What a Wrong Number Costs, and Why Nobody Computes It

Every company knows what its reporting costs. The BI licenses, the analyst salaries, the warehouse bill. It's all in the budget, itemized, argued over every year.

Ask what a wrong number cost them last year and you get silence.

I don't think the silence is only because the math is hard, though the math is hard. The incentives mostly run the other way. The people closest to the number tend to have very little reason to compute it, so the math stays undone until something outside the company forces it.

The one time it got priced

On May 10, 2022, Unity told investors it had two problems in its monetization business. A fault had reduced the accuracy of its Audience Pinpointer ad-targeting tool, and the company had lost the value of a portion of its ML training data, in its own words, "due, in part, to us ingesting bad data from a large customer." The official release cut full-year guidance. The prepared remarks put the combined 2022 impact at approximately $110 million, and described a recovery that had to run through data rebuilding and model retraining before revenue came back, with planned revenue features delayed while teams cleaned up.

Two caveats worth being straight about. The $110 million was a company estimate covering two intertwined problems, a management judgment rather than a forensic accounting of what one bad dataset cost. And Unity's case is a data problem hurting revenue directly through a product, which sits adjacent to my subject here rather than dead center on it. Dead center would be an executive making a costly call off an internal report that was wrong, and that version rarely produces a public number at all.

Unity had to put a number on it because the company was public and the impact was material. Most companies never get that forcing function. The pipeline gets fixed, the quarter comes in soft, and everyone moves on.

Where the cost hides

Most wrong numbers cost nothing. Fair enough. They get caught in review, or they're wrong in a direction no one was going to act on anyway.

The expensive ones are different. They don't look like data incidents. They look like business decisions.

A bad number becomes expensive when somebody acts on it. Maybe they change pricing. Kill a product. Hire too fast. The damage shows up months later in another department, and by then the report the decision ran on has been refreshed forty times, so the version that mattered is gone. The post-mortem blames the market.

Here's a made-up example. Say your contribution margin report runs three points high because refunds never make it into the transformation layer. Nothing crashes. Every chart renders. On the strength of that report you greenlight a promotion on a $10 million product line that was thinner than it looked. The promotion runs for two quarters before finance reconciles against the ledger and finds the gap. Whatever the mispriced promotion cost, it appears nowhere as a data quality cost. It's booked as a marketing result that underperformed.

The error's cost got laundered into a judgment call. And arguing "the number was wrong" after the fact sounds like an excuse even when it's true, so mostly nobody argues it.

The incentives, briefly

The analyst who finds an error in their own report has little upside for surfacing it. Best case, nothing happens. Worst case, everything they've ever shipped is now suspect. The executive who made the call has no appetite for reopening a decision that's already been defended. The vendor sells speed and self-service, and correctness doesn't demo.

The budget process finishes the job. Checking numbers is a visible cost, recurring, sitting in somebody's budget where it gets challenged every year. Being wrong is an invisible cost, attributed elsewhere when it lands. So verification reads as overhead, and being wrong gets treated as weather.

The number you can compute

You probably can't calculate what wrong numbers cost you last year. The attribution trail is gone, and I'd distrust anyone who claims they can rebuild it.

You can estimate what you're exposed to next year, and the exercise fits on one page. Pick the five or six numbers that drive real decisions at your company. For each one, write down four things: the decision it feeds, how the number could plausibly be wrong, what it costs if somebody acts on it while it's wrong, and who checks it today, how often. A filled-in row looks something like: churn rate, feeds renewal pricing, breaks when the CRM dedupe misses a merge, a bad quarter of it misprices every renewal in the book, and the current check is one analyst eyeballing the trend monthly. Then rough out the exposure per number: how likely wrong, times how expensive if acted on. You'll estimate both badly. Estimate them anyway. The page tends to reorder the priority list on its own, because the number with the sloppiest pipeline is rarely the number with the biggest blast radius.

That page is the honest version of a data quality budget. Not "how clean is our data," a question with no stopping condition, but "which numbers can hurt us, and what are we spending to keep those specific ones true."

Unity's number exists because disclosure rules forced the math. Yours exists too. No one is forcing you to compute it, which is a different thing from it being zero.