Short answer. Shadow AI is the use of AI tools, mostly public chatbots and generative assistants, by employees without approval, security review, or governance. Inside financial institutions in 2025, ungoverned AI tools outnumbered sanctioned ones by roughly four to one, and one in five organizations that suffered a data breach traced it to shadow AI. It is the AI your risk register does not know exists.
Shadow AI is a governance gap, not a technology. It appears the moment a capable tool is one browser tab away and the approved alternative is slower, narrower, or missing. An employee is not trying to cause harm. They are trying to finish a report, summarize a document, or draft an email faster, and a free assistant does it in seconds. The organization never sees the request, the data that went with it, or where that data now sits.
The pattern is consistent across banks and insurers. Staff reach for AI in exactly the tasks that carry the most sensitive data, because those tasks are the most tedious. Common examples:
Each action feels small. Together they create unmonitored data flows, inconsistent privacy protection, and no record of what left the building. The tools operate outside procurement, security review, and the model inventory, so the institution cannot answer a basic question: what sensitive information has already been shared, and with whom.
For a bank or insurer, shadow AI collides directly with two live obligations. Article 4 of the EU AI Act has applied since 2 February 2025 and requires providers and deployers to ensure a sufficient level of AI literacy among staff and anyone operating AI on their behalf. An employee pasting client data into an unvetted chatbot is the exact failure that literacy duty exists to prevent, and breaches of the Act can carry administrative fines of up to 15 million euros or 3% of global annual turnover, whichever is higher.
DORA adds a second edge. Since it began applying to financial entities on 17 January 2025, every firm must maintain a Register of Information covering its ICT third-party arrangements and manage the risk those providers carry. A public AI tool that no one approved cannot be in that register, cannot be assessed, and cannot be given an exit plan. Shadow AI is, by definition, unregistered third-party ICT risk. GDPR sits on top of both: personal data sent to a consumer AI service is a transfer the institution can neither document nor control.
Shadow AI is unapproved, ungoverned AI use, and in 2025 it outnumbered sanctioned AI in financial firms by roughly four to one.
Larger than most boards assume. Surveys through 2025 found that 49% of workers use AI tools their employer has not sanctioned, and inside financial services the ratio of ungoverned to monitored AI tools ran near four to one. The governance side has not kept pace. IBM reported that 63% of breached organizations had no AI governance policy at all, and only 37% had any policy to manage AI use or detect shadow AI. The result is a wide gap between how much AI is in daily use and how much of it anyone is watching. The tools concentrate in the high-volume back office, treasury operations, payments processing, and document-heavy claims work, where staff carry heavy workloads and the approved tooling is thin. That is precisely where the most sensitive data sits, which is why the exposure is larger than a headcount of AI users would suggest.
IBM’s 2025 Cost of a Data Breach Report put the average shadow AI breach at 4.63 million dollars, about 670,000 dollars more than a breach without it. The extra cost comes from three things: slower detection, wider exposure, and missing controls. Shadow AI incidents took longer to find, spread across more environments in 62% of cases, and 97% of organizations breached through an AI system lacked proper AI access controls. The legal exposure is now concrete too. In the United States, a financial firm filed the first SEC Form 8-K triggered by unauthorized AI use, judging the incident material because of the volume and sensitivity of the non-public information involved.
The model can be identical. The difference is control. Sanctioned AI runs inside a boundary the institution defines: an approved vendor, a data-processing agreement, logged access, retention rules, and a place in the model inventory and the DORA register. Shadow AI has none of that. The same large language model reached through an enterprise contract is governed; reached through a personal account it is not. Shadow AI is therefore not a different technology but the same technology used without the accountability layer that makes it defensible to a regulator. When a supervisor asks which AI systems touch customer data, a governed deployment answers with a register entry and an audit trail. Shadow AI answers with silence, and silence is what turns a routine inquiry into a finding.
Banning AI does not remove it, it drives it further out of sight. The workable sequence is to discover, then replace, then govern. Discover what is already in use through network and SaaS visibility rather than a survey. Replace the strongest shadow tools with a sanctioned alternative that is fast enough to be the default choice, because usage follows convenience. Then govern: add approved tools to the DORA Register of Information, set access controls, and run the AI literacy training Article 4 requires so staff know which data may go where. The target is simple: every AI tool in use is one the institution can see.
Shadow AI is a symptom of a missing structure, so we treat it as a governance problem, not a policing one. Ableneo shipped 34 production AI projects in 2025, 94% of them using large language models, and roughly four in five reached production and stayed there. That work in regulated Central European financial services, with institutions of the profile of ČSOB, Erste, and UNIQA, is built on putting AI inside a boundary of access control, logging, and a clear inventory from the first day. The fastest way to shrink shadow AI is to give people a sanctioned tool that is genuinely better than the one they smuggled in. For the training duty that sits underneath this, see our explainer on AI literacy under the EU AI Act.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.