14 Oct 2026 · Margot Reyes (editorial persona) · Governance and leadership · on threat.cafe

Shadow AI's $670,000: how to read IBM's breach-cost figure before it goes into your budget request

IBM's $670,000 shadow-AI figure is an average difference in breach cost. Among 600 organisations that suffered a breach between March 2024 and February 2025, those with high levels of shadow AI paid about $670,000 more per breach than those with low or no shadow AI. It describes breached organisations only. It does not measure how much a governance programme saves, and it does not show that shadow AI caused the extra cost. It is also one edition old: IBM's 2026 report came out on 29 July 2026, and its press release does not repeat the figure.

The headline metric that changes the question

The number comes from IBM's Cost of a Data Breach Report 2025. The study was conducted by the Ponemon Institute and "sponsored and analyzed by IBM," and it is "based on data breaches experienced by 600 organizations globally from March 2024 through February 2025" (IBM press release, 30 July 2025). In IBM's words, organisations with high levels of shadow AI saw "an average of $670,000 in higher breach costs" than those with low levels or none.

Read the sentence slowly and three limits appear. The population is breached organisations. The statistic is an average. The comparison is between two groups that differ in shadow-AI use and, very likely, in other ways.

So the useful question is not "how much will governance save us?" It is "what does this study let me say to a finance partner without being caught out?"

Why the common reading is incomplete

Most budget decks quote the number as if it were a return on investment: govern shadow AI, avoid $670,000. That reading fails three ways.

It ignores the denominator. Every one of the 600 organisations had a breach. The figure tells you about the cost of a breach that happened. It says nothing about whether shadow AI made a breach more likely, which is the half of the risk equation most budget requests need.

It treats a correlation as a cause. Organisations with widespread unsanctioned AI may also be larger, hold more data, sit in costlier industries or have weaker controls generally. Any of those confounding variables could carry part of the gap. The press materials read for this article do not say whether the comparison was adjusted for them.

It hides the group definitions. Neither IBM's release nor TechRepublic's report on the study defines the threshold between "high" and "low or no" shadow AI, or says how many of the 600 fell into each group. Without group sizes, there is no way to judge how stable the average is.

The figures around it that are safer to cite

The same 2025 study produced governance numbers that make a cleaner argument, because they describe prevalence rather than cost differences:

Figure (2025 edition)                                        Population                        Source
-----------------------------------------------------------  --------------------------------  -----------
63% lack an AI governance policy or are still developing one  Breached organisations (n=600)    IBM release
13% reported breaches of AI models or applications            Organisations in the study        IBM release
97% of those 13% had no AI access controls in place           Orgs with AI model/app breaches   IBM release
1 in 5 reported a breach due to shadow AI                     Organisations in the study        IBM release
37% have policies to manage AI or detect shadow AI            Organisations in the study        TechRepublic
34% run regular audits for unauthorised AI use                Organisations in the study        TechRepublic
Global average breach cost: $4.44 million                     Breached organisations            TechRepublic
US average breach cost: $10.22 million                        Breached organisations (US)       TechRepublic

Note the nested denominator in the third row. The 97% applies only to the subset that reported AI model or application breaches, roughly 13% of the sample. It is not 97% of all organisations.

The data has moved: what the 2026 edition says

IBM published the next edition on 29 July 2026. The study covers 602 organisations and breaches "between March 2025 and February 2026," again conducted by Ponemon and sponsored and analysed by IBM (IBM press release, 29 July 2026). The release puts the global average breach at $4.99 million, up from $4.44 million, and says AI-enabled malicious breaches cost $6 million on average. It does not mention shadow AI.

That absence matters for anyone writing a 2027 budget. A reviewer who knows the newer edition exists will ask why the request quotes the older one. Before citing $670,000, check the full 2026 report for a shadow-AI comparison. If it has one, cite that. If it does not, say the figure is from the 2025 edition and that the 2026 edition did not repeat it.

A second, independent source tracks the prevalence side. Verizon's 2026 Data Breach Investigations Report, which "uses 2025 data," says frequent use of AI tools by employees went "from 15% to 45% of employees in a single year," in a passage about unapproved AI tools at work (Verizon, 19 May 2026). Verizon measures usage, not cost. Paired with IBM, it supports a narrower claim that holds up: unsanctioned AI use is common and growing, and breached organisations that had a lot of it paid more on average.

Competing explanations a CFO will raise

Expect three objections, and answer them before they are asked.

"Big firms have more shadow AI and costlier breaches." Possible. The press materials do not show an adjustment for size or industry. Concede it, and lean on the prevalence figures instead.
"This is a vendor's study." IBM sponsors and analyses it and sells security products. Ponemon conducts it. Say so in the footnote; a disclosed sponsor is not a disqualification, but a hidden one is.
"An average of what distribution?" Breach costs are skewed by a few very expensive incidents, which pull a mean upward. The press materials report averages, not medians. Without the distribution, treat $670,000 as an order of magnitude.

The how-to-cite box

Paste this wording into the request, adjusting the edition if the 2026 report carries a newer figure:

HOW TO CITE
"In IBM's Cost of a Data Breach Report 2025 (Ponemon Institute; 600 breached
organisations; breaches March 2024 - February 2025), organisations with high
levels of shadow AI reported breach costs averaging about $670,000 more than
those with low or no shadow AI. The comparison is among breached organisations
and is not adjusted for causation; we use it to prioritise AI governance, not
as a projected saving. The 2026 edition (602 organisations) reports a global
average breach cost of $4.99 million."

DO NOT WRITE
"Governing shadow AI will save us $670,000."
"Shadow AI adds $670,000 to every breach."
"97% of organisations lack AI access controls."

What the figure cannot do: a worked illustration

Illustrative example: suppose a firm believes it has a 10% chance of a material breach in a year. Even if the whole $670,000 gap were caused by shadow AI, and none of it by size or industry, the expected annual difference would be 10% of $670,000, or $67,000. Real figures will differ, and the causal assumption is the generous one. The point is the structure, not the number. A cost-given-breach figure only becomes a budget figure once multiplied by a probability, and this study does not supply the probability.

Decision consequences, and what would change this reading

Use the figure for prioritisation. It is fair evidence that shadow AI belongs on the list of things a breached organisation wishes it had governed. Do not use it as the return line of a business case. Use the prevalence and governance-gap figures for the argument, and the cost figure as supporting context.

Three things would change this guidance. A full-report table showing group sizes and a size-adjusted comparison would make the figure stronger. A 2026 shadow-AI comparison would replace it. A study that measured breach probability by shadow-AI level, rather than cost given breach, would finally supply the missing half of the equation.

This is general information about reading a published statistic, not financial advice; for a budget model, work through it with your finance partner and the full IBM report.

The threat.cafe Weekly Brew covers agent-security numbers with the source and denominator attached, if you want the next edition's figures read the same way.

Sources

#shadow-ai #breach-cost #ibm #business-case #ai-governance

All posts