# The Cost of Bad Data in Marketing: Calculate Your Actual Loss

Canonical URL: https://www.martechdevs.com/blog/cost-of-bad-data

By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo)

Data Engineer and Co-founder, martechdevs

Published: 2025-06-18

Last updated: 2026-09-29

Poor data quality costs you money when it creates unnecessary work, wasteful actions or worse decisions. Estimate that cost from specific failures in your business: duplicate outreach, incorrect audiences, broken lead routing, repeated reconciliation and unreliable revenue definitions. A famous industry statistic can establish that the problem matters. It cannot tell you what your team is losing.

A founder does not need another slide claiming that bad data costs trillions. They need to know which failure to fix first, what the repair will cost and whether the business will notice the difference.

This guide checks the widely cited numbers, then shows how to build a practical loss register. The worked financial example is hypothetical. Its purpose is to make the calculation reproducible, not to imply these losses occur in every company.

## What the famous data-quality statistics actually say

### Gartner's $12.9 million figure comes from 2020 research

Gartner's current data-quality page attributes an average annual cost of at least $12.9 million per organization to research from 2020. Cite that amount and date together. Do not silently replace it with $15 million or describe it as a new 2026 survey. [Gartner's data-quality guidance](https://www.gartner.com/en/data-analytics/topics/data-quality)

An organization-level average does not establish the expected loss for a five-person SaaS company, a marketing department or an individual CRM. The public summary does not provide enough detail to scale the figure to your revenue or headcount.

Use it to introduce the issue. Build your financial case from your own process and cost records.

### IBM's $3.1 trillion estimate is older than many citations suggest

An IBM Redbooks publication states an estimated annual US economic cost of roughly $3.1 trillion from poor data quality. That is an economy-wide estimate, not a marketing budget benchmark. [IBM publication, printed page 192](https://www.redbooks.ibm.com/redbooks/pdfs/sg248238.pdf)

The publisher dates that book to August 2015. Describing the number as something first established by a 2016 marketing-data study would therefore be misleading. [IBM publication details](https://www.redbooks.ibm.com/abstracts/sg248238.html)

Do not divide a national estimate by the number of businesses and present the result as your company's loss. Different sectors, business sizes, definitions and overlapping consequences make that shortcut unsuitable for an investment decision.

### The 15% to 25% revenue range is an estimate, not your forecast

Thomas Redman's November 2017 MIT Sloan Management Review article estimates bad-data costs at 15% to 25% of revenue for most companies, drawing on research and consulting work. It discusses costs arising from corrections, checking and mistakes. [Redman's original article](https://sloanreview.mit.edu/article/seizing-opportunity-in-data-quality/)

That does not mean your marketing function is wasting a quarter of company revenue. Nor does it mean a cleanup project can recover that amount. The cost of a problem, the portion you can prevent and the benefit you can actually realize are different quantities.

For the same reason, a published record-error percentage is not a percentage of revenue lost. One wrong routing field may delay an important opportunity; thousands of harmless formatting inconsistencies may have little economic effect. You need the consequence of the error, not just its presence.

## Start with the decision or workflow the data serves

“Improve data quality” is too broad to assign an owner or calculate a return.

A useful scope sounds like: “Route an eligible inbound demo request to the correct active owner within our agreed response window, without creating a second open task for the same request.”

That scope identifies the data that matters: request ID, account association, territory, eligibility, owner availability, event time and existing task state. It also identifies the business consequence of failure.

Other useful scopes include excluding current customers from a particular acquisition campaign, reconciling settled orders with reported purchases or identifying accounts approaching renewal. They require different checks and different evidence.

For each workflow, write down who uses the output, what they do with it and when the data becomes too old to support that action. A company address can be perfectly valid while being irrelevant to the decision. A subscription status can be correctly formatted while being a week out of date.

The [tracking-plan guide](https://www.martechdevs.com/blog/tracking-plan) helps define events and responsibilities. The [CRM data-hygiene guide](https://www.martechdevs.com/blog/crm-data-hygiene) covers practical record controls. Use a bounded process from either as the first investigation, rather than declaring the entire stack broken.

## Separate four kinds of cost

### Cash paid unnecessarily

These are expenses you can trace to invoices, usage charges or transactions: enrichment requests repeated because records lack stable identifiers, duplicate paid sends, contractor cleanup invoices or an avoidable software tier increase caused by billable duplicate records.

The last example requires the vendor's actual billing rule. Removing 10% of records does not automatically reduce a fixed annual subscription by 10%. A saving may begin only at renewal or after crossing a tier boundary.

Refunds and remediation can also create cash costs, but use the correct treatment. Do not count the same refunded order as both full lost revenue and a separate full cash loss without reconciling what was previously recognized and which variable costs were recovered.

### Staff capacity consumed

Count the time spent diagnosing, correcting, reconciling and repeating work. Use the hours people actually spend, including the reviewer who must recheck a repaired report.

Multiply by an agreed cost or capacity-value rate, then label the result. Unchanged payroll is not a cash saving. Reclaimed hours have operational value when there is useful work to redirect them toward; they do not automatically increase profit by the same dollar amount.

