Decisions that cost time or revenue before they paid

Sequencing a domestic ERP by process readiness, not by deadline

In 2021 Ruoshan ran its Chinese and overseas companies on a European ERP suite installed in 2009, heavily customised and due to lose the vendor's mainstream maintenance for our version in 2027. Our largest shareholder, a provincial state-owned investment company, was encouraging its portfolio companies to move core systems to domestically developed software, and our vice-chairman, who represents it on the board, proposed a single domestic platform for every entity by the end of 2025.

Three options reached the executive committee: upgrading the foreign suite everywhere for about RMB 420 million; replacing it everywhere with domestic software by the end of 2025 for about RMB 510 million; or moving our Chinese entities and plants in an order set by readiness while overseas subsidiaries stayed on the foreign suite until at least 2027. I recommended the third. Our inventory had found 1,180 custom programs, 43 per cent of which had not run in two years, and several plants could not describe in writing how they issued material or recorded rework. Moving those processes to new software on a fixed date would have carried the confusion into it.

The vice-chairman thought the phased plan timid. Our CFO, Shulan Meng, wanted China finance delayed until after an H-share audit cycle. The board approved the phased plan at RMB 470 million, plus about RMB 36 million a year of dual running. Because I had worked at Zhiluo Software, one of the bidders, I declared it and left the commercial scoring; the selection committee chose Qiyao Software for the ERP core and Zhiluo's execution system for the plants.

China finance moved in January 2023 and six plants followed by December 2025. Close fell from eight working days to five, custom programs to 270, and China licence and support costs by RMB 41 million a year. The Huai'an structural plant's cutover in August 2024 failed: errors in converted bills of materials for 3,400 weld assemblies forced nine days of manual pick lists and delayed 310 excavators by up to two weeks. The end-2025 single-platform milestone was missed, and I said so to the board before it was asked.

I would rehearse every plant cutover with a full month of real orders rather than a sample. I would also set the date for the overseas decision at the start, instead of leaving it open for four years.

Keeping dealers' customer data in their own regions

When I became CIO in 2019, our overseas dealers ran 46 different dealer management systems, or none at all. Parts orders arrived as spreadsheets and warranty claims as paper forms with photographs sent through messaging apps, and a claim took 41 days on average to settle. First-pass parts fill from our hubs was 81 per cent. Dealers in Europe, the Gulf and Southeast Asia were losing service business to rivals with better parts availability, and we could not see why, because we could not see their stock.

I proposed a dealer portal for parts, machine registration, warranty and service history, with an optional hosted dealer management system for dealers that lacked one. The contested question was where it would run. Our head of international sales wanted one platform in China holding full end-customer records, so that Changzhou could follow up finance contracts and campaigns directly. Our security team preferred a single platform for control. I recommended two regional platforms, in Frankfurt and Singapore, with end-customer personal data kept in the region and Changzhou receiving machine, parts and claim data with customer identities pseudonymised. Dealers own the customer, their own countries' laws govern those records first, and moving them to China would have required consents their customers had never been asked for.

The cost was RMB 57 million more over five years than a single platform, and Changzhou gave up direct marketing to end customers abroad unless a dealer agreed. Nine large dealers refused the hosted system and connected through interfaces instead, which took a year longer to build.

By the end of 2025, 148 dealers selling 84 per cent of our overseas volume used the portal and 61 used the hosted system. Claims settled in 16 days, first-pass fill reached 92 per cent, and dealers' parts stock turned 3.4 times a year instead of 2.6. Twenty-three small dealers abandoned the hosted system within two years, because it demanded more training than their workshops could spare.

I would sign data-sharing agreements with dealers before building anything. I would also offer the hosted system only to dealers with enough service volume to justify the effort of learning it.

Stopping the Hefei plant for 58 hours after ransomware

In November 2023 our security monitoring flagged file encryption on two engineering workstations in the welding shop of our Hefei wheel-loader plant. A robot integrator had connected a laptop to commission a new welding cell, and the laptop carried ransomware. The workstations held programs for 38 robot controllers and had network paths to the line PLCs and the plant's execution-system server. As far as we could tell in the first hour, nothing on the controllers had changed.

The plant director, Zhaoming Kang, wanted to keep the lines running and rebuild the workstations one at a time, since the robots were still welding correctly. Our head of international sales reminded me that 180 loaders for a Gulf customer were due to ship before the quarter closed. I recommended isolating the plant's production network from the corporate network and from the other plants, stopping welding and assembly, rebuilding every engineering workstation and comparing every robot and PLC program with a known copy before restart. The president agreed. My reasoning was that we could not yet prove the controllers were clean, and an altered weld program on a loader frame is a safety question before it is a quality one.

The stop lasted 58 hours. We deferred 410 loaders and RMB 96 million of revenue into the next quarter, spent RMB 7.2 million on recovery and overtime, and shipped the Gulf order eleven days late. The comparison also exposed a weakness of our own: 31 of 412 robot programs had no current copy, and the integrator spent three weeks re-teaching them, with every affected weld checked by ultrasonic testing.

No other plant was affected and no controller program had been altered. Over 2024 and 2025 we segmented all eight plants by line at a cost of RMB 78 million, replaced permanent remote connections with brokered, recorded sessions for 52 vendors and integrators, and now verify 4,100 offline program copies every quarter. Plant directors report their operational-technology risk measures each month.

I would have made program backups a condition of every integrator contract years earlier. And I would have called the Gulf customer myself on the first morning, instead of letting the delay reach them through the dealer.

Restricting a service AI assistant to answers it could cite

Overseas dealer technicians called our Changzhou technical hotline about 8,600 times a month in 2023, in seven languages, often through interpreters and usually about fault codes, hydraulic settings or parts compatibility. Our head of international service proposed a generative AI assistant that would answer technicians directly in their own language, and his business case assumed calls would fall by 45 per cent.

We piloted it from March to August 2024 with 140 technicians at seven dealers. Senior service engineers reviewed 1,800 answers: 72 per cent were correct and complete, 19 per cent incomplete and 9 per cent wrong. Forty-one of the wrong answers stated a relief pressure, torque or flow setting that appeared in none of our documents. A technician who set a main relief valve to an invented pressure could damage a machine or injure someone standing near it.

I recommended restricting the assistant to retrieval. Every answer must quote the manual or service bulletin it came from; numeric settings may appear only verbatim from a cited document for the machine's model and serial range; anything else is routed to the hotline. The head of international service argued, fairly, that this would make the assistant far less useful and that technicians would give up on it. Under the restriction it could answer only 61 per cent of questions.

Most of the work turned out to be documentary. We converted 2,600 service bulletins and 380 manuals into structured documents tagged by model and serial range, and found 140 bulletins that contradicted one another. Twenty-one senior engineers now audit 500 answers a month.

In 2025 hotline calls fell 19 per cent, to about 7,000 a month, and first-time fix rates at participating dealers rose from 77 to 82 per cent. In 6,000 audited answers we found two wrong settings, both copied faithfully from errors in our own bulletins, which we corrected. The programme cost RMB 13 million. The 45 per cent ambition was never met, and the board heard that from me in the same paper that reported the gains.

I would have cleaned the documents before building the assistant, since the assistant mostly exposed their faults. I would also have given technicians a way to flag a bad answer from the first day.