Cost of bad data statistics: 82 figures on poor data quality, traced to the source

The cost of bad data is usually quoted as $12.9 million a year for the average organization, a Gartner figure from 2020, or $3.1 trillion a year for the US economy, a number credited to IBM that we traced to one writer's rough extrapolation in 2011. Newer surveys measure it in other ways. In 2023, data professionals estimated that data quality issues affected 31% of their company's revenue, and a 2024 survey tied a loss of 6% of revenue to AI models built on bad data.

Below are 82 statistics on what poor data quality costs, grouped by where the damage shows up. Each one is numbered so you can link to it, and each links to the organization that published it. We checked every figure on the publisher's own page or report in September 2026.

The famous cost of bad data figures, traced#

Most articles on the cost of poor data quality repeat the same few numbers. Several are more than ten years old, and a few measured something narrower than the version that gets quoted. We followed each one back to where it started.

Even Gartner's own number has moved. Its research has put the average annual cost of poor data quality anywhere from about $8 million to $15 million since 2009. Each figure came from a different group of organizations, so read them as separate estimates rather than a trend.

What bad data costs organizations#

1. Organizations estimate that poor data quality costs them $12.9 million a year on average, according to Gartner's 2020 survey of 154 reference customers of 16 data quality vendors. It's the most quoted cost figure, and it's a self-estimate by companies already buying data quality software (Gartner, 2020).

2. Gartner's 2017 Data Quality Market Survey put the average annual cost of poor data quality at $15 million per organization (Gartner, 2018). A Gartner article in January 2017 had cited $9.7 million (Gartner, 2017), and earlier surveys found $14.2 million in 2013 (Gartner, 2013) and more than $8 million in 2009 (Gartner, 2009).

3. Nearly 60% of organizations don't measure the annual financial cost of poor data quality, a Gartner survey found in 2017. Gartner's current guidance says 59% don't measure data quality at all (Gartner, 2018).

4. More than a quarter of data and analytics employees who see poor data quality as an obstacle to data literacy estimate their organization loses more than $5 million a year to it, and 7% put the loss at $25 million or more (Forrester, 2024).

5. Thomas Redman estimates that bad data costs most companies 15% to 25% of revenue, building on Experian research and the work of two consultants. He also estimates that about two-thirds of that cost can be found and removed for good (MIT Sloan Management Review, 2017).

6. Data professionals estimated that data quality issues affected 31% of revenue on average in 2023, up from 26% in 2022, in a survey of 200 people commissioned by Monte Carlo (Monte Carlo and Wakefield Research, 2023).

7. 95% of businesses have seen impacts from poor data quality. They include less reliable analytics (36%), a worse customer experience (32%) and damage to reputation and customer trust (32%) (Experian, 2021).

8. 62% of marketing organizations say they lose revenue directly because of poor CRM data quality, and nearly a third of teams spend six or more hours a week fixing and reconciling data, in a survey of 500 marketers (Validity, 2026).

9. Thomas Redman's rule of ten holds that a unit of work costs 10 times as much to finish when its data are flawed as when they're perfect. By that rule, 11 flawed records in a batch of 100 nearly double the cost of the batch (Harvard Business Review, 2017).

What it costs the US economy

10. The $3.1 trillion a year figure for the US economy, usually credited to IBM, first shows up in a 2011 SOA World Magazine article whose author extrapolated it from a healthcare estimate (Tibbetts, 2011). An IBM infographic repeated it by 2013 without a method (IBM, archived 2013), and it spread after Thomas Redman cited IBM in Harvard Business Review in 2016.

11. TDWI estimated in 2002 that poor customer data cost US businesses $611 billion a year, often rounded to $600 billion. The estimate covered only postage, printing and staff time wasted on bad name and address records (TDWI, 2002).

How much data is wrong#

12. 47% of newly created data records had at least one critical error, on average, across 75 data quality assessments collected over two years from a wide range of organizations. Only 3% of the assessments scored 97% or better, the bar the authors set for acceptable (Nagle, Redman and Sammon, Business Horizons, 2020).

13. Between 10% and 25% of customer and prospect records include critical data errors at any given time, from strong organizations to typical ones, according to a SiriusDecisions brief from 2008 or earlier (SiriusDecisions).

