[Case Study] 20,000 annual order checks: 3-person workflow reduced to 1 — Nabel Co., Ltd. × Leach Totsugo.com

Mr. Fujibayashi: I'm in the Sales Engineering Department, overseeing both Sales Engineering (field sales) and Sales Support (internal operations). Totsugo.com is used in Sales Support — the internal operations side.

The main work of Sales Support falls into four areas.

Mr. Fujibayashi: The team structure: Sales Support is 4 members in Mie and 2 in Yamaguchi — 6 total. Including Sales Engineering, the Sales Engineering Department has 6 in Mie and 3 in Yamaguchi in total. Annual order volume is around 20,000, with at least 20,000 confirmed for both last year and this year. Thanks to a V-shaped recovery in sales, volume is actually trending up.

Mr. Fujibayashi: Order channels are about 60% analog (FAX/email) and 40% EDI. EDI used to be around 20%, and we've pushed it above 40% by working with customers, but more than half is still analog.

For analog orders, someone manually keys the purchase order into our in-house ERP system. After entry, a printed order-check list is laid next to the customer's original purchase order, and we check delivery date, drawing number, quantity, unit price, customer order number, ship-to, and so on — one line at a time.

Mr. Fujibayashi: We run a two-stage check — primary and secondary, each performed visually by a different member. On peak days that's 100 pages of order-check list × 100 pages of purchase orders. Even on lighter days, 50 × 50.

Mr. Fujibayashi: We see three recurring kinds of mistakes.

Mr. Fujibayashi: Quantity and drawing-number errors have an especially large impact. If a tight-lead-time order is short on product, we have to re-ship within weeks, which throws manufacturing into chaos — and, more importantly, damages the company's credibility.

Volume-wise, on any given day's orders, the order check surfaces an average of 2–3 errors. But "we don't check assuming it's correct" is our operating principle — every cycle, we have to approach each line thinking "is this really right?" Sustaining that kind of attention for decades, while staying in good physical health, is simply not realistic.

Mr. Fujibayashi: In the past, we did run trials with OCR to digitize incoming FAX purchase orders. The OCR of that era couldn't reliably handle handwritten entries or smudged characters, so we didn't adopt it. The fact that our customers' purchase-order formats vary widely was also a barrier.

Still, the question of "can we use AI for this daily order check?" was something we kept raising with Mr. Tominaga.

Mr. Fujibayashi: We wanted to use data and AI to cut down on a task that happens every day and carries heavy mental weight. Even a paid custom build — we wanted it solved. The first quote Mr. Tominaga came back with was a custom-build quote around ¥4M.

But as the conversation went on, it became clear to us that prompt tuning and additional requests were going to keep coming. Rather than a one-time custom build, a monthly subscription with ongoing maintenance and updates felt like the better fit. That feedback led to the solution being delivered as Leach's SaaS product "Totsugo.com."

Mr. Fujibayashi: We supply to many industries in high-mix, low-volume production, so our daily product-registration volume is high and customization is a given. Because AI products evolve, the subscription model, where updates come in and additional requests are absorbed into the service, fits much better.

Mr. Nagai: When the front line said "even on a custom-build basis, please take it on," my first reaction was joy.

As Japan's working-age population shrinks and even new hires from nearby high schools are harder to come by, I'd been thinking: indirect work like purchase-order checking and data entry must be shifted to AI, so we can raise productivity and sales. Hearing the same view from the front line was genuinely welcome.

Adoption itself depends on ROI. If a pilot demonstrated value and contributed to productivity, I wanted to go ahead and adopt it. And I also hoped this system would become the catalyst that accelerates AI adoption across the company.

Mr. Fujibayashi: The current flow is straightforward.

Ms. Kitadera: At first, I wanted the AI to match every field — customer, ship-to, delivery date, customer contact, sales rep, drawing number, quantity, unit price, factory, manufacturing group, order number, notes, product memo… everything.

Mr. Fujibayashi: But in actual testing, the AI's output became inconsistent. Working with Mr. Tominaga, we narrowed things down step by step, and are currently running on this fixed prompt.

Mr. Fujibayashi: Remaining fields are covered in the secondary human check. The AI's accuracy is higher than I expected — grouping by order number and matching purchase-order numbers are working notably well.

