Short answer: what should AI do in a finance team?
AI fits best where work is repetitive, rule-based and checkable: reading invoice data, suggesting ledger accounts, matching bank statements to the books, flagging unusual transactions and drafting management-report commentary. Work that creates legal or cash consequences — posting final entries, filing and paying tax, approving payments, signing financial statements — must still be reviewed and approved by an authorised person.
That is also the principle in Vietnam's Law on Artificial Intelligence No. 134/2025/QH15: Article 4(2) states that AI serves people and does not replace human authority and responsibility, and that human control and the ability to intervene in system decisions must be maintained. The rest of this guide moves from a do / don't list to a control matrix and the Vietnamese legal milestones businesses should know in 2026.
This guide is written by EYC, an AI and technology company that builds AI automation and AI Agent workflows. We are not an accounting, audit or tax advisory firm.
Five kinds of finance work AI does well
What they share: structured or semi-structured inputs, outputs that can be checked against source documents, and errors that can be caught before they cause harm — provided a review step exists.
1. Capturing and extracting invoice and document data
E-invoices in Vietnam are already structured data in the tax authority's standard format, so the main job is getting them into the right place, checking for duplicates and linking them to purchase orders and contracts. For scanned or PDF documents (payment vouchers, contracts, statements), recognition and language models can extract fields such as number, date, tax code and amount. Each field should carry a confidence score so low-confidence fields go to a reviewer.
2. Bank, receivables/payables and intercompany reconciliation
AI or rules can match statement lines to receipts, payments, sales invoices and open balances by amount, date and transfer description. Confident matches are marked; mismatches and partial matches go to a work queue for an accountant, with a suggested reason.
3. Suggesting categories and ledger accounts
Using posting history and the chart of accounts the business applies, AI can suggest accounts, cost codes and cost centres. These are suggestions, not journal entries: an accountant confirms them before posting, especially when the company is moving to a new accounting regime or changing its chart of accounts.
4. Flagging anomalies
Signals worth a second look: duplicate invoice numbers, a new supplier receiving a large payment, a change to a supplier's bank details, costs far above the usual range, transactions at odd hours. AI helps prioritise what to check; deciding whether something is an error or fraud remains a human call.
5. Drafting reports and explanations
Language models can draft commentary on revenue, cost and cash-flow movements for management reports, or summarise new circulars and official letters for the chief accountant. Every number in a draft must trace back to the books, and every legal citation must be checked against the source text, because models can misquote or rely on rules that are no longer in force.
What AI should not decide on its own
The line is not technical difficulty but accountability: who answers to the law, the tax authority, auditors and owners when the result is wrong.
- Posting final entries and closing the period: AI prepares, an accountant confirms and an authorised person closes.
- Preparing, signing and filing tax returns; issuing and digitally signing e-invoices: digital signatures carry the legal responsibility of the company and its representative and should not be handed to an autonomous agent.
- Approving and releasing payments, especially when the payee's bank details have just changed.
- Accounting estimates and professional judgement: provisions, accruals, impairment, revenue recognition on complex contracts.
- Signing and issuing financial statements or reports to banks and investors.
- Credit decisions, credit limits or declining transactions that directly affect customers and partners.
A control matrix for AI in the accounting workflow
The table below is a starting point for internal discussion. It does not replace each company's financial policies and internal control procedures. The "Automation level" column describes the maximum worth considering in an early phase.
General principle: separate the person who configures the AI, the person who confirms its output and the person who approves payments, and log every action by both AI and people. See also access control and audit logs for AI Agents.
| Process | What AI does | Automation level | Approver | Key controls |
|---|---|---|---|---|
| Inbound invoice capture | Pulls data, checks duplicates, matches to orders | Automatic, exceptions to a person | Payables accountant | Duplicate-number check; match against invoice data on the tax authority's system; logging |
| Scanned/PDF document extraction | Reads fields, scores confidence | Semi-automatic | Accountant | Confidence threshold; periodic sample checks; keep originals |
| Bank reconciliation | Matches transactions, proposes reasons for differences | Auto-match confident items; differences to a person | Treasury/payments accountant | Daily or period-end difference report; second-person review |
| Posting suggestions | Proposes accounts and cost codes | Suggest only | General ledger accountant | No posting without confirmation; track override rate |
| Anomaly flags | Scores risk, prioritises a queue | Alert only | Chief accountant / internal control | Documented handling of each alert; periodic rule review |
| Payment approval | Checks supporting documents, warns on changed bank details | Not automated | Authorised signatory per policy | Two approvers; verify bank-detail changes through an independent channel |
| Tax filing, invoice signing | Prepares figures, cross-checks | Not automated | Chief accountant, legal representative | Digital signature held by a person; reconcile ledger to return before filing |
| Management reporting | Drafts commentary on movements | Semi-automatic | Chief accountant / CFO | Figures trace to the ledger; label AI-drafted text when shared internally |
Vietnamese legal milestones to know in 2026
This section summarises rules that directly affect the data AI will process. Regulations change quickly; always check what is in force at the time you apply them.
