EY launched a global multi-agent framework for its assurance division on Tuesday, embedding it directly into EY Canvas, the platform used daily by the firm’s 130,000 auditors. The goal, per Marc Jeschonneck, EY’s global assurance transformation leader, is to have 100% of audit activities supported by AI agents by 2028.
The initial framework ships with four agents: a core assistant plus three specialized tools for searching and summarizing documentation and automating administrative tasks. Two more agents are in near-term rollout: one that reviews auditors’ work papers and suggests improvements, and another focused on reconciliation documentation. The system has around 20 core modular capabilities, with that number expanding based on what data is connected and how capabilities are combined.
Jeschonneck told Business Insider the key difference from tools like Microsoft Copilot is integration depth. Unlike Copilot, which requires manual file uploads, the EY system acts as a “one-stop shop” with native access to audit data. The main impact: AI capabilities stop being fragmented across separate tools.
The Junior Staff Problem
The adoption friction EY is being transparent about is instructive. Jeschonneck acknowledged that auditors will “need to have quite a level of experience” to effectively review what reconciliation agents produce. For entry-level staff, the transition period is harder: the repetitive tasks that traditionally taught new auditors how audit work functions are the exact tasks the agents are taking over.
EY’s response is a shift in training method. Instead of learning through repetition across multiple client engagements, new hires will work through realistic audit scenarios with adaptive learning tools and embedded short videos. The logic: skilled graduates “don’t need to do certain things a thousand or 10,000 times before they finally understand how it works.”
On headcount, EY is drawing a line from McKinsey’s playbook. In January, McKinsey CEO Bob Sternfels said his firm now runs 25,000 agent employees. Jeschonneck pushed back directly: “If somebody builds thousands of agents, they probably have not understood how it works.” EY is betting on a small, well-integrated fleet rather than raw agent count.
The firm says it is not reducing headcount. Historical audit volumes may require fewer people, Jeschonneck said, but EY expects to need more capacity on the technology and regulatory complexity side. PwC, by contrast, has cut entry-level US recruitment by a third through 2028. KPMG is piloting vibe coding for tax and compliance this month.
What This Tells Agent Builders
The EY rollout is one of the clearest enterprise production deployments of a multi-agent system at scale published so far this year. A few specifics worth noting for teams building similar systems:
Tight agent count, broad capability. EY launched with four agents, not dozens. Modular capabilities (around 20 at launch) compose on top of a small agent core. This mirrors guidance from Anthropic, OpenAI, and Google to prefer single-agent or lean multi-agent architectures over sprawling fleets.
Deep data access beats surface integration. The distinction Jeschonneck draws between Copilot and EY’s system is native data access. Agents that require users to manually provide context lose the productivity gains. The value is in automatic access to the underlying data store.
The learning curve is real and costs time. EY is building training infrastructure specifically for the transition period. For enterprises deploying agents in any domain where junior staff historically learned through repetition, that transition cost doesn’t disappear: it shifts from learning-by-doing to learning-by-simulation. Budget for it.