Decisions that cost us something, and what they bought

Ending our own CNC controller after three years and ¥3.4 billion

In April 2019, when I took over product development, our in-house CNC controller programme was in its third year. Our former president, by then chairman, had launched it in 2016 to reduce our dependence on outside controller makers and to set our machines apart in Europe. It employed 52 engineers and had spent ¥3.4 billion. The first customer shipment had slipped twice, from 2018 to late 2020, and our own field trials showed a controller that was competent but no better than the ones our customers already knew, trained their operators on and kept spares for.

The stakes went beyond money. Every new machine family was waiting on the controller's schedule, our development lead time had stretched to 34 months, and two more years would cost roughly ¥2.4 billion.

In September 2019 I recommended stopping the controller and keeping the parts of it that were genuinely ours: the operator interface, the thermal compensation and the machining cycles, rebuilt to run on the controllers customers already used. The programme's chief engineer opposed me in writing, arguing that we would never fully govern our machines' behaviour without our own controller. Our head of European sales wanted the distinction a proprietary controller offered. The chairman told me privately that he thought I was wrong. Our president took my paper and the dissent to the board, which agreed to stop.

We gave up the option of owning the controller, and the ¥3.4 billion was spent. Thirty-one engineers moved to the operator software, twelve to the remote-service unit and four to test automation; five left the company.

The operator software was released in October 2021 and now ships on 78% of our new machining centres. With paid remote-service contracts it earned ¥4.1 billion in the year to March 2026. Development lead time for a new machine family fell from 34 months to 27, because no project now waits on a controller release.

Every platform proposal now states which of its parts a customer could buy elsewhere. What I would do differently is the timing and the manner. My own test data pointed to this conclusion about a year before I acted, and I waited partly because the chairman had sponsored the programme. I also told the team by memo on the morning after the board meeting. Two of the five who left said that was why.

Listing every value our machines send home

Our remote service went live in April 2015 for new machining centres. In its first year only 11% of customers buying new machines agreed to connect them, against the 40% our service division had planned, and finance expected ¥600 million of subscription revenue by March 2018. Visiting customers, I heard what our brochures had not said. A contract machining shop feared its cycle times and utilisation might reach its own customers through us. An aerospace supplier was forbidden by its customer to connect anything. Most simply did not know what we would collect.

The stakes were the unit's survival and whether a machine maker could be trusted inside a customer's network.

In 2016 I proposed four rules. Every value a machine sends would be listed on its own screen. Owners could switch categories off. Part programmes and workpiece data would stay in the plant unless the customer chose otherwise. And we would not use utilisation data to plan sales calls without written consent. The general manager of our service division objected to the last rule, because utilisation was the best signal of when a customer would buy again. Finance objected to my second proposal: remote service free for three years with every new machine, with charges only afterwards.

We gave up the sales-planning use of the data and about two years of subscription revenue, and accepted a slower route to profit for a unit that grew from nine people to 96 by 2019.

The connection rate for new machines rose to 34% in the first year under the rules, 57% by 2018 and 74% in the year to March 2026. By 2019 first-visit fix rates had risen from 61% to 84%, as engineers arrived with the right parts, and average downtime per service call had fallen from 3.9 days to 1.6. Paid contracts began in 2019 and now cover 6,800 machines and ¥2.4 billion a year.

The data list is now part of every product specification, and our contracts describe retention and deletion in plain terms. What I would do differently is the on-premises gateway. We offered it only in 2018, three years after launch, and aerospace customers who could never connect outward waited that long. I should also have involved our legal team before the European launch, not after a customer asked for a data processing agreement.

Declining a software-limited export variant

In 2023 we were preparing a five-axis machining centre for aerospace and medical parts whose positioning accuracy placed it within the export-controlled category for machine tools. Shipping it to many destinations outside a small group of countries would need a licence under the Foreign Exchange and Foreign Trade Act, adding weeks to delivery. Our overseas sales head and the general manager of our Suzhou subsidiary proposed a variant whose accuracy would be capped by a controller parameter, so it could be sold without a licence yet share every part with the controlled model.

The stakes were commercial and legal. Competitors were quoting shorter deliveries, and a misjudged classification could expose the company to penalties and, worse, place a precise machine where it should not be.

Our export trade control office, led by Ayaka Fujinami, took the view that a cap any service engineer could lift by editing a parameter would not be a sound basis for treating the machine as uncontrolled. I agreed and proposed two alternatives. The first was a hardware-distinct variant, with different feedback devices and a separately qualified accuracy, for markets where licences took longest. The second was relocation detection, which stops a machine from running after it has been moved until our service centre confirms its new location, on every controlled machine shipped outside Japan. Sales argued that customers would read the lock as an accusation.

We gave up an estimated ¥5.1 billion of orders, lost or delayed while competitors delivered faster, and the variant took nine months and ¥380 million to develop. Between April 2024 and March 2026 we shipped 212 machines under licence. Relocation detection was triggered 38 times. Thirty-six were legitimate moves, re-authorised after confirmation; two were referred to the export trade control office, which handled them with the authorities under our compliance procedures.

Classification is now an agenda item at every concept review, and the export office has a vote at the design freeze. What I would do differently is the re-authorisation process. In its first months a legitimate move left a customer idle for an average of nine days while we confirmed the site, and two customers complained to their sales contacts before they told us. We redesigned it with customers, and it now averages three and a half days. We should have designed it with them before launch.

Keeping human hands on the final slideway while robots take the rough work

In 2022 our two Japanese plants had 44 technicians who hand-scrape the slideways and mating surfaces on which our machines' accuracy depends. Seventeen would turn sixty-five within five years, and our records showed that an apprentice needed about nine years before being trusted with final finishing on a precision slideway. We were taking two apprentices a year. Scraping had become the bottleneck for large horizontal machining-centre beds, which waited eleven days for it.

The stakes were both accuracy and output. Our manufacturing division proposed a ¥1.6 billion robotic scraping programme to automate the work on every bed by 2027, promising labour savings of about ¥420 million a year.

I recommended a narrower plan. Robots would take rough scraping on large beds, where the pattern is regular and the work is heavy. Final finishing would stay with people. And we would invest in measurement, imaging the contact pattern after each pass and scoring the spread and depth of the bearing points, so that apprentices could see what the masters saw. The head of manufacturing disagreed in writing, arguing that a half-automated process would keep the bottleneck. Several masters disagreed for the opposite reason: they felt the imaging reduced their judgement to a number. Tetsuo Miyabe, our most senior master, agreed to help design the scoring only if the masters could overrule it.

We gave up roughly two-thirds of the promised saving; the plan we chose cost ¥640 million and saves about ¥150 million a year. We raised apprentice intake to seven a year and introduced an instructor allowance for masters who teach.

By March 2026 robots carried 38% of scraping hours on large beds, and the wait for scraping had fallen from eleven days to seven. The first apprentices trained with measured feedback are on course for certification in final finishing after about five and a half years. Accuracy-related warranty claims have not risen.

The masters' scores now help us set the scraping requirements for each new design. What I would do differently is the beginning. One master retired early, telling me the imaging had been decided before anyone asked him, and he was right. I should have asked the masters to design the scoring before we bought the cameras, and the instructor allowance arrived six months later than it should have.