The moment a single invoice lands on the desk
Last month, I spent about two hours talking with the head of accounting at a manufacturing company. On one side of his desk sat a stack of paper invoices; next to it, a monitor with the Korean National Tax Service (Hometax) window open. "I'm going through these one by one. Whether the amount is right, whether the VAT classification is right, whether the vendor code is right." The invoices he pointed to with his mouse were less than half of what had come in that day.
If the previous post on the broken seams in logistics automation was about the gaps between pickup, tracking, and customs clearance, today is about the journey that single invoice takes through the accounting system — and how far along that journey we can hand things over to the machine.
Invoice processing — the first knot to come loose
Invoice processing was the first area where automation took hold. As OCR technology matured, the accuracy of machines extracting text — from paper invoices, PDFs, or email attachments — reached practical levels.
The numbers from overseas benchmarks are striking. According to one global accounting automation report, the cost of processing a single invoice dropped from about $16 to around $2, and processing time fell from an average of 10 days to roughly 3. "An 85% reduction" sounds abstract, but for a company processing 1,000 invoices a month, it means roughly 14 million won worth of labor and time can shift to other work.
The Korean environment, though, is a little different. Electronic tax invoices already dwarf paper ones, and from July 2025, corporations and sole proprietors with prior-year revenue above 80 million won are required to issue electronic tax invoices. So the more important question has become not OCR but "how do we automatically pull data from the National Tax Service system into our ERP?"
The current landscape of tax invoice automation
Domestic tax invoice automation solutions seem to fall into roughly three categories. There's the category that hooks into the National Tax Service in real time to pump issuance and receipt data straight into the ERP, the category that uses OCR to assist input from paper or PDF tax invoices, and lighter SaaS-style services aimed at freelancers and small businesses.
I heard about one mid-sized manufacturer. After deployment, finance costs dropped 20% and monthly reporting time shrank by 75%. When I first heard "75% of reporting time," it sounded a bit much, but as the story went on it became clear that "reporting" didn't just mean writing the closing report. For that one line the CEO expects when he asks "why did receivables go up this month?", the accounting team had been gathering, organizing, and reviewing materials — all of that time had been compressed.
That said, this kind of impact depends entirely on "what you defined as the target for automation." With the same tool, one company sees a 70% reduction and another sees almost no change. About 80% of that gap comes from the person designing which workflows to slice off and hand to the machine. Honestly, the tools themselves are starting to converge.
All the way to closing — where does the machine's job end?
This is where it gets genuinely tricky. Invoice entry is easy to automate; closing the books is not. Closing is less about "record-keeping" and more about "judgment" — and the share of judgment rises fast.
If we roughly carve up the work:
- Things machines do well: extracting invoice data via OCR, suggesting first-pass journal entries for vendors and accounts, automatic bank-account-to-ERP reconciliation, receivables matching, detecting duplicate entries for the same transaction, flagging anomalous transactions, generating routine month-end reports
- Things people still have to do: judging taxable vs. tax-exempt status, deciding accounting policies for new transaction types, year-end adjustments (inventory valuation, allowances, depreciation policy changes, etc.), responding to tax audits and external audits, and the holistic judgment that "this transaction is correct from an accounting standpoint but carries tax risk"
One accountant's comment stuck with me. "It's nice that the AI suggests journal entries, but for the first few months, verifying those suggestions actually took more time than before." Strictly speaking, the ROI of automation is lowest right after deployment. There's a stage where humans have to verify the machine's answers. Only after that stage does time actually start to fall.
AI-Human Loop — operating guide
So the setups that work well in the field are almost always similar. The AI produces a draft, a human reviews it, a human confirms it, and the confirmed result feeds back as training data for the machine — a loop. People call this the "AI-Human Loop," and in the end the success or failure of an automation rollout comes down to how cleanly this loop is designed.
A few principles that have helped in practice. First, design the UI so the machine's output is treated as a "suggestion" rather than a "decision." The simple act of a human clicking once to confirm clarifies where responsibility sits. Second, show a confidence score alongside each suggested journal entry, and route low-confidence items to humans first. Third, require humans to leave a one-line note explaining "why I changed it" whenever they correct a suggestion, and feed that into the next round of training. Those notes are what separate model accuracy six months later.
For reference, the 2026 corporate tax amendment includes a 25% tax credit for investments in AI and DX. If you're weighing the timing of a deployment, this is worth checking with your accounting firm or tax advisor.
Next up
Next time we'll look at clinic automation — how reservations, intake, and charts sit at the intersection of patient satisfaction and operational efficiency. If automating the "numerical records" of invoices and closings was relatively structured, clinic automation is the area where "human feelings" enter the picture most.
At 5years+, we've supported Korean and Japanese small and mid-sized companies in adopting AI and automation in the finance space. From invoice processing to closing support, the hardest decision has always been where to draw the automation line. If you're wondering "where should our company start," we'd be glad to sit down through 5years+'s automation services or a Free consultation and take a look at where you are. The decision to deploy comes later — usually it's more useful to first sketch out together where automation actually fits.