Why RPA seemed like the right answer
Between 2018 and 2023, Robotic Process Automation became the dominant approach to back-office automation across Vietnamese enterprises. The logic was sound: finance teams were spending 3–5 hours per day on manual data entry; RPA bots could execute the same keystrokes in seconds with no integration project required.
Deloitte's Global RPA Survey (2020) found that 53% of large organisations had already started their RPA journey and expected full payback within two years. Vietnamese banks, manufacturers, and distributors followed the same trend, purchasing licences from vendors including UiPath, Automation Anywhere, and Blue Prism at rates that outpaced regional averages.
The problem was not the technology. It was the assumption about the documents.
What RPA was not built for
RPA operates on rules. A rule says: find the value in cell B12 of this Excel export, copy it to field 4 in this ERP screen, and move to the next row. That works when the document format never changes, the source is always digital, and every vendor sends invoices in the same template.
Vietnamese B2B invoicing does not work that way. Vietnam has over 900,000 registered active enterprises (General Statistics Office, 2023), of which the overwhelming majority are small and medium businesses. Their invoices arrive:
- As scanned PDFs of handwritten documents
- In mixed Vietnamese–English formats with inconsistent field labels
- From suppliers using three or four different e-invoice templates (VNPT, MISA, Fast, and home-built)
- With totals written as both numeric and Vietnamese language text
- Under VAS (Vietnam Accounting Standards) requirements that mandate specific fields the international RPA templates do not recognise
"RPA handles the documents you designed. Invoice AI handles the documents you receive."
When an invoice falls outside the rules — which in Vietnamese supply chains happens on 15–30% of invoices — the RPA bot throws an exception. A human handles the exception. The finance team ends up maintaining two parallel processes: bots for the clean documents, people for the rest. Total cost reduction is a fraction of what the business case projected.
The maintenance tax
Even for the invoices that RPA handles successfully, the long-term economics erode. Every time a supplier changes their invoice template — a new logo placement, a reformatted line-item section, a new field required by a tax update — the rule must be rewritten. In practice, most enterprises need 0.5 to 1.0 FTE dedicated to RPA maintenance for every 10 bots in production.
Gartner's 2022 Market Guide for Intelligent Automation noted that organisations underestimate ongoing maintenance costs by a factor of three when evaluating RPA ROI. For Vietnamese finance teams that have grown their bot fleet over several years, maintenance now consumes a material share of the savings the automation was supposed to generate.
What AI-native invoice processing does differently
AI-native invoice processing — specifically systems built on large language models fine-tuned on Vietnamese accounting documents — approaches the problem differently. Rather than rules that describe where a field should be, the model learns what a field means. It can extract a tax identification number from an unstructured paragraph in Vietnamese, reconcile it against a vendor master, and flag discrepancies — regardless of document format.
- Exception rate: RPA typically handles 70–85% of invoices; AI-native systems handle 92–97% without human intervention
- Maintenance: RPA requires rule updates for every template change; AI models improve with new examples
- VAS compliance: AI can validate VAT registration numbers, check e-invoice authenticity via the Tax Authority portal, and flag non-compliant documents automatically
- Multi-language: AI handles Vietnamese, English, and bilingual invoices natively; RPA requires separate rule sets per language variant
- Audit trail: AI logs extraction confidence scores per field, creating audit-ready records; RPA logs executions but not extraction rationale
The transition question: rebuild or layer
Enterprises with existing RPA investments face a transition decision. The practical answer for most is to layer AI on top of RPA rather than replace it entirely. AI handles the extraction and classification of unstructured documents; validated, structured data is handed to the RPA bot for ERP entry. The bots handle clean, structured input reliably — which is what they were designed for. The exception rate drops. Maintenance costs stabilise.
Full AI-native replacement — where the AI connects directly to ERP via API rather than using RPA as the last mile — produces better long-term economics but requires an integration project. For organisations with modern ERPs (SAP, Oracle, Odoo), direct API integration is straightforward. For legacy Vietnamese accounting systems (MISA SME, Fast Accounting), the integration layer matters and should be evaluated carefully before committing to a vendor.
What to measure before and after
The three metrics that matter for any invoice automation project are: straight-through processing rate (invoices handled without human intervention), cost per invoice (fully loaded including exception handling and maintenance), and time-to-close (days from invoice receipt to approved payment entry). All three should be baselined before implementation and measured monthly for the first six months post-deployment. Vendors who cannot commit to these KPIs in their contract are selling software, not outcomes.
Sources
Gartner — "Market Guide for Intelligent Document Processing," Gartner Research, 2022.
Deloitte — "Global Robotic Process Automation Survey," Deloitte Insights, 2020.
McKinsey & Company — "Finance 2030: Four must-make transformations," McKinsey Global Institute, 2023.
General Statistics Office of Vietnam — Enterprise Census 2023, GSO Vietnam, 2023.
Vietnam Ministry of Finance — E-Invoice Implementation Guidelines (Decree No. 123/2020/NĐ-CP), 2020.