14. Organizations believe about a third of their customer and prospect data is inaccurate in some way, and only 50% think their CRM or ERP data is clean (Experian, 2021).

15. Data and analytics leaders estimate that 26% of their organization's data is untrustworthy, in Salesforce's 2025 survey of 7,652 people in 18 countries (Salesforce, 2025).

16. Poor data quality was the top challenge for 57% of data professionals in 2024, up from 41% in 2022 (dbt Labs, 2024). It was still the most reported problem in 2025, at 56% (dbt Labs, 2025).

17. 64% of data and analytics professionals called data quality their biggest data integrity challenge in 2024 (Precisely and Drexel LeBow, 2025 Outlook).

18. 81% of AI professionals say their company still has significant data quality issues, and 85% say leadership isn't addressing them, in a survey of 500 US data professionals commissioned by Qlik (Qlik and Wakefield Research, 2025).

19. At least 23% of an email list decays within a year, ZeroBounce estimates from the more than 11 billion addresses it checked in 2025, down from 28% in its previous report (ZeroBounce, 2026).

Time spent finding and fixing bad data#

20. Knowledge workers waste 50% of their time in what Thomas Redman calls hidden data factories, hunting for data, correcting errors and checking data they don't trust. It's his estimate, and the article gives no study behind it (Harvard Business Review, 2016).

21. Organizations say their data scientists spend 67% of their time preparing data rather than building AI models, in a 2024 survey of 550 respondents (Fivetran and Vanson Bourne, 2024).

22. Data professionals spent about 45% of their time loading and cleansing data in 2020 (Anaconda, 2020), and 37.75% on data preparation and cleansing in 2022 (Anaconda, 2022).

23. Data teams handled 67 data incidents a month on average in 2023, up from 59 in 2022 (Monte Carlo and Wakefield Research, 2023).

24. Resolving a data incident took 15 hours on average in 2023, up 166% in a year (Monte Carlo and Wakefield Research, 2023).

25. 70% of data leaders said data incidents take more than four hours to detect, in a 2024 survey of 200 US data engineers and leaders (Monte Carlo and Wakefield Research, 2024).

26. 74% of data professionals said business stakeholders found data issues first, all or most of the time, in 2023, up from 47% a year earlier (Monte Carlo and Wakefield Research, 2023).

How far people trust their data#

27. 67% of data and analytics professionals didn't completely trust the data their organization used for decisions in 2024, up from 55% a year earlier (Precisely and Drexel LeBow, 2025 Outlook).

28. A year later the same study surveyed only leaders, and 67% reported high trust, against 33% in the previous survey. The report says perspectives change when only leaders are asked, so the two editions don't make a trend (Precisely and Drexel LeBow, 2026).

29. 71% of organizations with data governance programs report high trust in their data, against 50% of those without one (Precisely and Drexel LeBow, 2026).

30. Only 35% of sales professionals completely trust the accuracy of their organization's data, in Salesforce's survey of 5,500 sales professionals in 27 countries (Salesforce, 2024).

31. Business leaders' confidence in the accuracy of their data is down 27% since 2023, and confidence in its relevance is down 18%, in Salesforce's survey of 552 US business leaders (Salesforce, 2025).

Bad data and AI projects#

AI puts a new price on data quality, because models and agents act on whatever they're given. Asked what holds AI back, most groups put data near the top of the list.

32. 63% of organizations don't have, or aren't sure they have, the right data management practices for AI, in Gartner's July 2024 survey of 1,203 data management leaders (Gartner, 2025).

33. Through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data, Gartner predicts (Gartner, 2025).

34. Organizations lose an average of 6% of global annual revenue, or $406 million, to underperforming AI models built on inaccurate or low-quality data, respondents to a 2024 survey estimated. Their organizations averaged $5.6 billion in revenue (Fivetran and Vanson Bourne, 2024).

35. Organizations whose AI initiatives succeed invest up to 4 times more, as a share of revenue, in foundations such as data quality and governance than those with poor AI outcomes. Gartner surveyed 353 data, analytics and AI leaders (Gartner, 2026).

36. 55% of organizations have avoided certain generative AI use cases because of data-related issues, in Deloitte's survey of 2,770 business and technology leaders (Deloitte, 2024).

