The Governance Imperative

Ledger, tax and audit systems support defined tasks. AI is different. It can reason across whatever information it receives. In an accounting firm, that may include ledgers, tax affairs, payroll and bank details. The decision to use AI therefore belongs within the firm's governance, not only its technology budget.

1The firm holds other people's confidential numbers

Clients entrust the firm with records that describe their finances in detail. This includes ledgers, tax positions, payroll, management accounts, transaction data and bank details. Protecting that information is central to the engagement.

AI introduces a new question. It is not only where a file is stored, but where its contents are processed and who can access that processing. A system may leave the original file in place while sending its contents elsewhere for inference.

Partners therefore need to know which model is used, where it runs and what happens to the information it receives. These are governance questions because the firm remains responsible for confidential client work.

2AI touches several professional duties at once

Data protection is only one part of the decision. AI use can also affect confidentiality, independence, competence, professional scepticism and anti-money-laundering obligations.

Confidentiality. The MIA By-Laws establish a professional duty of confidentiality. That duty is not the same as legal professional privilege. If the firm sends confidential client information to a third party, it must understand and govern that disclosure.

Independence. In audit and assurance work, AI may affect how evidence is prepared, assessed or reviewed. The firm must consider self-review risk, the evidence chain and whether its independence remains clear.

Competence and professional scepticism. A fluent answer can still be wrong. Accountants must check the source, treatment and figures. The accountant, not the tool, remains accountable for the work.

Anti-money-laundering obligations. Accountants are reporting institutions under Malaysia's AMLA when they carry out specified activities, such as company formation, nominee or trustee services, providing a registered office, or handling client money. Where these rules apply, customer due diligence, suspicious-transaction reporting and the tipping-off offence require particular care. KYC packs and facts relating to a suspicion should not enter a general AI path.

Data protection. Malaysia's amended Personal Data Protection Act places direct obligations on data processors and uses a risk-based approach to cross-border transfers. Singapore's PDPA includes a transfer-limitation obligation. AI processing must be assessed within these existing duties.

3Unofficial AI use needs a practical response

Staff may already use public AI for research, drafting or working with spreadsheets. A policy that assumes no such use exists will miss the real risk.

The concern is not the intention behind using the tool. It is that confidential client numbers may be disclosed without classification, approval or a clear record. The firm needs rules that staff can follow and a private AI option for confidential client work.

4Why a single prompt can be a breach

Pasting client material into public AI is a disclosure. Depending on the content and circumstances, it may engage confidentiality, data-protection or anti-money-laundering duties.

KYC material and facts relating to a suspicion require particular care because tipping-off rules may apply. The absence of legal professional privilege also matters when information is held by a third party.

Simple controls can reduce accidental disclosure. These include clear information classes, approved tools, access rules, staff training and a defined process for exceptions.

5The direction of travel is one way

Accounting information is becoming more structured and more widely processed. Malaysia's MyInvois programme is one example of that continuing digitisation.

At the same time, data-protection duties continue to develop. Clients may also ask where their information is processed and who can access it. Firms should be ready to answer with a clear data-flow map, policy and record of controls.

6The cost of doing nothing, and of doing it badly

A ban without an approved alternative may drive AI use out of sight. Unrestricted adoption can expose confidential client work. Neither gives partners the oversight they need.

A governed approach starts with the work and the information involved. Public AI may be appropriate for non-confidential research or drafting. Confidential client work needs a private route with access controls, review and clear accountability.

7Governance as the enabler, not the brake

Governance enables adoption by removing uncertainty. Staff should know which tools are approved, what information each tool may receive, who reviews the output and when an exception needs approval.

The papers that follow examine where data goes, how to classify it, how deployment affects control, and how retrieval can respect client boundaries. Later papers cover the threat model, professional obligations and the move from pilot to production.

The starting point is simple: the firm remains accountable for how AI is used with client numbers.

---