person for a second opinion before anyone acts on it.

That is why I think there is a more useful question than, “How clean is our data?”

Ask what the data is costing you.

Not your data warehouse costs or what you paid for Power BI licenses. How much are you spending because people have to verify, rebuild, reconcile, explain, or circumvent information the organization already paid to produce? How much employee time goes into producing the same answer twice? How many decisions wait because someone needs to check the number first? How often is new technology being purchased to compensate for a problem that is actually buried in business logic, ownership, or process?

Now layer AI on top of that environment and the economics get even more interesting.

Companies are increasingly aware that AI usage is not free. Depending on the system, they are paying through tokens, API calls, compute, platform licenses, or some combination of all four. That has made AI costs visible in a way that many older information problems never were. But before you spend too much energy optimizing your AI bill, it is worth asking a much simpler question: what are you paying the AI to read?

If your documentation is outdated, your metric definitions conflict, your knowledge base contains duplicate information, and critical business rules exist mostly inside the heads of employees who have been there for fifteen years, an AI system does not magically transform that mess into institutional wisdom. It processes what you give it. Worse, it can turn contradictory or incomplete information into an answer that sounds remarkably polished.

That is the part I find particularly important. The “old” version of a bad data problem was often visibly messy. Two spreadsheets disagreed. The dashboard looked strange. Somebody knew enough to raise their hand and say the number seemed off.

Now, bad data can hide in plain sight. AI can make the output look better without making the underlying information better. A chatbot can produce an elegant answer from conflicting source material and distribute it to hundreds of employees before anyone realizes the premise was wrong.

We have always had some version of “garbage in, garbage out.” AI doesn’t make that principle obsolete. It gives the garbage scale.

Imagine five analysts each spend three hours a week reconciling reporting discrepancies. Fifteen hours does not sound especially alarming in the context of a large organization, but over a year it becomes roughly 780 hours. At a hypothetical fully loaded labor cost of $75 an hour, that one recurring issue costs about $58,500 annually. And that number still ignores the manager reviewing the work, the developer investigating the discrepancy, the executives sitting in the meeting, and whatever value was lost while everyone waited for an answer they felt comfortable using.

This is also why I would not respond by announcing a massive initiative to “clean all the data.” Aside from sounding like a punishment invented specifically for data engineers, it is usually unrealistic. Not every imperfection matters equally. The more useful approach is to identify where unreliable information is already creating an expensive business behavior and start there.

Look for the places where people repeatedly redo work, delay decisions, maintain unofficial reporting systems, or require a small ritual of explanation before accepting a number. Put an approximate cost on that behavior. Then trace it backward. The problem might be the source data, but it might just as easily be an undocumented definition, an exception nobody accounted for, or a process that changed while the reporting logic stayed frozen in time. Fix the things that are actually expensive before trying to perfect everything.

That becomes even more important before automation. Companies are understandably excited about using AI to make information faster and more accessible, but speed and scale are only advantages when the thing being accelerated deserves to be amplified. If the source is unreliable, automation does not remove the problem. It gives the problem distribution.

So before you spend more money making your data faster, smarter, or more automated, figure out what it is already costing you. You may discover that you have been funding a bad data budget for years. The difference is that unlike your AI subscription, nobody ever thought to put a spending limit on it.

If your team is spending too much time reconciling reports, rebuilding calculations, or feeding AI systems information nobody fully trusts, ThinkFig can help. I work with organizations to identify where those costs are coming from, untangle the underlying data and business logic, and prioritize what is actually worth fixing before investing in the next layer of technology.