A diary recorded close to the work is stronger evidence than asking people to remember the previous six months. Avoid multiplying a rough estimate from one unusually bad incident across the entire year.

### Commercial outcomes potentially lost

Examples include an ignored qualified request, an unnecessary customer discount or a campaign aimed at the wrong audience. These can matter more than cleanup time, but their value is harder to establish.

A misrouted lead is not automatically a lost sale. A wrong audience member is not proof that every dollar spent reaching them was wasted. Identify the causal mechanism and estimate a range supported by comparable outcomes or a suitable experiment.

Track these amounts as uncertain opportunity exposure until you have evidence. The [attribution ROI guide](https://www.martechdevs.com/blog/attribution-roi) explains why an increase in reported conversions alone does not establish additional sales.

### Exposure to future incidents

A broken rule can remain dormant until the next import or product release. Record the affected workflow, plausible consequence, detection gap and recovery work.

Do not monetize every exposure. If you cannot defend an incident probability or financial severity, present the operational risk clearly without an invented expected-loss number. A critical failure can warrant prevention even when its dollar value is not estimable.

## Build a loss register that does not count the same cost twice

Use one row per distinct consequence, linked to a shared incident ID when several consequences arise from the same failure.

Record the process, defect, affected period, evidence location, number of occurrences, cost type, calculation, owner and confidence. Add whether the amount is observed, estimated or conditional. Keep personal customer data out of a broadly shared financial worksheet; link to an appropriately restricted evidence location instead.

An incident might create a contractor invoice and internal review time. Both can be legitimate, separate costs. But if the contractor invoice already includes the review work, adding the same hours again inflates the total.

Similarly, a team member may report “five hours rebuilding the dashboard” while another reports “three hours reconciling the dashboard.” Establish whether they are describing separate people and tasks or overlapping descriptions of the same work.

Give every proposed saving its own realization condition. A duplicate-record cleanup might require a billing-tier change. A reporting repair might require retiring the old manual report. If the team continues both processes indefinitely, the proposed labor saving may never arrive.

## A worked monthly example for a small marketing team

Assume a hypothetical team documents these distinct recurring costs over a representative month:

- External reconciliation invoice: **$600**.
- Unnecessary billable data requests: **$120**, verified against the usage ledger.
- Internal report repair: **12 hours** at an agreed capacity value of **$60 per hour**, or **$720**.
- Repeated lead-assignment corrections: **8 hours** at **$50 per hour**, or **$400**.

The observed cash cost is **$720**. The internal capacity value is **$1,120**. The combined monthly value burden is **$1,840**, with the two categories kept visible.

The team also finds 15 late demo requests. Leave their potential commercial loss outside the total until there is evidence about qualification, response and what would have happened with correct routing. Do not multiply all 15 by average customer lifetime revenue.

A proposed repair costs **$4,000** upfront and **$150 per month** to maintain. Its initial planning assumptions are a 75% reduction in the identified cash waste and a 50% reduction in the identified internal work.

Those assumptions imply **$540** of monthly cash benefit and **$560** of monthly capacity value. After the $150 maintenance cost, the monthly net cash benefit is **$390**. The combined monthly net value is **$950**.

If benefits start immediately and persist for 12 months, first-year costs are **$5,800**. Cash benefits are **$6,480**, leaving **$680** and a cash ROI of approximately **11.7%**. Including usable capacity gives **$13,200** in benefits, **$7,400** net value and approximately **127.6%** value ROI.

Do not present the second percentage as a cash return. The arithmetic depends on the reduction assumptions, and neither is a result until measured.

Under the same immediate-benefit assumption, cumulative cash crosses break-even at the end of month 11. The broader value view crosses at the end of month 5. A two-month implementation delay, with maintenance still charged throughout the year, reduces first-year cash benefit to $5,400 and makes the first-year cash result **negative $400**.

That timing difference can change the approval decision. The [warehouse ROI guide](https://www.martechdevs.com/blog/data-warehouse-roi) shows how to extend this into a monthly cash-flow and discounted-value model.

## Measure defects against a defined population

Select records from the actual workflow and period you intend to improve. Define critical fields before inspecting results, and preserve the selection method. A hand-picked set of known failures is useful for diagnosis but cannot estimate the population error rate.

For an initial diagnostic, you could review 200 eligible requests sampled across the relevant days and channels. If 24 distinct requests fail at least one critical check, the observed record-failure rate is **12%**. That is a sample result with uncertainty, not proof that precisely 12% of every future request will fail.

Suppose the checks identify 18 ownership errors and 10 missing eligibility values, with four requests in both groups. There are **28 defect occurrences but 24 affected requests**. Adding the two percentages and calling the result a customer failure rate would double count the overlap.

Keep field-level, record-level and outcome-level measures separate. Improving field completeness is useful only if the value is correct enough for the action. Replacing every unknown company size with zero makes a completeness score look better while introducing false information.

Repeat measurement with the same definitions after the fix. Compare comparable periods and channels, and record volume or process changes. Include the cost of running the new controls so the improvement is not overstated.

