Jay Limburn is Chief Product Officer at Ataccama, building AI-powered data products that deliver trust and business impact.
Enterprise AI is changing the economics of bad data by shrinking the time organizations have to find and correct a problem before it shapes a decision. For years, companies absorbed a surprising amount of inconsistency because business processes built in delay. Information passed through reports, reconciliations and analyst review before it reached anything consequential, creating room to catch discrepancies before they spread.
AI removes much of that lag. A system that interprets information and recommends or initiates an action can do in seconds what used to pass through several layers of human review. As that cycle speeds up, a problem that once stayed contained can start shaping purchasing, pricing, risk or operational decisions almost as soon as it appears.
Take a manufacturer carrying an incorrect lead-time figure for a critical component. Sitting in a planning system, the error might go unnoticed until a weekly review catches it. Feed that figure to an AI-driven supply chain application adjusting procurement volumes continuously, and its significance changes. By the time anyone finds it, the business may already be paying for expedited freight, idle capacity, excess inventory or missed customer commitments.
Bad data hasn’t become harder to fix. What’s changed is how fast the cost of finding it late can climb once a system can act on it directly. Large organizations will always carry incomplete or outdated information, and treating every defect as equally urgent is neither practical nor economically sensible. The useful question is how much time the organization has before a defect turns into an economic consequence.
The Window For Intervention Is Getting Smaller
Every material data problem has two important moments.
1. The first is the earliest point at which the organization has enough information to know something is wrong. A missing required field is identifiable when a transaction is entered. A duplicate customer record may only become visible once data from several systems can be compared. An unusual transaction needs historical behavior to be recognized as unusual.
2. The second is when letting the defect persist starts creating unacceptable risk. The space between those points is the organization’s window for intervention, and AI is compressing it across a growing share of business processes. A record that once moved through days of review can now be consumed almost immediately by a system that recommends a price, approves a transaction or reallocates inventory.
That doesn’t mean every control should move as far upstream as technically possible. Moving detection earlier carries its own cost: redesigned systems, added validation, slower transactions and new infrastructure to maintain. The economic question is when the expected cost of letting an error travel further exceeds the cost of catching it sooner. For data feeding a high-value automated decision within seconds, the answer may be almost immediately. For information used in a monthly planning process with layers of human review, there’s considerably more room. The architecture should follow the risk rather than the reverse.
Match The Control To The Decision
This is why the standard advice to “shift left” needs qualifying. Finding problems closer to where data originates is often cheaper than tracing them later, but some defects can’t be caught at creation because the evidence needed to recognize them doesn’t exist yet. A potential duplicate might only become visible once records from several business units are brought together. Pushing checks closer to creation simply asks a system to judge before enough context exists to do so reliably.
The right calculation rests on four things: how quickly information can reach a meaningful decision, what it costs if that decision is wrong, how soon enough context exists to catch the defect reliably and what it would cost to move detection earlier. Some decisions are both fast and hard to reverse, including fraud detection, credit decisions and dynamic pricing. The tolerance for a known defect feeding those is correspondingly thin. Other automated decisions reverse easily or carry limited consequences, while plenty of processes still move slowly enough that later validation remains rational.
What matters isn’t whether a system runs on AI. It’s the combination of speed, consequence, detectability and reversibility behind the decision it’s making. Perfect data was never achievable across a large enterprise, and AI hasn’t changed that. What it has changed is when imperfect data starts costing real money.
Controls Can Be Distributed But Definitions Can’t Drift
Quality checks will inevitably run in different parts of the data lifecycle because different risks become visible at different points. The exposure comes when each location owns its own definition of good data. If one application updates its definition of a valid region while a downstream system keeps running the old one, the same record can pass one control and fail another. When both systems make automated decisions based on those diverging definitions, the disagreement stops being a reporting discrepancy and starts producing conflicting actions.
A quality check can execute in several places at once. The business meaning behind it has to stay the same everywhere it runs. This matters more as AI systems pull information from across the enterprise, because a model has no way of knowing that two teams use slightly different definitions of an active customer or an acceptable risk level. It will act on whichever version it receives.
Start With The Decision
AI hasn’t redefined data quality. What’s changed is how much time remains between a data problem appearing and becoming operational.
For every consequential AI-enabled process, leaders should be able to say which data can change the outcome, when a defect first becomes detectable, how quickly the system can act on it and how hard that action would be to unwind. Those answers determine where quality controls belong and how much it’s worth spending to move them earlier.
The advantage won’t come from pushing every check as far upstream as technology allows. It will come from understanding an error’s economic life well enough to intervene before the cost of correcting it becomes the cost of the decision itself.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?