Avi Santoso details Q3 research on AI finance automation

Perth software engineer and finance workflow automation consultant Avi Santoso has published a roundup of his Q3 2026 research into using AI to streamline accounts payable and bookkeeping. The work tested invoice tools, read-only bookkeeping workflows and a tuned local model, showing both efficiency gains and current limits for finance teams.

Why it matters: - Santoso’s research points to a practical path for finance teams that want to use AI without giving up human control. - The findings focus on accounts payable and bookkeeping, two areas where small efficiency gains can reduce review time and manual work. - The testing also shows where AI tools still miss details, which matters for businesses handling taxes, receipts and cross-border payments.

What happened: - Perth-based software engineer and finance workflow automation consultant Avi Santoso published a roundup of his third-quarter 2026 research on automating finance workflows. - Santoso studied how businesses can use AI to streamline finance-related work and improve accounts payable and bookkeeping processes. - He said the full Q3 2026 summary is available in his website’s library. - Santoso also made a starter guide available for finance teams looking to use AI and other technologies to improve operations.

The details: - One study compared invoice processing tools including Xero, Hubdoc and Dext across multiple invoice files. - All three tools totalled the invoices correctly. - Only Hubdoc and Dext captured the GST value on every invoice. - Xero missed the GST value on those invoices. - All three solutions misrecorded Australian dollar references on an overseas invoice that was payable in Japanese yen. - A second test examined Xero and Codex on read-only bookkeeping workflows. - The workflow aimed to suggest account codes, flag missing receipts and draft emails requesting documents. - Santoso estimated that review time would fall from 1 hour to around 5 minutes under that approach. - The systems did not post or match automatically and instead waited for human approval. - A third study fine-tuned a small local AI model, Llama 3.1 8B, on 600 synthetic supplier bills. - The model’s accuracy for sorting expense bills into 8 expense categories rose from 70.5% to 99.5% on 200 held-back test examples.

Between the lines: - The research suggests AI can take on narrow finance tasks well before it can run fully automated workflows end to end. - Santoso’s quoted framework emphasizes gradual adoption, with testing, limit-setting and process review before more tools are added. - That approach lowers risk for finance teams that need accuracy more than speed alone. - The results also show that international transactions and tax handling remain common failure points.

What's next: - Santoso’s guidance points toward phased adoption, where teams expand AI use only after checking outputs and tightening processes. - The published starter guide gives finance teams a reference point for future implementation work. - Businesses exploring AI in finance are likely to focus first on review support, document chasing and categorization before moving to higher-risk automation.