Managing Data as an Asset: What It Means, and Bad Data Management Examples With Their Costs

Managing data as an asset means giving each important dataset what a company already gives its equipment and cash: a named owner, an entry in an inventory, measured quality, a lifecycle with retention and disposal dates, and a known cost. Citibank and its parent paid $535.6 million in civil money penalties to the OCC and the Federal Reserve in 2020 and 2024, with data governance named in both actions. Unity Software told the SEC that problems including "ingesting bad data from a large customer" would cost it about $110 million in 2022. Public Health England left 15,841 positive COVID-19 cases out of its daily figures. NASA lost the Mars Climate Orbiter to one ground software file that reported thrust in the wrong units.

What managing data as an asset means in practice

The governance structure around these practices (councils, stewards, policy) is covered in the data governance framework. This piece is about what the practices prevent.

Data is rarely an asset on the balance sheet

"Asset" here is a management term. It is usually not an accounting one.

Under IFRS, IAS 38 Intangible Assets defines an intangible asset as "an identifiable non-monetary asset without physical substance." The IFRS Foundation's summary of the standard says the cost of generating an intangible internally "is often difficult to distinguish" from the cost of running the business. For that reason, internally generated customer lists "and similar items" are not recognized as intangible assets. The same summary lists licenses among its examples of intangible assets, so data a company buys under license can reach the balance sheet at cost.

This article does not cover US GAAP. Ask your controller how your own data purchases are treated.

The practical consequence: the value of most enterprise data appears in no financial statement. If you want to know what a dataset is worth, or what it costs when it fails, you have to track that yourself.

Bad data management examples, with documented costs

Each case below comes from the regulator, company or agency involved.

NASA does not give a separate price for the orbiter. Its September 1999 arrival press kit puts the Mars 98 project at $235.9 million for the orbiter and the Mars Polar Lander together, excluding launch vehicles.

What each case says about the practices

Citibank: an owner who cannot see the reports

The 2020 OCC order named data governance alongside risk management and internal controls. Four years later the OCC found that the bank had failed "to meet remediation milestones." Acting Comptroller Michael J. Hsu said "certain persistent weaknesses remain, in particular with regard to data." The Federal Reserve said Citigroup "has made insufficient progress remediating its problems with data quality management" and had "failed to implement compensating controls."

The specific gap the OCC priced in 2024 was a missing link between data quality and regulatory reports. A bank could know a source system had a defect and still not know which filings it touched. That is a lineage problem. Owners need to see, for each known defect, which reports carry it.

On December 11, 2025, the OCC terminated the 2024 amendment. The termination order states that the amendment supplemented, and did not replace, the 2020 consent order.

Unity: no quarantine for data from outside

Unity's 10-Q says challenges with its Operate products "included the consequences of ingesting bad data from a large customer" and "reduced the efficacy of such products." It expected the impact to be "approximately $110.0 million in 2022, with no carry-over impacts in 2023." For scale, Operate Solutions revenue in that quarter was $184.0 million, according to the earnings release.

The filing does not describe the data. It does establish where the failure entered: at ingestion, from a party outside the company. Validate data from customers, partners and vendors against expected ranges and distributions. Hold it apart until it passes. Know which models and products it feeds before it reaches them.

Public Health England: a silent limit in the pipeline

PHE's statement says "some files containing positive test results exceeded the maximum file size" of the process that loads them into central systems. The records that did not fit were not loaded. Of the 15,841 cases, 11,968 should have been reported between September 30 and October 2. All outstanding cases reached the contact tracing system by 1am on October 3. PHE put in a fix that splits large files and started "a full end to end review of all systems."

Press reports at the time attributed the limit to Microsoft Excel's older XLS format. PHE's statement does not name the software.

The control that catches this is a reconciliation count. Records received must equal records loaded plus records rejected, and a mismatch must stop the run or raise an alert. Every hand-off in a pipeline needs one.

Mars Climate Orbiter: a definition nobody enforced

The units were written down. The Mishap Investigation Board (page 16) says the Software Interface Specification required newton-seconds, and "the SIS, which was not followed, defines both the format and units" of the file. The navigation software therefore "underestimated the effect on the spacecraft trajectory by a factor of 4.45." At Mars the spacecraft was about 170 kilometers lower than planned (page 7).

The warning came early. The report (page 18) says the correctly formatted files were not in use until April 1999 and that "almost immediately (within a week)" they showed anomalous data. The orbiter was lost five months later. The board's recommendation for the next mission was "a software audit for SIS compliance on all data transferred" between the two teams.

Check data against the definition where it crosses the boundary. Give each anomaly an owner who has to close it.

The $3 trillion figure

The claim that bad data costs the US about $3 trillion a year is usually cited to a Harvard Business Review article by Thomas C. Redman, published September 22, 2016, and titled "Bad Data Costs the U.S. $3 Trillion Per Year." The article is behind a paywall, and we could not read the method behind the figure.

A national estimate is not a measurement of any company's losses. Leave it out of a business case. Per-organization averages credited to analyst firms deserve the same test: ask for the survey, the sample and the method, and leave the number out if you cannot get them.

Practices to start with

1. List the ten datasets that feed external reports or revenue. Regulatory filings, financial statements, pricing, and anything a model trains on. That is the first inventory.
2. Name an owner for each one. A person, with authority over definitions and the time to act on quality findings.
3. Write the definition, unit and format of every field that crosses a team or company boundary. Put it in the interface specification or data contract, and test incoming data against it automatically.
4. Reconcile counts at every hand-off. Records in must equal records out plus records rejected.
5. Quarantine external data until it passes validation. Customer, partner and purchased data does not reach production models or reports first.
6. Map lineage from each owned dataset to the reports it feeds. When a defect is found, the owner can say which reports carry it.
7. Give every anomaly a ticket and an owner. Track time to resolution.
8. Record what each dataset costs and what it supports. Acquisition, storage, remediation hours, and the processes that stop without it.
9. Set retention dates and apply them.

Scale it through domain owners once the first ten work. The federated operating model describes how.

Start with the data an outsider will check

Citibank's data problems were priced by its regulators. Unity's showed up in revenue. NASA's showed up as a lost spacecraft. Choose the datasets whose errors a regulator, customer or the public will see first. Give each one an owner and a test at every boundary. Record what a failure would cost. That record is the business case for extending the practices to everything else.

Frequently asked questions

What does managing data as an asset mean? Each important dataset has a named owner, an inventory entry, measured quality, written definitions and lineage, a retention schedule, and a known cost and use, managed with the same discipline as physical or financial assets.

Is data an asset on the balance sheet? Usually not. Under IAS 38, internally generated customer lists and similar items are not recognized as intangible assets. Licensed or purchased data can be.

What are examples of bad data management? Citibank's $400 million OCC penalty in 2020 and $135.6 million in further penalties in 2024; Unity Software's roughly $110 million 2022 impact from problems including bad customer data; Public Health England's 15,841 unreported cases; and NASA's Mars Climate Orbiter, lost to a units mismatch.

Related Reading

• Data Governance & Catalog Platforms
• Data Catalog & Metadata Management
• Data Quality & Observability
• Master Data Management (MDM)
• Data Privacy & Consent Management
• Data Governance Framework: A Practical Blueprint for the Enterprise
• Data Governance at Scale: Building a Federated Governance Operating Model
• DataOps and Data Observability: Ensuring Trust in Data Pipelines