EY cybersecurity study: AI spending plans outpace proof

EY’s March 2026 cybersecurity study documented a striking mismatch: US security leaders expected AI to transform their defences, yet fewer than half expressed strong confidence in resisting a major AI-enabled breach. Respondents planned substantial increases in AI’s share of cybersecurity spending over the following two years. These were self-reported expectations, not evidence that larger AI budgets had produced better protection or financial returns.

What happened

According to EY US’s Cybersecurity Roadmap Study release, 96% of senior security leaders considered AI-enabled attacks a significant threat to their organisation. Some 48% estimated that AI had enabled at least a quarter of the cybersecurity incidents their organisation experienced during the preceding year. That measures executives’ assessments of incidents, rather than an independently verified tally of AI-assisted attacks.

The spending intentions were substantial. EY reported that 9% of organisations were allocating at least a quarter of their cybersecurity budget to AI solutions at the time of the survey; respondents expected that proportion to reach 48% two years later. The projected increase concerns the number of organisations crossing a spending threshold. It does not mean cybersecurity budgets, or cash spending on AI, would rise fivefold.

Among senior security leaders already using AI in cybersecurity, 85% said their current cybersecurity budget was insufficient to meet AI-enabled threats. That denominator matters: the finding describes AI users within the sample, not necessarily every respondent. It also records a judgement about adequacy, not a calculation of the funding required to close a specified security gap.

EY reported 99% confidence that strategic use of AI would transform proactive cybersecurity strategies, and separately 99% confidence concerning defensive strategies. Those near-unanimous expectations sit alongside the weaker confidence in defending against a major AI-enabled breach. Belief in a technology’s future usefulness is evidently easier to secure than confidence in today’s controls.

The background

Who took part in EY’s cybersecurity survey?

The online survey covered 500 senior security leaders: 216 C-suite executives and 284 leaders below C-suite level. All were full-time US employees at director level or above, responsible for their organisation’s information security, including data and systems. Their employers had annual revenue of at least $500m. Fieldwork ran from December 19, 2025, to January 8, 2026.

Respondents came from 12 industries, with individual industry samples of either 30 or 50. EY applied weighting to distribute the sample evenly across industries and reported a margin of error of plus or minus four percentage points at the 95% confidence level for the total sample. That stated uncertainty does not turn respondents’ estimates into verified incident data or make their forecasts observed outcomes.

For UK companies and smaller US businesses, the sample sets a clear boundary. This was not a survey of British employers or small firms. Its findings illuminate the priorities of security leaders at large organisations; they do not establish an appropriate AI spending ratio for a business with a much smaller security team or budget.

Does the study show a return on AI cybersecurity spending?

The supplied release contains no measured financial return, quantified reduction in losses or before-and-after comparison of defensive performance. It therefore cannot establish that returns had failed to materialise, either. The narrower, defensible finding is that returns were not demonstrated in the published evidence. Absence of a payback calculation is not a negative payback calculation.

This distinction belongs to the broader evidence gap facing AI buyers. A spending forecast can reveal demand without validating a product. In this case, the missing connection is between allocating money to AI and achieving a measurable improvement in security; the survey describes the former much more clearly than the latter.

What people are saying

EY’s Americas Consulting Cyber Risk Practice Leader, Ganesh Devarajan, argued that adding AI tools to existing defences was insufficient. His stated position was that organisations needed to redesign security architecture around AI, rather than merely automate established defensive tasks. That is EY’s prescription for the confidence gap, not a result tested by the survey.

Devarajan also presented larger budgets as an opportunity to adopt agentic systems capable of complex, multi-step actions across products and systems. The distinction is operationally important: software that takes actions poses a different management question from software that merely highlights a suspicious event. The relevant purchasing decision becomes not just what the system detects, but what it is authorised to do.

Respondents were similarly enthusiastic. According to EY, 97% agreed that their organisation’s competitive advantage over the following two years would be directly linked to the maturity of its agentic cybersecurity defences. That is an unusually strong expression of strategic belief. The release supplies no comparison showing that organisations with more mature agentic defences actually outperformed others.

EY’s position on governance was equally expansive: Devarajan called for governance, compliance and ethics to be integrated with cyber defences and risk management. The release also directed readers towards EY’s AI services. Its recommendations are therefore those of a consulting provider discussing a market it serves, rather than an independent evaluation of competing security architectures.

What happens next

Where did leaders expect agentic AI to expand?

EY recorded the following proportions of leaders saying that each function was, or would be, largely run using agentic AI. The first figure describes respondents’ position during the survey; the second is their expectation for two years later:

• Advanced persistent threat detection: 30% to 62%.
• Real-time fraud detection: 32% to 58%.
• Identity and access management: 23% to 51%.
• Third-party risk management: 25% to 50%.
• Data privacy and compliance: 27% to 48%.
• Deepfake and impersonation defence: 23% to 42%.

These forecasts encompass access permissions, compliance and fraud, not simply faster alert handling. For business owners, that expands the approval question from a technology purchase to the delegation of consequential decisions. The distinction between generating an answer and authorising an action is also relevant to prompt-injection risks in automated purchasing.

How mature was AI cybersecurity governance?

Virtually all respondents reported having an AI cybersecurity governance framework. Among those with one, 98% agreed it had proved essential to responsible AI use. Yet only 20% described their framework as optimised and embedded in organisational culture. Another 51% reported a defined framework implemented in key processes, while 26% reported full rollout and integration across relevant business units.

Those categories distinguish the existence of policy from its organisational reach and maturity. They do not justify treating every organisation outside the top category as ungoverned. For a buyer assessing delegated security decisions, however, a framework operating in selected processes provides different assurance from one integrated throughout the relevant business units.

The release provides no subsequent results against the two-year forecasts. Testing those expectations would require follow-up evidence on actual spending, deployment and outcomes, rather than another statement of confidence. In particular, a larger AI budget share would verify a purchasing shift; it would not, by itself, establish that attacks were stopped more effectively.