/

Data

The cost of bad data depends on how long it goes unnoticed

The cost of bad data depends on how long it goes unnoticed

Data

Data

Mahesh Chandran

Mahesh Chandran

CEO

CEO

Guide cover: The cost of bad data grows until it is found. Why detection time decides what a data error costs, and how to shorten it.

For three weeks in the spring of 2022, a coding issue at Equifax changed how some credit scores were calculated, and lenders made loan decisions on them. The coding fix went in on April 6. The harder work came after, because every score sent in those three weeks had already been used.

The cost of bad data is usually described as a single large number. Gartner’s estimate of at least $12.9 million a year for the average organization is the one most people quote. A number that size is hard to act on, because it says nothing about where the money goes or what would reduce it.

A more useful way to look at it is per incident. Each wrong number costs very little on the day it appears and more on every day after, as more people read it, more decisions rest on it and more systems copy it. This guide shows why detection time drives the cost of bad data, how to measure it for the datasets a business depends on, and how to shorten it.

The short answer

The cost of a data error grows with two things: how long it stays wrong and how far it spreads in that time. Most of the cost builds up before anyone notices, as reports, decisions, customers and downstream systems take in the wrong number. To reduce the cost of bad data, measure how long errors in each important dataset go unnoticed, then shorten that time for the datasets with the highest-stakes destinations first. Fixing errors faster helps less, because by the time an error is found, most of its cost has already been paid.

Key terms

  • Data incident: any period during which a dataset is late, incomplete or wrong in a way that affects someone downstream.

  • Time to detection: the time from when the error first appears in the data to when someone who can fix it knows about it.

  • Time to resolution: the time from detection to when the data is correct again.

  • Exposure window: the whole period during which the wrong data was available to be read, copied or acted on. It runs from the first wrong value to the last downstream correction.

  • Blast radius: everything the wrong data reached during the exposure window, such as reports, decisions, customers, models and other tables.

The method at a glance

  1. Keep a log of every data incident, with when the error started, when it was detected, who detected it and when it was resolved.

  2. Work out the exposure window and the blast radius for each incident.

  3. Rank datasets by where their data lands, and set a detection target for each tier.

  4. Add checks at the source of each Tier 1 dataset, and route their alerts to a named owner.

  5. Review the log every quarter and compare detection times against the targets.

Why does detection time drive the cost?

A wrong number does nothing on its own. It costs money when someone reads it and acts on it, when a system copies it into another table, or when it reaches a customer. Each of those things happens again every day the error stays in place, so the exposure window decides how many of them pile up.

The exposure window has three parts:

  • Before detection. The error is live and nobody knows. Reports go out, models train, invoices print and other pipelines copy the values. This part is usually the longest, and it is where the blast radius grows.

  • Detection to fix. Someone knows and is working on it. New damage slows, especially if readers are told to hold off.

  • Fix to cleanup. The source is correct again, but every copy, report and decision made during the window still carries the old value. Each one has to be found, rerun, reissued or explained.

The exposure window runs from the first wrong value to the last downstream correction, through before detection, fixing and cleanup. The blast radius grows until someone notices.

The cleanup part scales with the first. A fix to the source can take minutes, while finding and correcting everything that read the wrong data can take weeks, and that work grows with every day the error went unnoticed.

Three examples

The published cases show the same pattern.

Public Health England, 2020. A problem in the data load process that moved positive COVID-19 test results into reporting meant 15,841 cases from 25 September to 2 October were left out of the daily figures. The cases were transferred to contact tracing within hours of the problem being found. The delay that mattered was the week before, when those people were not on the list to be contacted.

Equifax, 2022. The coding issue affected some credit scores calculated between March 17 and April 6, 2022. Lenders had already used scores from that period for loan decisions, and Equifax spent the following weeks working with customers on the impact. In 2025, the New York Attorney General secured $725,000 from Equifax over the inaccurate scores.

Unity Software, 2022. Unity told investors it had lost the value of part of its training data, in part because it ingested bad data from a large customer. It estimated the impact at about $110 million in 2022, spread over three quarters while it rebuilt the data, retrained its models and recovered the revenue. When bad data reaches a model’s training set, the cleanup includes rebuilding the model as well.

In all three, the fix to the source was the small part. The cost sat in everything that happened while the error went unnoticed and in the work of undoing it.

The 1-10-100 rule

Quality management has a long-standing rule of thumb for the same idea. The 1-10-100 rule, proposed by George Labovitz and Yu Sang Chang in 1992, says that an error costing $1 to prevent costs $10 to correct later and $100 if it is never caught. The exact ratios are illustrative. The direction holds in every case above.

Where the data lands matters as much as how long

Two errors that go unnoticed for the same week can cost very different amounts. An error in an exploration sandbox reaches a handful of analysts. The same error in a table that feeds invoices, a regulatory filing or a production model reaches customers, regulators or every prediction the model makes. That is why detection targets should be set by dataset, based on where each dataset’s numbers end up. How to decide which datasets your business depends on sets out a method for ranking them.

How do you measure the cost of bad data?

