Only 8% of large organisations consistently enforce their data protection policies when staff use AI tools, a survey of 1,064 cybersecurity practitioners has found. The 2026 Unified Data Security Report covered organisations with at least 2,500 employees.
Research firm Cybersecurity Insiders ran the survey in early 2026 with support from security company Netskope. Almost all respondents (98%) said their organisation uses AI, and one in five said AI is already built into business-critical work.
The report describes a wide gap between the rules companies write and what they can enforce. Security teams also struggle to trace sensitive data as it moves between applications, and many need days to rebuild what happened after an incident.
AI Policies Lack Technical Enforcement
Two thirds of respondents said their organisation has a formal AI policy. Only 14% enforce that policy through inline technical controls, which sit in the path of the data and act on it directly. Another 28% rely on written policy and guidance, and 25% said enforcement is inconsistent.
Confidence is also low. Just 7% are very confident that sensitive data is not reaching AI tools without proper controls. A further 39% suspect such data is reaching AI tools but cannot confirm it.
Controls are stronger for files than for the text people type into AI tools. More than half (53%) consistently govern file uploads to cloud services, against 25% for copy-and-paste into AI prompts or web forms. For prompts and text submitted to AI tools, the figure falls to 20%.
Personal AI tools used for work are the hardest unapproved activity to detect, according to 51% of respondents. AI assistants built into existing software followed at 40%, and autonomous AI agents and connected tools at 37%. These connected tools include ones linked through the Model Context Protocol (MCP), a standard for connecting AI models to outside data and apps.
Telling approved AI use from unapproved use is also difficult. Only 10% said they can do so clearly across most environments.
Security Tools Remain Fragmented
More than half of the organisations surveyed (58%) run 11 or more separate data security tools or platforms, yet just 7% describe their set-up as completely unified. Six in ten called their approach moderately or highly fragmented.
The report says this fragmentation slows response. Nearly half of respondents (45%) said analysts must check six or more systems to look into a single incident. Analysts can investigate nearly all sensitive-data alerts at 9% of organisations, while 56% investigate half or fewer.
Visibility is weakest for AI tools. Over a third of respondents (36%) reported strong visibility over email, against 10% for AI tools. Real-time tracking of sensitive data between applications is possible at 7% of organisations.
Tracing Data After An Incident
Few teams can rebuild an incident quickly. Just 4% can work out within minutes where sensitive data came from and where it went. For 68%, the task takes days or longer, or the data needed is rarely available.
Breach-reporting rules set tight deadlines. The EU’s General Data Protection Regulation (GDPR) requires organisations to notify regulators of a personal data breach without undue delay and, where feasible, within 72 hours of becoming aware of it. More than a quarter of respondents (27%) take weeks or longer, or rarely have the data, so cannot reconstruct a sensitive data path before such a deadline closes.
Evidence for auditors is also hard to gather. A full chain of custody, a record of who handled data and when, can be produced quickly at 12% of organisations. Only 9% can quickly find all of the data they hold on a given individual across their systems.
Spending Set To Rise
Most respondents expect their data security to change soon. Some 79% said changes are underway or planned within 12 months, and 72% expect investment to increase. Better visibility was the top priority, cited by 44%, ahead of consistent enforcement (39%) and automated investigation (37%).
Respondents cited integration complexity (35%) and budget limits (31%) as barriers. Two in five (41%) also said their own controls create enough friction to push employees towards unapproved tools.
Netskope cited the findings when it unveiled its DataSec Command Center in August. The product pulls data security signals from across a company’s systems into one console.
“The problem isn’t a lack of tools — it’s the massive volume of data, the signals it generates…” said Sanjay Beri, Netskope’s co-founder and chief executive.
IBM Data On AI Breaches
IBM’s 2025 Cost of a Data Breach Report studied breaches at 600 organisations. Some 13% of them reported breaches of AI models or applications, and nearly all of that group (97%) said they lacked proper AI access controls.
Breaches at organisations with high levels of unapproved AI use cost about $670,000 more on average. IBM also found that 63% of breached organisations had no AI governance policy or were still writing one.
Both reports describe AI in wide use at large companies, while control over the data going into it lags behind. Most firms have written rules, but few can enforce them inside AI tools, trace data across their systems or rebuild an incident quickly.
