Agentic AI has crossed the line from experiment to engagement tool in Asia’s audit market, with software agents already planning, executing and summarising procedures on live engagements. The open question that remains – and the one regulators are now organising around – is how audit firms prove that judgment, and the accountability that goes with it, still sits with a person.
The Big Four have settled the first part of the AI deployment challenge. KPMG began integrating agents into its Clara audit platform in April 2025, starting with expense vouching and unrecorded-liability searches. EY embedded a Microsoft-built multiagent framework in its Canvas platform in April 2026, targeting end-to-end coverage by 2028. Deloitte’s Zora and PwC’s Agent OS follow the same design. Meanwhile, the judgment-proving second part of the challenge is being worked out engagement by engagement, in front of regulators who have started to spell out what evidence they expect to find in the file.
‘AI can increasingly handle the first step of labour-intensive activities’
At KPMG China, roll-out is imminent. ‘In the next 12 months, the Hong Kong audit practice will start to take advantage of KPMG Clara intelligence,’ says Alan Yau FCCA, the firm’s audit innovation leader. The enhanced platform bundles ‘agentic planning, data-driven risk assessment, procedure agents and an engagement lifecycle orchestrator’ into day-to-day work, he explains.
What agents do
Andrew Wong, partner at KPMG China, is precise about where the technology sits. ‘AI can increasingly handle the first step of labour-intensive activities, analysing quantitative and qualitative information from our clients and external sources, identifying anomalies and patterns, checking consistency and cross-references, and executing selected standardised procedures.’
Because agents apply the same logic to comparable data, the firm can also ‘take a more standardised approach in the risk assessment and audit response globally, and in the audits for clients in the same sectors’, he adds. Teams can then ‘focus more on exceptions, judgment and the evidence that matters most’.
Vishaan Kumarasamy, audit manager at EY Singapore, sees the same shift from the engagement floor. ‘The work is becoming less about processing information and more about interpreting it,’ he says. However, he emphasises that interpretation cannot be delegated. ‘AI does not make an audit a push-button exercise. When a tool flags something, we still need to ask why.’
‘AI can highlight a risk, but it is our responsibility to understand that risk’
Humans still in the loop
For regulators, the question is exactly where in the chain that ‘why’ gets asked – and by whom.
Wong describes a deliberate hand-off. The agentic planner suggests risks of material misstatement and candidate procedures, but ‘our auditors exercise professional judgment to determine the appropriateness of such risk assessment and select the procedures’. A separate procedure agent then executes those steps, and its outputs still pass through human review before they count as audit evidence.
Kumarasamy draws the line in exactly the same place. ‘The key distinction is between assistance and judgment,’ he says. ‘I am comfortable using technology for high-volume and structured tasks, but significant audit judgments still require close human review. AI can highlight a potential risk; it cannot replace our responsibility to understand that risk.’
This is the vocabulary Singapore’s Infocomm Media Development Authority has written into policy. Its model AI governance framework for agentic AI, launched in January 2026 and updated in May, sets out four dimensions: bounding risks upfront, making humans meaningfully accountable, implementing technical controls and enabling end-user responsibility.
The trail, not the tool
In December 2025, Singapore’s Accounting and Corporate Regulatory Authority issued guidance on delivering quality audits in a technology-driven environment. It treats audit technology as part of a firm’s system of quality management under SSQM 1, meaning firms must manage their tools with the same rigour as they do their people and methodologies.
In its 2025–27 strategic priorities, Hong Kong’s Accounting and Financial Reporting Council has committed to ‘assess the opportunities and risks arising from the use of AI by audit firms,’ while flagging data security, bias and transparency as new risks.
‘The traceability of auditors exercising professional judgment is essential’
Malaysia offers a glimpse of what enforcement looks like when the underlying discipline slips. The Audit Oversight Board’s 2025 annual inspection report sanctioned three firms and eight auditors, criticising failures to assess whether assumptions were reasonable. It predates agentic tools but describes the failure mode regulators fear: an auditor accepting an output without testing what sits behind it.
Wong says KPMG has built its documentation around that risk. ‘The use of firm-approved AI solutions at KPMG is transparent to regulators. The traceability of auditors exercising professional judgment is essential in demonstrating that accountability remains with the audit team.’ Where AI supports an engagement, ‘our documentation needs to make clear what information is used, what audit evidence is obtained and how the final conclusion is reached’.
The junior pipeline
There is a quieter tension underneath all this. The tasks that agents are absorbing first – vouching, cross-referencing, consistency checks – are the same that have traditionally taught associates what evidence looks like. But if the machine does the first pass, where does an auditor learn to distrust it?
‘There is more time to investigate exceptions and evaluate evidence’
Yau argues the trade is favourable. Juniors spend less time on summarisation, basic drafting, cross-referencing and consistency checks, and more on ‘interacting with clients for understanding the business, investigating exceptions and evaluating evidence’.
Kumarasamy adds a competency that did not feature in the traditional training pipeline. Because ‘good AI starts with good data’, auditors need to interrogate what goes into a tool as much as what comes out: which data is used, how it is processed, and how reliable the result is. Data literacy, he says, ‘will become increasingly important for auditors’.
More information
Find out more about the use of AI in the workplace at ACCA’s Accounting for the Future conference, including the session ‘Agentic AI: where it works and where it fails’. Register to watch live on 24-26 November or on demand.
Visit ACCA’s AI knowledge hub with resources and guidance.