A count of last quarter’s data incidents is easy to produce. It is harder to say how long each one went unnoticed, which is the number that explains the cost. A simple incident log fixes that.

Keep an incident log

Record every incident in one place, including small ones. Each entry needs a few fields:

  • Started: when the first wrong, late or missing value appeared. This usually has to be worked out after the fact from load times and query history.

  • Detected: when someone who could act on it found out.

  • Detected by: the data team, an automated check, or a person downstream such as a finance analyst or a customer.

  • Resolved: when the source data was correct again.

  • Cleaned up: when the last downstream copy, report or decision was corrected.

  • Reached: the reports, teams, customers, models and tables the wrong data touched.

These columns can go straight into a spreadsheet. The rows show one invented incident:




From those fields come three measures:

  • Time to detection: detected minus started. In the example, just over seven days.

  • Exposure window: cleaned up minus started. In the example, three weeks.

  • Detection source: the share of incidents found by a check or the data team, against those found by someone downstream. When customers and finance analysts find most incidents, the checks are in the wrong places.

Put a value on each incident

The exact figure matters less than having one that people agree on. For each incident, count the cost that can be counted, such as hours spent on the fix and cleanup, invoices or filings that had to be reissued, refunds and credits, and decisions that had to be reversed. Note the costs that cannot be counted next to it, such as lost trust in a dashboard. Over a few quarters, the log shows which datasets produce expensive incidents and how much of each incident’s cost built up before detection.

How do you shorten detection time?

Shortening detection time for every dataset at once spreads the effort too thin. Start with the datasets where an unnoticed error costs the most.

Set detection targets by tier

Rank datasets by where their data lands and what happens when it is wrong. Then give each tier a detection target that matches the stakes. For example:

  • Tier 1: data that feeds billing, finance, filings, customer-facing products or production models. Errors should be found within hours, before the next downstream run uses them.

  • Tier 2: data behind dashboards that teams act on. Errors should be found within a day.

  • Tier 3: exploration and ad hoc data. Errors can wait for the person using it to notice.

Detection targets by tier: within hours for Tier 1, within a day for Tier 2, and when a user notices for Tier 3.

Check at the source

An error caught where it enters the pipeline has reached nothing yet. The same error caught in a dashboard has already reached everything between. For each Tier 1 dataset, add checks at the point where its data arrives, and on every table it is built from. The upstream pass finds those tables.

The checks worth having first are the ones that catch the errors behind the examples above:

  • Freshness: the data arrived when expected.

  • Volume: the row count is within the normal range. The Public Health England records were dropped during a data load, which is the kind of gap a volume check against the source is built to find.

  • Reconciliation: totals match the source system, such as the sum of invoice lines against the billing system.

  • Distribution: values such as scores, prices or rates have not shifted beyond their normal range overnight.

Send alerts to someone who will act

A check that fires into a shared channel nobody watches adds nothing to detection time. Route each Tier 1 alert to a named owner who can stop downstream runs or warn readers, and keep Tier 3 alerts out of the way so the important ones stand out.

Tell readers early

Once an error is found, telling the people who read the data shortens the exposure window straight away. A short note saying a dataset is wrong, what it affects and when it will be fixed stops new decisions being made on it while the fix is underway.

Review the log every quarter

Compare detection times against the targets, look at who found each incident, and move checks toward the datasets and sources that produced the most expensive ones.

Questions people ask

What is the average cost of bad data?

The most quoted figure is Gartner’s estimate that poor data quality costs the average organization at least $12.9 million a year, from research published in 2020. It is an average across organizations of different sizes, so it is a poor guide to any one company. An incident log gives a figure based on the company’s own data.

Why is it called bad data when the problem is often late or missing data?

From the reader’s side, the effect is the same. A dashboard showing yesterday’s numbers as today’s, or a total missing a day of records, leads to the same wrong decision as an incorrect value. The incident log should count all three.

Is faster fixing or faster detection more important?

Detection, for most incidents. The fix to the source is often quick once someone knows what is wrong. The time before detection is when the wrong data spreads, and the cleanup afterward grows with it.

Do we need to monitor every dataset to catch errors sooner?

No. Checks on every dataset produce more alerts than any team can read, and the important ones get lost. Put the closest monitoring on Tier 1 data and the tables it is built from, lighter checks on Tier 2, and let Tier 3 data rely on its users.

Who should own detection targets?

The data team owns the checks and the incident log. The business owner of each destination, such as finance for revenue data, should agree on the detection target, because they know what a day of wrong numbers costs them.

Where Dataring fits

We built Dataring around the idea that the cost of a wrong number depends on how long it goes unnoticed. It ranks every dataset by how much the business depends on it, so the closest checks and the fastest alerts go to the data behind billing, finance and customer-facing products. Tests cover the datasets the business depends on, and the pipelines nobody depends on are retired within limits each company sets, under its supervision.

Book a free call to see it on your data estate. For the wider process this guide fits into, read How to audit your data estate in seven steps.

Newsletter

Product news by email. We use your address for nothing else.

© 2026 ShareData Inc.Last updated October 29, 2026