Mr. Fujibayashi: What's interesting is that the AI can group by order number, even recognizing handwritten numbers our staff wrote on the purchase orders after entry. It matches those handwritten numbers with the printed order numbers on the check list, so each product is correctly tied to its slip.

Unit-price variations like "1 man (10,000)" vs. "10,000" vs. "10000" come back as a "confirm" status, which is easy to scan with the eye. Purchase-order numbers come back as "match" consistently, so we can confidently move on.

Mr. Fujibayashi: In concrete numbers, the 3-person daily check became effectively a 1-person workflow. Ms. Kitadera handles change and analog checks; the EDI side uses one more person. The previous secondary checker's 40 to 60 minutes per day are now fully freed up and redirected to other work.

Data from roughly the first month shows total time for primary + secondary checks has been cut by about half. There's still room: if we move to an operation where "items the AI flags as a match are listed, and humans skip those," we can shave off another layer.

Ms. Kitadera: Using it every day, just seeing the AI say "match" takes a real weight off. The pressure of chasing every single line with "am I missing something?" has clearly decreased.

Mr. Fujibayashi: From just one month of data, the savings are roughly at the level of our monthly fee. We aim to push well beyond that.

When I report to the company, I lead with the financials. But what genuinely helps me the most is that the "variation by individual" and the "worry about continuing this for years" have eased. Analog checking varies by how the person is feeling on a given day, but with the AI running the primary pass, work time levels out. And the worry about sustaining the same work daily for decades — that getting lighter is, behind the scenes, the biggest benefit.

Mr. Nagai: On headcount cost, the current monthly fee is already producing sufficient return, from our recognition.

Purchase-order confirmation and mapping to our in-house drawing numbers sits at the very top of our operational flow. An error there voids everything downstream — parts sourcing, manufacturing, inspection, packaging, shipping. That's why the people doing the checking were spending enormous time and nervous energy on it.

With this system in place, time and mental headroom opened up, and the front line itself started raising new "we want to AI-ify this process too" requests. Beyond cost benefit, that was the moment I felt we'd earned the real value of adoption.

Mr. Fujibayashi: We're already struggling to hire, so we have to get the work done with the people we already have. The worry of "can we really keep doing this?" was constantly there.

Meanwhile, sales are on a V-shaped recovery and order volume is climbing sharply. Being able to evaluate and adopt this at exactly this moment was a real stroke of good timing. The busier the operation gets, the more decisive the system's value becomes. Without it, we'd have had to expand the team, and that wasn't realistic.

Mr. Nagai: Without adoption, we'd have kept burning critical people on confirmation and data-entry work.

And in all likelihood, AI adoption wouldn't have progressed, and our team wouldn't have had its moment of realizing AI's importance and necessity. Company-wide AI adoption would have been further delayed.

Mr. Fujibayashi: Next, we want to bring Totsugo.com into our invoice-reconciliation work. Customers close the books at month-end and issue invoices at the start of the next month, so from around the 1st to the 15th we're flooded with delivery/receipt documents that have to be matched against our internal records.

The number of fields to verify is smaller than for order checks, so the AI's effect should be even more pronounced. We want to start with a trial.

Mr. Nagai: We first applied it to the Sales Engineering Department's purchase-order verification, but Procurement, Engineering, Production Management, Quality Assurance, and General Affairs all have substantial verification and data-entry work. The application space for Totsugo.com is still large.

We want the recognition of "AI is necessary here" to spread to every department, and for that to drive company-wide energy, efficiency, and revenue growth — that's our near-term direction.

Mr. Fujibayashi: I genuinely felt that, as long as the fields are defined, AI can handle the matching.

Almost no SME doesn't have some kind of document-check work. There's a staffer somewhere doing this, every day, treating it as just how things are. Don't give up — please raise it with Mr. Tominaga and tell him what you want. Every second saved on a task you do every day turns straight into profit.

Mr. Nagai: In high-mix, low-volume production like ours, the standard view was that automating matching was impossible — that manual work and multiple verifications were simply how it's done. Even so, a staffer held a gut-level concern about this status quo and raised their voice for change.

What matters is that even small automations get tried and experienced. A single successful automation spreads the idea through the department, and then past it to the rest of the company. For us, Totsugo.com was exactly the right first step.