37. 84% of data and analytics leaders say their data strategy needs a complete overhaul before their AI ambitions can succeed (Salesforce, 2025).

38. Only 26% of chief data officers are confident their data can support new AI-enabled revenue streams, in IBM's study of 1,700 senior data leaders (IBM Institute for Business Value, 2025).

39. 57% of data leaders see data reliability as a key barrier to moving AI projects from pilot to production (Informatica, 2026).

40. Only 12% of data and analytics professionals said their data was of sufficient quality and accessibility for AI in 2024 (Precisely and Drexel LeBow, 2025 Outlook).

41. 88% of data leaders said their data was ready for AI in 2025, yet 43% named data readiness as their biggest obstacle to AI (Precisely and Drexel LeBow, 2026).

42. Data quality is the most cited challenge to launching AI in production at large financial institutions, named by 84%, ahead of skilled staff at 77% (IIF and EY, 2026).

43. 38% of large enterprises name data quality and readiness as the top hurdle to scaling agentic AI, just ahead of integration with existing systems at 37% (UiPath, 2026).

44. 46% of sales professionals whose teams use AI agents say data quality issues hurt their sales, and they rank manual errors as the top data problem (Salesforce, 2026).

Fines and findings in financial services#

Banks report data to supervisors every day, and supervisors have started pricing the errors. Most of these cases involve reporting data, where one wrong field repeats across millions of records.

45. The OCC fined Citibank $400 million in October 2020 over long-standing weaknesses in risk management, compliance, data governance and internal controls (OCC, 2020).

46. In July 2024, Citi was fined another $135.6 million, $75 million by the OCC and $60.6 million by the Federal Reserve, for insufficient progress on the 2020 orders. The OCC cited a lack of processes to monitor how data quality problems affected regulatory reporting (OCC, 2024).

47. Only 2 of 31 global systemically important banks were fully compliant with all of the Basel Committee's principles for risk data aggregation and reporting, nearly ten years after the principles were published (Basel Committee on Banking Supervision, 2023).

48. None of the 25 significant euro-area banks in the ECB's 2016 review fully followed those principles. In 2024 the ECB called progress since then generally insufficient (European Central Bank, 2024).

49. An internal audit of Washington Federal's 2016 mortgage data found errors in 40 of 100 files, a 40% error rate. The CFPB traced the errors to a lack of appropriate staff, insufficient training and ineffective quality control (CFPB, 2020).

50. The CFPB ordered Freedom Mortgage to pay $3.95 million in 2024 because its 2020 mortgage data submission had widespread errors across numerous data fields. It was the company's second order over this kind of data (CFPB, 2024).

51. Goldman Sachs paid a $6 million SEC penalty in 2023 after more than 22,000 deficient blue sheet submissions over about ten years. Errors of 43 types left missing or inaccurate trade data for at least 163 million transactions (SEC, 2023).

52. Citadel Securities didn't report a 0 in one field for fully canceled orders, which made 31.2 billion of its order events inaccurate in the Consolidated Audit Trail. They were part of about 42.2 billion misreported events from 2020 to 2022, and FINRA fined the firm $1 million (FINRA, 2024).

When bad data made headlines#

Most bad data costs a little every day. Now and then a single wrong value costs a lot at once.

53. Unity Software told the SEC in May 2022 that problems with its Operate ad and monetization products, including the consequences of ingesting bad data from a large customer, would hurt its 2022 business by about $110 million (Unity quarterly report, 2022).

54. A Citigroup trader in London meant to sell $58 million of stocks in May 2022, made an input error, and created a basket worth $444 billion. Controls stopped most of it, but $1.4 billion of stocks were sold before the order was canceled. UK regulators fined the firm £61.6 million (FCA, 2024).

55. Samsung Securities put 2.81 billion of its own shares into 2,018 employees' accounts in April 2018, 1,000 shares for each share held, when it meant to pay a cash dividend of 1,000 won a share. Sixteen employees sold 5.01 million of them, and the stock fell as much as 11.7% that morning (Financial Services Commission, Korea, 2018).

56. A coding issue in an Equifax server changed how some credit scores were calculated for three weeks in 2022, and fewer than 300,000 consumers saw their scores shift by 25 points or more (Equifax, 2022).

