Yamada-san (not his real name) dreaded the 25th of every month.
He runs a specialty trading company in Hyogo Prefecture — about 30 employees. For him, the last week of the month meant one thing: invoice hell. His two-person accounting team would spend three solid days doing nothing but processing invoices. Everything else ground to a halt. Overtime was a given because it collided with month-end closing. During busy seasons, they sometimes worked past midnight.
"Every month, the same thing over and over," he told me during our first meeting. "Open an invoice, read the numbers, type them into the accounting software. If you make a mistake, fix it and check again. I kept thinking — is this really something humans should be doing?"
I found myself nodding. I'd heard this exact frustration from small business owners more times than I could count.
The never-ending manual labor called "invoice processing"
When we took a closer look at Yamada-san's operation, the scale of the problem became clear.
His company receives about 150 invoices every month. The real issue? They arrive in every format imaginable:
- PDF attachments via email: ~60%
- Paper invoices by postal mail: ~20%
- Faxed invoices: ~10%
- Excel files: ~10%
Every vendor has a different layout. Company A uses a landscape format. Company B splits line items across two pages. Company C sometimes includes handwritten corrections. All of this had to be read by human eyes and manually typed into accounting software. Every single month.
And naturally, mistakes happened.
Mistyped digits. Incorrect tax rates. Wrong vendor codes. Yamada-san's team was seeing roughly a 5% error rate on monthly invoice processing — about 7 to 8 errors out of 150 invoices. That might sound small, but finding and fixing those errors ate up even more time. Worse, payment errors that slipped through damaged trust with business partners.
"There was this one time we transferred the wrong amount because of a data entry mistake. When the vendor called, I was drenched in cold sweat."
The total time Yamada-san's accounting team spent on invoice processing? Over 20 hours every month. Annualized, that's 240 hours — the equivalent of one full-time employee spending a month and a half doing nothing but data entry.
What we built: an AI-OCR invoice processing pipeline
"Automating invoice processing with AI" can sound like a big, intimidating concept. Massive systems, huge budgets, months of implementation — that's the image many people have.
What we actually built was surprisingly straightforward.
Here's the simplest way to think about it:
The work a human was doing — looking at an invoice, reading the content, and entering it into accounting software — is now done by AI instead.
In slightly more detail, the system works in three steps.
Step 1: Intake. Whether an invoice arrives by email, scanner, or fax, it gets pulled into the system. PDFs are ingested directly. Paper and fax invoices get scanned to digital format. This part is semi-automated through email forwarding rules and scanner configuration.
Step 2: AI-OCR reads and extracts data. This is where the AI does its work. OCR (Optical Character Recognition) technology reads the invoice content, and the AI identifies and extracts the relevant fields: "this is the invoice date," "this is the vendor name," "this is the total amount," "this is the consumption tax." The big difference from traditional OCR is that AI understands context. It handles handwritten corrections and messy layouts. When line items span two pages, or when discount information is buried in the remarks field — the things a human would process intuitively — the AI picks those up too.
Step 3: Sync with accounting software. The extracted data flows into whatever accounting software the company already uses. But — and this is important — it's not fully unattended. The AI's extracted data appears on the accounting team's screen as "pending review." The staff member sees the AI's reading alongside the original invoice, compares them, and either approves with one click or makes corrections on the spot.
The key shift: "entering data from scratch" becomes "reviewing data the AI prepared." That single change transforms the entire workload.
The implementation process: faster and lighter than you'd expect
"Can a small company like ours really do this?"
That was Yamada-san's first question. He assumed this was enterprise-only technology.
Here's how the actual implementation went.
Week 1: Assessment. We spent the first week mapping his current invoice processing workflow together. What vendors send invoices, in what formats, and how many? What accounting software does the company use? What does the accounting team's day-to-day actually look like? At this stage, we barely talked about technology. The conversations centered on "Where do you spend the most time?" and "Where do mistakes tend to happen?"
Weeks 2-3: Prototype and testing. Using real invoice samples from Yamada-san's company, we tested the AI-OCR's reading accuracy. It wasn't perfect from day one. Certain vendor formats weren't parsed correctly. Handwritten amendment recognition was weak initially. But with iterative tuning on sample data, we reached production-grade accuracy within two weeks.
Week 4: Accounting software integration. We built the data bridge to Yamada-san's existing accounting software using its CSV import functionality. The goal was to integrate without disrupting the existing workflow. No new software to learn — the same familiar screens, just with data already filled in.
Week 5: Staff training and go-live. We trained the two accounting team members using real invoices. The actual workflow is simple: pull invoice into system, review AI's reading, approve or correct. Both team members had the basics down after a one-hour training session.
Total timeline: about five weeks. This wasn't a massive system overhaul. It was more like inserting "AI eyes" into an existing business process.
Results: the numbers speak for themselves
Three months after go-live, here's what we measured.
Invoice processing time: from 20+ hours/month to under 3 hours/month. Per-invoice processing dropped from an average of 8 minutes (manual entry) to about 1 minute (review and approve). Month-end overtime for the accounting team dropped to nearly zero.
Error rate: from ~5% to near 0%. The AI's reading accuracy exceeded 97%. The remaining 3% were caught during the human review step, so errors no longer made it out the door. In the three months since launch, the number of payment mistakes reported by vendors: zero.
ROI: payback in 3 months. When we added up the savings — reduced overtime, eliminated error correction costs, and recovered productivity — the implementation cost was fully recouped within three months. From month four onward, it was pure upside.
The numbers alone tell a compelling story. But what Yamada-san seemed most excited to share was something else entirely.
What surprised us: the changes that don't show up in spreadsheets
"Honestly, the biggest change was the look on my accounting team's faces."
That caught me off guard.
With "invoice hell" behind them, the two accounting staff found themselves with time for work they'd never been able to touch before. One started analyzing payment cycles by vendor and put together a cash flow improvement proposal. The other began tracking expense trends in Excel and started presenting cost reduction ideas at monthly meetings.
"Before, they were 'data entry staff.' Now they've become 'analysts.' They both tell me work has actually become enjoyable."
This was a change we hadn't predicted. Automation often triggers fears about "replacing people," but what actually happened was the opposite. Freed from repetitive manual work, people gravitated toward higher-value tasks that made better use of their skills.
Another comment from the accounting team stuck with me: "I'm not afraid of month-end anymore." The dread that used to build every time the 25th approached was simply gone. That psychological relief translated into better work quality and a noticeably better team atmosphere. Hard to quantify, but unmistakably real.
Invoice automation isn't just for big companies anymore
A few years ago, AI-powered invoice processing was firmly in enterprise territory. Implementations cost millions of yen. Monthly operating costs ran into the hundreds of thousands. It was genuinely out of reach for small and mid-sized businesses.
That has changed dramatically in the past two to three years. Cloud services have matured. Open-source OCR models have gotten remarkably good. API-based architecture makes flexible, modular builds possible. The result: for a company processing around 150 invoices a month, automation is now realistic at SMB-friendly budgets.
You don't need a multimillion-yen system. What you need is a partner who understands your operations, integrates with your existing accounting software, and builds something your team can actually use.
At SolidTech, we approach every project the way we did with Yamada-san's company — by looking at the painful parts of your workflow first. It might be invoice processing. It might be something else entirely. The important thing is starting from the problem, not from the technology.
If you've been thinking "we spend way too many days on invoices every month" or "there has to be a better way to handle month-end accounting" — let's have a 30-minute conversation about it. We'll give you an honest assessment of whether AI is a good fit for your situation. Not a sales pitch — just a straightforward discussion between people who've seen this problem before and business owners living with it every day.
