India emerged as one of the fastest growing enterprise adopters of artificial intelligence in 2025, reaching 82.3 billion transactions—a 309.9% year-over-year increase. The data comes from a research report by Zscaler titled “ThreatLabz 2026 AI Security Report 2026.

“India’s growth aligns with continued government-backed digital transformation efforts in 2025, alongside major public and private investment in AI infrastructure and skills development. An expanding AI-enabled workforce, combined with cloud-first architectures that enable fast, scalable deployment of AI services, likely contributed to the country’s outsized growth relative to prior years,” the report stated.

India also drove nearly half of all AI/ML activity, accounting 46.2% of regional traffic, driven largely by the Technology and Communication sector (31 billion transactions).

Globally, the geographic distribution of AI/ML activity remained broadly consistent in 2025, with subtle shifts at the margins. AI is firmly established in the United States—the epicentre of enterprise AI development and deployment— and the country continues to claim the largest share of AI/ML traffic volume, but AI usage grew significantly across several international markets. Although the U.S. continued to lead in absolute usage (218.9 billion AI/ML transactions, accounting for 37.6% of global activity), AI adoption expanded faster year-over-year elsewhere.

Key application findings

Grammarly emerged as the most active AI/ML application in enterprise environments, accounting 38.7% of total transactions, overtaking ChatGPT in total transaction volume. ChatGPT remained a dominant general-purpose assistant, recording 14.2% of total usage, used broadly across roles for research, drafting, and analysis, making it a common touchpoint for enterprise data.

Codeium entered the top five (5%), showing how AI has become a regular part of software development work where source code and proprietary logic are routinely processed.

Blocked transaction trends:

The report also highlighted how organisations also tightened oversight on enterprise AI in 2025. Data exposure, privacy, and compliance concerns pushed them to block 39.2% of total AI/ML transactions, reinforcing AI governance as a standard part of daily security operations.

The applications most impacted by enforcement controls were also among the most widely used AI apps in the enterprise. Grammarly comprised the single largest share of blocked activity—171.2 billion blocked transactions, which amounted to 44.2% of all blocked AI/ML transactions.

“Broad-use AI applications remained under scrutiny as well. ChatGPT and Microsoft Copilot were frequently blocked, seeing 5.7 billion and 4.1 billion transactions blocked, respectively, as access to unstructured data continues to raise the risk of sensitive enterprise information being shared unintentionally,” the report stated.

AI coding assistants, including Codeium and Tabnine, were also commonly blocked to limit exposure of proprietary code and development artifacts. Language and content transformation tools, such as QuillBot and DeepL, faced similar controls, reflecting broader efforts to limit content sharing with external models.

Which AI companies recorded the highest data exposure risks?

ThreatLabz also examined the amount of data transferred between enterprise environments and AI/ML applications. Over the past year, enterprise data transfer to AI/ML applications continued to rise, reaching 18,033 terabytes (TB)—a 93% increase year-over-year.

Grammarly remained the top application by this measure, with 3615 TB of data transferred. It was followed by ChatGPT, with 2021 TB, OpenAI with 865 TB, DeepL with 625 TB, and Codeium, with 387 TB.  

The report also highlighted rising data loss prevention (DLP) violations linked to AI tools. ChatGPT DLP violations increased 99.3% year-over-year, generating 410 million data loss prevention (DLP) policy violations, affirming enterprise risks tied to high-context AI assistants. The most common violations specific to ChatGPT included name leakage and national identifiers—possibly customer records or identity details.

Enterprise DLP violations tied to Codeium increased 100% year-over-year, suggesting increased leakage risk for source code and proprietary logic.

Where failures happen most often

The findings are based on exploit data generated through Zscaler red teaming exercises conducted across more than 25 enterprise environments, covering over 222,000 adversarial attacks, of which nearly 199,000 were completed successfully without error.

“Bias, off-topic responses, and manipulation vulnerabilities remain among the most common failure categories across AI systems,” the report noted. “These weaknesses highlight the gap between enterprise expectations for policy compliance and the current reliability of AI models.”

Bias accounted for 49% of failures, followed by off-topic responses at 47% and manipulation vulnerabilities at 45%. Competitor checks, intentional misuse, and Q&A stability each recorded failure rates between 44% and 45%, underscoring persistent reliability and governance challenges in enterprise AI deployment.