57. Public Health England left 15,841 positive COVID-19 cases out of its daily figures between September 25 and October 2, 2020, after a technical issue in the process that loaded lab results into its dashboards (Public Health England, 2020).

58. JPMorgan's task force on its 2012 trading losses found that a new value-at-risk model ran on Excel spreadsheets filled in by copying and pasting from one sheet to another, with data uploaded manually without enough quality control. The errors understated risk, and the report doesn't blame them for the losses themselves (JPMorgan Chase, 2013).

59. Fannie Mae corrected its third-quarter 2003 results after finding computational errors made while applying a new accounting standard. Stockholders' equity went from $16.39 billion as first reported to $17.52 billion, a difference of about $1.1 billion (Fannie Mae filing with the SEC, 2003).

60. NASA lost the Mars Climate Orbiter in 1999 because one ground software file gave thruster data in pound-seconds when the navigation software expected newton-seconds, so each value was off by a factor of 4.45 (NASA Mishap Investigation Board, 1999).

61. Meaning to pay about $7.8 million of interest for Revlon, Citibank wired almost $900 million of its own money to Revlon's lenders in August 2020. Three employees signed off, each believing that setting one field to an internal account would keep the principal from going out. The system needed three fields set (US District Court, Southern District of New York, 2021).

62. A clerical error left mismatched bids in a TransAlta spreadsheet, and the company submitted them in a 2003 auction for New York transmission contracts. It expected the mistake to cost $24 million before tax (TransAlta filing with the SEC, 2003).

Bad data by industry#

Healthcare

63. A duplicate patient record adds an average of $1,950 to an inpatient stay and more than $800 to an emergency department visit, according to 1,392 users of patient-matching software at US hospitals and health systems (Black Book Market Research, 2018).

64. Respondents blamed inaccurate patient identification or information for 33% of denied claims, which they said cost the average hospital $1.5 million in 2017 (Black Book Market Research, 2018).

65. Only 22% of health information professionals had brought their organization's duplicate record rate to 1% or less, and 29% didn't know their rate (AHIMA, 2020).

66. 50% of US healthcare providers name missing or inaccurate data as a top reason claim denials are rising, up from 46% a year earlier, in a survey of 250 people who manage billing and claims (Experian Health, 2025).

67. Medicare fee-for-service had an improper payment rate of 7.66% in fiscal 2024, or $31.70 billion, and 6.55%, or $28.83 billion, in fiscal 2025 (CMS, 2025). CMS says most improper payments across its programs happen when reviewers can't tell whether a payment was proper because the documentation was insufficient (CMS, 2024).

Mortgage and lending

68. The critical defect rate in post-closing mortgage quality control rose to 1.71% in the first quarter of 2026, from 1.38% the quarter before (ACES Quality Management, 2026).

69. Legal, regulatory and compliance problems made up 26.02% of critical mortgage defects in that quarter, just ahead of income and employment problems at 20.07% (ACES Quality Management, 2026). For how lenders check income documents, see mortgage income verification.

Insurance

70. 78% of life insurers say data readiness is the biggest challenge in getting value from AI (LIMRA and Equisoft, 2025).

71. Commercial underwriters spend 17% of their week on manual admin, re-keying and moving between systems, and say about 8% would be reasonable, in a survey of 350 underwriting leaders and underwriters in the US and UK (hyperexponential and Coleman Parkes, 2026). More on reading submissions is on our insurance automation page.

Retail and supply chain

72. 65% of nearly 370,000 inventory records across 37 stores of one retailer were inaccurate when checked against what was on the shelves (DeHoratius and Raman, Management Science, 2008).

73. 69% of orders shipped from brands to retailers contained data errors when their barcode records were checked against RFID tag reads, in a one-year study of more than 1 million items across 5 retailers and 8 brand owners (Auburn University RFID Lab and GS1 US, 2018).

Accounts payable and finance

74. 18.4% of invoices hit an exception on average, and processing one costs AP teams $9.84 and takes 8.2 days (Ardent Partners, 2026).

75. High exception rates tie with slow approvals as the top accounts payable challenge in 2026, each cited by 48% of AP leaders (Ardent Partners, 2026).