## Repair the cause and the downstream copies

A one-time cleanup does not stop a faulty form, import or integration from recreating the defect. Find the earliest controllable point where the bad value enters or the correct value changes meaning.

For a duplicated purchase, that might be the event identifier shared between collection paths. For a wrong owner, it might be an outdated territory map. For inconsistent revenue, it may be a definition dispute rather than a broken pipeline.

Then inventory the affected destinations. Correcting a warehouse row does not necessarily retract an audience member, reverse a CRM task or repair a historical report. Define whether each downstream consequence needs recomputation, removal, annotation or no action.

![A defined workflow leads to measured defects, separate cost categories, source and destination repairs, and monitoring of recurring defects and realized benefits.](https://www.martechdevs.com/assets/blog/cost-of-bad-data.svg)

Automated tests help catch recurrence. dbt, for example, supports uniqueness, non-null, accepted-value and relationship checks, along with custom business assertions. Passing those checks proves only the assertions you implemented; it does not establish that every recorded value matches reality. [dbt data-test documentation](https://docs.getdbt.com/docs/build/data-tests)

Assign an owner and response for every consequential alert. A daily red dashboard with no action path becomes another task people learn to ignore. Prefer a smaller set of meaningful checks with clear consequences over hundreds of unowned warnings.

## AI makes source discipline more valuable

An AI assistant can turn inconsistent inputs into a fluent explanation. If the underlying definition of an active customer differs between billing and CRM, a well-written answer can still recommend the wrong audience or misstate churn.

Research on high-stakes AI has documented how data issues compound into downstream failures. Its setting differs from marketing, so it does not supply a marketing loss percentage. It does support taking the data pipeline and organizational handoffs seriously. [Google Research's data-cascades study](https://research.google/pubs/everyone-wants-to-do-the-model-work-not-the-data-work-data-cascades-in-high-stakes-ai/)

For a marketing assistant, include the source, definition, freshness and unresolved conflicts with the answer. Test questions where the correct response is to stop, qualify the result or request a missing definition. A system that always returns a number is not necessarily more useful than one that recognizes insufficient evidence.

Our [AI-ready data work](https://www.martechdevs.com/ai-ready-data) focuses on those inputs and controls. If the recurring defect is in identity, modeling or movement between systems, [warehouse and CDP implementation](https://www.martechdevs.com/data-warehouse-cdp) is the relevant scope to investigate.

## A checklist before funding the repair

- [ ] Name the workflow, decision and owner affected by the defect.
- [ ] Define the eligible population and critical checks before measuring.
- [ ] Keep defect occurrences separate from distinct affected records.
- [ ] Attach evidence to observed invoices, usage charges and work hours.
- [ ] Separate cash savings, usable capacity and uncertain commercial exposure.
- [ ] Remove overlapping costs and benefits from the total.
- [ ] Include maintenance, adoption effort and the delay before benefits start.
- [ ] Fix the entry point and identify downstream corrections.
- [ ] Compare recurrence and realized benefit after implementation.

Start with a defect whose evidence is accessible and whose owner can act. A smaller correction that permanently removes a recurring cost is more useful than a broad quality initiative with no measurable consequence.

## Frequently asked questions

### What is the cost of poor data quality?

It is the unnecessary expense, work and adverse outcomes caused by data that is unsuitable for its use. Calculate it from specific processes and evidence. Keep observed costs separate from estimated lost opportunities, and distinguish the total problem from the portion a proposed repair can prevent.

### What does Gartner say poor data quality costs?

Gartner's public guidance cites an average annual cost of at least $12.9 million per organization from 2020 research. That historical average is context, not a forecast for a particular company or marketing team.

### Does bad data cost every business 15% to 25% of revenue?

No company should assume that range applies to it. The widely cited figure is an estimate published by Thomas Redman in 2017. A defensible budget uses the organization's own failure frequency, costs, affected decisions and realistic improvement assumptions.

### How do I value the time spent fixing data?

Measure distinct tasks and hours, then apply an agreed rate. Label unchanged payroll as capacity value, and specify the useful work that recovered time enables. Claim cash savings only when spending actually changes, such as reduced contractor invoices or avoided paid overtime.

### Should we buy a data-quality tool first?

First identify the failure and its cause. A tool can help monitor and enforce checks, but it cannot settle an undefined revenue metric or assign business ownership by itself. Choose technology once you know the controls, scale and response process you need.

## Sources

1. [Gartner data-quality guidance](https://www.gartner.com/en/data-analytics/topics/data-quality)

2. [IBM Software Defined Environment publication](https://www.redbooks.ibm.com/redbooks/pdfs/sg248238.pdf)

3. [IBM publication date and details](https://www.redbooks.ibm.com/abstracts/sg248238.html)

4. [Thomas Redman on data-quality costs](https://sloanreview.mit.edu/article/seizing-opportunity-in-data-quality/)

5. [dbt data tests](https://docs.getdbt.com/docs/build/data-tests)

6. [Google Research data-cascades study](https://research.google/pubs/everyone-wants-to-do-the-model-work-not-the-data-work-data-cascades-in-high-stakes-ai/)