EY Built 50,000 AI Agents. Only 150 Were in Production.
EY Built 50,000 AI Agents. Only 150 Were in Production.
EY has developed more than 50,000 AI agents on its internal EY.ai platform over nine months. As of late 2025, roughly 150 were live in actual production use, according to a third-party analysis of EY earnings materials; EY has not independently confirmed that figure. The gap — 150 versus 50,000 — is the most honest metric in the Big Four firm's enterprise AI story, and the one its marketing materials don't lead with.
The numbers that do get promoted are real. EY.ai EYQ is deployed to more than 300,000 of EY's 400,000 professionals, and the firm says 80% of its people have completed foundational AI training. Its audit AI is embedded in EY Canvas across 130,000 assurance professionals in 150+ countries, covering 160,000 audit engagements and processing over 1.4 trillion lines of journal entry data annually, according to an April 2026 International Accounting Bulletin report. EY's AI-related revenue grew 30% year-over-year in fiscal 2025, and the firm invests more than $1 billion annually in AI platforms and products, per EY's October 2025 earnings release. It has NVIDIA infrastructure, a proprietary agentic operating system, and a stated target of 100,000 live agents by 2028, according to EY's own case study.
None of that contradicts the deployment gap. It frames it.
The question no enterprise buyer or investor should skip is: why? Why would a firm with $53.2 billion in annual revenue, 400,000 professionals, and a dedicated nine-figure annual AI budget build 50,000 agents and ship only 150? The answer is that the hard part of enterprise AI was never the building. Competitors offer their own version of this same dilemma: KPMG has publicly described its approach as "trust-first," meaning it will not deploy an agent until governance controls are airtight end-to-end; PwC has faced comparable integration challenges across its own client-facing agent rollout, per industry reporting. EY's answer has been "platform-first" — build the infrastructure at scale and solve governance in parallel — which produces more agents in development but the same production bottleneck.
Governance, integration, change management, and client sign-off are where agentic AI deployments die — not in the lab. EY is simultaneously an auditor and an AI vendor. Its clients are being sold agents that would, in theory, augment or replace the same professional judgment EY charges for when it performs an audit. The structural conflict is real, even if it is rarely named that way in the marketing materials.
The "service-as-a-software" pricing model EY has begun describing — per Raj Sharma, EY's Global Vice Chair for Tax — is the relevant diagnostic. If EY can charge a fixed software fee rather than billable hours for work previously done by professionals, it has genuinely restructured its own business. If "service-as-a-software" is accounting language for the same time-and-materials engagement with a different invoice line, the AI is an add-on to the existing model, not a replacement of it. That distinction is not yet settled in either direction.
What is settled is that EY's deployment gap is not unique to EY. Gartner forecasts that 40% of agentic AI projects will be cancelled by the end of 2027 due to governance failures, cost overruns, and inability to demonstrate business value. Ninety-four percent of enterprises report that AI sprawl is already increasing complexity, technical debt, and security risk; only 12% say they have centralized control over their agentic AI systems, per LinkedIn research synthesis. EY is not failing. It is running ahead into a problem every enterprise will face.
The 150-live figure is a snapshot, not a verdict. EY Canvas is live at scale in audit. The firm's audit AI is covering real engagements in real countries with real data. By 2028, EY expects its agentic system to underpin all end-to-end audit activities. That is a credible trajectory — and it is also a reminder that "deployment" is not a single moment. It is a sequence, and EY may be halfway through it.
But for anyone evaluating enterprise AI claims — as a buyer, a competitor, or an investor — the number worth remembering is not 50,000. It is 150.