76. The median organization sends 92.0% of its customer invoices out error-free the first time, so about 1 in 12 needs rework, across 2,342 organizations (APQC).

77. 18% of accountants make financial errors at least daily, a third make several a week, and 59% make several a month, in Gartner's survey of 497 people in controllership. The errors tracked with low capacity (Gartner, 2024). For invoice-level checks, see accounts payable automation and our accounts payable statistics.

Government payments

78. US federal agencies reported about $186 billion in improper payments for fiscal 2025, up about $24 billion from fiscal 2024, across 64 programs at 15 agencies. Agencies count a payment as improper when documentation is too thin to show it was proper (GAO, 2026).

Where bad data starts, and the 1-10-100 rule#

79. In TDWI's 2001 survey of data warehousing professionals, 76% named data entry by employees as a source of their data quality problems, the most common answer, ahead of changes to source systems at 53% (TDWI, 2002).

80. The 1-10-100 rule says it takes $1 to verify a record as it's entered, $10 to cleanse and de-dupe it later and $100 if nothing is done. SiriusDecisions described it as a rule known in data management circles, and the idea is usually credited to a 1992 book on quality costs by Labovitz, Chang and Rosansky. Treat it as a rule of thumb about when to catch errors, not a measured cost (SiriusDecisions).

81. People keying data once got 0.95% of values wrong in a controlled study where each keyed 1,260 values, against 0.03% with double entry (Barchard and Pace, 2011).

82. Copying data from medical records into research databases had a pooled error rate of 6.57% across 93 studies, against 0.29% for single data entry (Garza et al., 2025).

The calculator below prices only the errors someone finds and fixes, which is the cheap end of the 1-10-100 rule. Put in your own volume. Full error rates by entry method are in our data entry error statistics.

Document AI moves the check to the point of capture. Docsumo reads each field of an invoice, bank statement or form with a confidence score and sends low-confidence fields to a person, with the source line highlighted. It reports 99% field-level accuracy on 250+ document types and 95%+ straight-through processing. On the Enterprise plan, cross-document validation checks one document against another, such as an invoice against its purchase order and receipt.

Our take. In lending and AP, the errors that cost the most are often two documents that disagree, like a pay stub and the deposits on a bank statement, or an invoice and the purchase order behind it. A perfect read of each page can still leave that mismatch in place. Judge any fix by whether the whole file agrees with itself.

How to put a number on your own cost of bad data#

Industry averages rarely survive a budget meeting, and your own numbers will. Start with the errors your team already finds. Count how many get fixed in a month, from a review queue, a ticket log or a sample, then multiply by the minutes each fix takes and a loaded hourly cost. That's the visible part, the $10 tier of the 1-10-100 rule.

Next, list what those errors did downstream in the same month. Payments reversed or sent twice, loan files sent back to underwriting, claims denied and resubmitted. These take longer to price, and they're the costs leadership notices.

Then estimate what nobody caught. Pull 100 recent records, check each against its source document, and apply the error rate you find to your monthly volume. Run the same audit a quarter later, after any change, so you measure the effect instead of assuming it.

How we picked and checked these figures#

We checked every figure on the publisher's own page or filing on September 27, 2026, and used the Internet Archive where a page has moved or blocks automated reading. Most figures are from 2024 to 2026. Older ones are here because people still quote them, and each carries its year and what it measured.

Many sources sell data products. Monte Carlo, Fivetran, Informatica, Qlik, Salesforce, Validity, Experian, ZeroBounce and Precisely sell data software or services. UiPath, Equisoft and hyperexponential sell automation or insurance software, Ardent Partners' AP research is sponsored by AP software vendors, and Black Book and ACES serve the industries they survey. We kept their figures because they're the newest data on these questions, and each figure names its source. Many are respondents' own estimates rather than measured losses, which matters most for the revenue figures.

You're welcome to use any figure here. Link to the original source, and to this page if it saved you the search.

If your bad data starts in documents#

Every survey above counts the damage after bad data is already in a system. If your team keys values from invoices, statements or claim forms, measure how many fields a person corrects today, then test automation against that number on your own files. Book a demo with a few real documents, or start a free trial.

Frequently asked questions#