CPA Canada: Why AI agents may be the next big thing in accounting
October 8, 2025
As agentic AI evolves, new possibilities for the future of the profession are emerging
Picture an accounting team of the near future. Alongside human staff, a team of digital colleagues works in the background: one agent processes payables, another reconciles invoices and a third monitors compliance. Instead of waiting for prompts, they spot tasks, act and call for oversight only when needed. This is the promise of agentic AI—systems that can make decisions and act independently rather than simply waiting for prompts or responding to requests.
Balancing risk and opportunity
A recent Deloitte poll of finance and accounting professionals found that 80 per cent say AI-powered tools, such as agents and chatbots, could become standard tools for the profession within five years, and nearly 15 per cent predict adoption in under 12 months. The optimism increasingly centres on agentic AI: technology capable of not only executing routine but complex processes, freeing accountants to focus on strategy and judgment-based tasks.
The challenge is trust. Efficiency alone won’t win over professionals or clients wary of ceding control to autonomous systems. “AI makes mistakes and even if the system itself is designed right, it evolves over time. Security holes and unintended behaviours can emerge,” says Yvonne Zhu, EY Canada’s AI assurance leader and partner, responsible AI risks and controls. “If people over-rely on AI, they won’t be able to distinguish reliable output from flawed output, and that creates risks for the business and the community.” She stresses that skepticism is healthy—and necessary—to avoid this overreliance. Without review, professionals risk leaning on tools they don’t fully understand, undermining the very confidence needed for adoption.
Modernizing processes
So far, the practical applications of agentic AI are limited. Most agents today handle simple tasks once given to junior staff: creating accounts, granting access and processing approvals, which can take minutes instead of hours. The potential lies in scaling up to multi-agent systems that could transform entire workflows. Zhu points to invoice and warehouse reconciliation, where a network of agents could exchange data across systems, flag breaks and resolve them automatically. “When you tie together data agents, workflow agents and knowledge agents, you’re not just automating steps, you’re reshaping the process itself,” she says.
At CPA Canada, Melissa Robertson, principal, research and thought leadership, describes agentic AI as a junior staff member. Where older automation stalled at predictable inputs, agents can now fill missing invoice fields, reconcile mismatches and gather information across systems before closing the loop. AI agents can be instructed to take their own actions, which may include judgment-like decision-making, executing sequences of tasks that can help reduce bottlenecks in ways traditional tools can’t. But the extent of agency the agent has would depend on the constraints placed on them, whether through instruction or programming.
Both experts emphasize the limits of the technology. Mistakes happen and evolving systems carry new risks. Oversight remains essential, though its depth depends on the stakes. Routine reconciliation might be checked weekly, while systems handling entitlements or sensitive approvals demand closer monitoring. Zhu frames it as three layers: governance from the top, preventative testing before deployment and detective controls such as monitoring, alerts and audit trails. These guardrails can mean the difference between a promising pilot and a production-ready tool.
Accountability adds another layer of complexity. If an AI agent makes a costly error, who is accountable? It’s a tough question, Zhu says, but ultimately the organization deploying the system is responsible. Vendors and contractors may be part of the picture, but firms integrating AI remain answerable to stakeholders. That makes third-party risk management as vital as with any outsourced service.
Assuring oversight and trust
Regulators are moving carefully. It is unknown whether the AICPA will issue new standards, but progress is evident based on its development of an AI policy template for nonprofits, educational initiatives and collaborative reports with CPA Canada, who recently partnered with AICPA on a series of AI papers.
In contrast, the ISO/IEC 42001 AI management system (AIMS) has been published, with companies firms seeking certification to demonstrate to clients that they take governance seriously. The measured pace reflects a broader tension: too little oversight undermines trust, too much too soon could stifle innovation.
The workforce impact is another open question. If agents take on repetitive work, what happens to training for junior staff members? PwC has already signalled plans to hire fewer graduates into entry-level roles, citing that the rapid pace of technological change is reshaping how we work. But Zhu argues the shift is less about eliminating roles than changing what firms look for. Critical thinking, communication and oversight skills will matter more—qualities technology can’t replicate. “AI needs subject-matter knowledge to guide it,” she says. “Accountants bring insight into how effective processes should look. That’s not being replaced, it’s being amplified.”
Agentic AI is arriving quickly, but its success depends as much on proper governance, oversight and working in tandem with humans. Firms that treat these systems as black boxes risk eroding the trust on which accounting rests; those that ground adoption in sound assurance practices stand to gain faster, more reliable processes without sacrificing confidence.
Most professionals see adoption as inevitable. The question is whether firms can build enough trust for AI agents to become reliable colleagues. If they can, the next five years may reshape accounting as profoundly as the arrival of spreadsheets a generation ago.
[CPA Canada]
