Setting the boundaries on AI

In September, corporate reporting, taxation and audit professionals came together virtually for the ACCA Technical Symposium 2026, a one-day event that brought ACCA members in Singapore and Malaysia alongside leading experts to unpack key global trends and translate them into practical local and regional insights.

Designed as an efficient, high-impact day of learning, the symposium ranged across IFRS updates and Scope 3 emissions reporting, but it was AI – and the question of how accountants can harness it responsibly – that captured much of the delegates’ attention.

One step at a time

Opening the event with his keynote, ‘AI-Driven Financial Analysis: Tools for Accountants in the Digital Era’, Andrew Chong FCCA, CEO of uPledge AI, drew on a decade of experience building language models for finance, including a stint at EY, where he worked alongside IBM and Microsoft during the earlier ‘deep-learning’ hype cycle.

‘Adopting AI and actually using it properly are two very different things,’ he explained. ‘We forget that AI is software.’ Rather than throwing 20 use cases at a tool, he suggested, firms are better placed to choose one or two, fleshing out the requirements properly and testing with a small group before rolling out more widely.

Chong went on to describe five tiers of AI systems: rule-based automation, machine learning, large language models (LLMs), agents and custom small language models. ‘Do you need an agent? Why do you need an agent?’ he asked. ‘You must test out level one and two before you ever go to number four.’

In his experience, LLMs ‘tend to be highly inaccurate and also highly inefficient’ for many finance tasks, and agents are best used ‘sparingly’, given the control they are handed. An agent-based workflow, he noted, could easily cost ten times more than a simple orchestration flow doing the same job.

When assessing tools, Chong advised looking for retrieval-augmented generation, code-based calculation, built-in verification and native integration with Microsoft or Google environments. In terms of governance, he urged members to read the documentation closely and favour tools that keep metadata logs only.

AI is a journey

Chong then moderated the panel discussion, ‘AI Reporting Revolution – Transforming Financial Reporting, Analysis & Use Cases’. He opened by asking each panellist to share their worst AI output.

Senthilnathan Sampath FCCA, partner, assurance risk and quality at PwC, recalled hallucinations delivered ‘with extraordinary confidence’, with AI citing ‘standards that didn’t exist, paragraphs that were never written’. As he put it, ‘AI can be brilliant at sounding right, which is very different from actually being right.’

Azleen Waris FCCA, chief financial and strategy officer at Ekuinas, told the story of an automated billing project that turned out ‘a bit more correct’ than anyone intended. It overcharged clients and prompted debit notes, although in the end it did help clean up the contract base.

Ngiam Kim Fong FCCA, CFO APAC at IBM Consulting, took a more hopeful view. At IBM’s annual watsonx Challenge, she explained, colleagues’ initial reluctance quickly turned into enthusiasm once they tried the tools for themselves. ‘The more you befriend it,’ she said, ‘it becomes a friend.’

It also hints at where the profession sits. Kim Fong put adoption at ‘five to six’ out of ten, noting that ‘much of this adoption remains at the individual level, rather than at the workflow level’. Encouragingly, she said, the conversation is shifting from ‘How do I use AI to do my job faster?’ to ‘How do we redesign work?’, which she sees as ‘a game-changing opportunity’ for finance to become a growth enabler.

AI in action

Azleen drew on two examples from her own experience. At Permodalan Nasional Berhad, an in-house electronic know-your-customer build let young Malaysians register for Amanah Saham Nasional Berhad funds without visiting a counter. At a nine-country tower company, robotic process automation and Microsoft Power BI anomaly tools delivered site-level profitability analysis.

From the audit side, Sampath explained that the most mature use case is turning unstructured client information into structured, reviewable evidence. ‘Today’s winner is document intelligence,’ he said. PwC’s ChatAssurance virtual assistant, he added, avoids hallucination by answering only from the firm’s own manuals.

Scaling up, though, comes down to trust. Azleen stressed that the data has to be clean. ‘Otherwise, your output is garbage in, garbage out,’ she said. ‘Copilot is not autopiloting the work; it is our co-pilot or co-analyst.’

Asked about inexperienced staff taking hallucinations at face value, Sampath suggested the answer lies not in banning AI but in strengthening ‘training, supervision and verification’. ‘Modern aircraft can fly largely on autopilot,’ he added, ‘but the pilot still remains accountable for the flight.’ Kim Fong agreed, noting that ‘AI should be supporting the process, not owning the decision’.

What’s the problem?

For SMEs, the panel’s advice was consistent. ‘Don’t start with AI. Start with the business problem to solve,’ Kim Fong advised, describing IBM’s approach of eliminate, simplify, automate and only then embed AI. Azleen’s own framework is ‘people, process, platform’; she pointed out that lean teams can rely on software-as-a-service tools such as Xero rather than a full SAP suite.

On governance, Sampath said that ethics cannot simply be handed to a committee. ‘Culture is what people do when no one is checking the checklist,’ he said. Kim Fong recommended keeping an inventory of AI use, with controls aligned to risk, while Azleen shared that Ekuinas had just approved a group-wide AI governance policy supported by data stewards in each department.

Kim Fong closed with a compelling example. IBM’s revenue-recognition tool, Luca, initially had limited accuracy in identifying non-standard contract terms. Through sustained training and refinement, it evolved into a highly reliable solution and now analyses 90% of contracts within 15 minutes. ‘AI is not magic,’ she said. ‘It is a journey.’