E-invoices: Decree 254/2026/ND-CP replaces Decree 123/2020/ND-CP
Decree 70/2025/ND-CP amended Decree 123/2020/ND-CP and took effect on 1 June 2025. From 1 July 2026, however, Decree 254/2026/ND-CP (issued 30 June 2026 under Tax Administration Law No. 108/2025/QH15) is in force and, under Article 43(2), ends the effect of Decree 123/2020/ND-CP, Article 1 of Decree 41/2022/ND-CP and Decree 70/2025/ND-CP. On the same day the Ministry of Finance issued Circular 91/2026/TT-BTC with implementing guidance.
The practical consequence for AI projects: invoice-checking rules, prompt templates, internal documents and the knowledge base an AI assistant searches must be updated to the new legal basis. An assistant fed with old documents can answer confidently under Decree 123 — an error that is hard to spot without human review.
Retention of accounting records: Article 41 of the Accounting Law
Article 41(5) of Accounting Law No. 88/2015/QH13 (unchanged in the 2026 consolidated text) sets retention periods: at least 5 years for records used for management and operations, including vouchers not used directly for bookkeeping and financial statements; at least 10 years for vouchers used directly for bookkeeping and preparing financial statements, accounting books and annual financial statements, unless other law provides otherwise; and permanent retention for records of historical significance. Article 41(3) requires records to be archived within 12 months after the end of the annual accounting period.
For AI systems this means: original documents must be kept intact and not replaced by AI-extracted data; extracted data, posting suggestions and approval history should be retained for the same period as the documents so you can explain decisions during inspections; and software and cloud providers must be able to support these retention periods.
AI Law 134/2025/QH15: an 18-month transition for the finance sector
The Law took effect on 1 March 2026 (Article 34). Under Article 35(1)(a), AI systems already in operation before that date in health, education and finance have 18 months for providers and deployers to meet their compliance obligations; other systems have 12 months (point b). That article does not define what counts as the "finance sector", so businesses using AI for internal accounting should watch implementing guidance to see which group they fall into.
A company using AI in its business is a "deployer" under Article 3(5). Its specific obligations depend on the risk level (Article 9) assigned by the provider. See the full analysis in Vietnam's AI Law 2025.
A 5-step rollout for small and mid-sized businesses
- Step 1 — Pick one narrow, measurable process: for example inbound invoice capture or reconciling one bank account. Record current processing time and error counts as a baseline.
- Step 2 — Clean and connect data: supplier master data, chart of accounts and cost codes must be consistent across accounting software, bank feeds and purchasing. See integrating CRM, ERP and accounting before AI and the data readiness checklist for AI.
- Step 3 — Design the human checkpoints: set confidence thresholds, amount thresholds and exception types that must go to a person. See the human-in-the-loop model.
- Step 4 — Run in parallel: let AI process while accountants keep the existing method for at least one closing period, and compare results line by line before removing manual steps.
- Step 5 — Operate and review: track the override rate of suggestions, true and false alerts and incidents; review rules whenever new regulations or accounting-regime changes arrive.
Common risks and how to reduce them
- Fabricated figures or citations: require every number to link to a ledger line or source document; let a regulation assistant search only a vetted document library that records effective dates.
- Data leakage: keep customer, payroll and contract data out of public AI tools — ICAEW advises keeping client and confidential internal data off public AI tools. Data containing personal information is also subject to personal data protection law.
- Over-reliance: the IESBA Code (section R112.1) explicitly names compromised objectivity caused by undue reliance on technology; keep manual sample checks as a habit.
- Payment-redirection fraud: never let AI or automated email change a supplier's bank details; verify through an independent channel.
- Outdated rules: tag every rule with its effective date and schedule a review when new decrees and circulars are issued.
Frequently asked questions
Can AI replace accountants?
Not in the sense of carrying responsibility. AI cuts data entry, matching and drafting, but professional judgement, approval and signing still need an authorised person. Vietnam's 2025 AI Law (Article 4) also states that AI does not replace human authority and responsibility.
Should we use ChatGPT or another public chatbot for bookkeeping?
Do not put vouchers, customer data or payroll into public tools. If you use a language model, use a business environment with access control, logging and data-processing commitments, and treat the output as a draft to be checked.
Which decree governs e-invoices in Vietnam from 1 July 2026?
Decree 254/2026/ND-CP, with Circular 91/2026/TT-BTC. Decree 123/2020/ND-CP and Decree 70/2025/ND-CP ceased to be effective on that date. Update this legal basis in the rules and documents your AI uses.
Can AI-extracted data replace original documents for retention?
It should not be treated as a replacement. Article 41 of the Accounting Law sets retention periods (5 years, 10 years or permanent depending on the record); original documents must be preserved, and extracted data is a supporting layer for search.
Where should a small business start with AI in accounting?
Start with one high-volume, low-judgement process such as capturing incoming invoices or reconciling bank statements. Measure processing time and errors beforehand, run in parallel with the old method for at least one closing period, and only then reduce manual steps.