Shadow AI use clouds returns and exposes data

Employees’ growing use of unauthorised artificial intelligence tools is making it harder for companies to calculate returns on technology spending while exposing confidential information to systems outside corporate control.

Businesses that invested heavily in generative AI are now confronting a widening gap between formal deployment plans and how staff use the technology. Workers are often turning to free consumer platforms or personal accounts because approved systems are slower, restricted or poorly suited to the task at hand.

This informal adoption, widely described as “shadow AI”, can produce productivity gains that remain invisible to management. It can also distort measurements of whether licensed corporate tools are delivering value because employees may switch between approved and unapproved services without recording how either affects output, costs or working time.

Surveys indicate that 86% of employees use AI for work-related tasks at least once a week, while 58% rely on unapproved or publicly available services instead of company-provided platforms. About 63% consider it acceptable to use AI without clearance from information technology departments, and 60% believe meeting a deadline can justify the security risk.

The behaviour is not usually driven by malicious intent. Staff use AI to summarise documents, analyse spreadsheets, rewrite emails, generate computer code and prepare presentations. The danger arises when contracts, customer records, financial projections, employee details, software source code or strategic plans are entered into services that have not been assessed by corporate security teams.

Research into workplace practices has found employees sharing datasets, personnel information and financial or sales material with unauthorised tools. Such disclosures can breach confidentiality obligations, privacy rules and contractual restrictions even when no conventional cyberattack has occurred.

Companies also struggle to detect the activity because AI platforms operate through ordinary web traffic and resemble standard productivity applications. Traditional data-loss prevention systems were designed to identify files being copied, emailed or transferred. They may not recognise sensitive information pasted into a chatbot, converted into prompts or embedded within AI-generated workflows.

The governance problem has intensified as boards demand evidence that spending on models, cloud computing, licences and specialist staff is producing measurable results. Reported returns vary sharply. Some early adopters claim positive and quantifiable gains, while other surveys show only a minority of organisations achieving significant returns from generative AI or autonomous agents.

A study of large listed companies found that deep integration of AI into business processes remained limited despite widespread experimentation. Only 11% of S&P 500 companies had deeply embedded AI in operations by 2025, while another 10% used it directly in producing goods or delivering services. Researchers found no clear difference in capital expenditure or productivity across the wider sample, highlighting the difficulty of separating real transformation from corporate promotion.

The measurement challenge reflects the way AI is being introduced. Many companies count licences, users or chatbot interactions rather than tracking changes in revenue, error rates, processing time, customer retention or operating costs. Time saved by one employee may also be offset by another worker checking inaccurate outputs, correcting fabricated information or reviewing AI-generated code.

Generative systems produce probabilistic answers rather than predictable results, making their value dependent on the task, data quality and level of human supervision. Gains may appear in faster decisions or improved service rather than direct cost reductions, while benefits can take months to emerge after training and workflow redesign.

Security specialists argue that simply banning public tools can push usage further underground. Employees are more likely to comply when approved platforms offer comparable speed and functionality, with straightforward access and clear rules on what information may be submitted.

Companies are responding by creating inventories of AI applications, restricting access to high-risk services and introducing enterprise accounts that prevent customer data from being used to train public models. Some are also installing systems that identify sensitive prompts, redact confidential details and maintain records of employee interactions.

Training is becoming another priority because many workers do not distinguish between personal and corporate accounts or understand how data retention policies differ between services. Clear guidance is particularly important in finance, healthcare, legal services and human resources, where a seemingly routine request for a summary may expose regulated or personally identifiable information.