Case Study / Manufacturing AI
Factory simulation work just went
from 8.5 to 1.8 person-days.
Handing a factory equipment simulation model to the autonomous coding agent "Devin" cut development time by 78.8% for small-scale models at Kewpie, Japan's largest mayonnaise maker. Following Uber and Salesforce case studies abroad, a domestic company disclosing concrete person-day figures is unusual. Here's what actually made the difference.
The Bottleneck
Factory simulation used
to be a craft skill in code
Kewpie validates its factory equipment using Siemens' "Tecnomatix Plant Simulation," an industrial simulation tool that models conveyor speeds and equipment utilization in a virtual space. Its behavior is written in a dedicated scripting language, "SimTalk" — every robot arm and every conveyor belt has to be defined as code.
Until now, in-house software engineers hand-wrote SimTalk, ran it, and iterated on the discrepancies — a repetitive cycle that ate up significant time. Even a single small-scale model took an average of 8.5 person-days for design, implementation, and verification combined. It's not an unusual situation — a shortage of engineers who can handle a proprietary scripting language is a common bottleneck in manufacturing DX.
The trickier problem is that a proprietary industrial-software language like SimTalk doesn't have nearly the documentation or shared know-how of a general-purpose language. When the one engineer who can write it moves teams or leaves, that expertise tends to walk out the door with them — leaving behind a pile of "only-this-person-could-build-it" simulation models and a standing key-person risk.
How It Works
Two Devins, cloud and local,
split the job
Kewpie's setup pairs Cognition's autonomous coding agent Devin with itself, in a two-stage division of labor.
Why It Matters
Real, measured person-days
finally show up domestically
Devin's track record so far has mostly come from overseas case studies at companies like Uber and Salesforce, so a Japanese company disclosing work hours in actual person-days stands out. According to Innovatopia's report, the effort was supported by DeNA AI Link. This isn't a one-off demo — it's a number that came out of an actual production-equipment verification workflow, which makes it a useful reference point for other manufacturers.
Who benefits. Engineering firms short on simulation staff in production-technology departments get a directly usable playbook. If you control PoC budgets, though, discount the "78.8% on small models" framing a bit — larger, more complex models aren't guaranteed the same reduction rate. Teams that don't outsource software development get a concrete number to weigh against the cost of training in-house engineers. It's also worth noting how broadly applicable this is — any manufacturer running its own proprietary production equipment, from food and chemicals to auto parts, is a plausible fit.
What to do next. Three realistic first steps: (1) inventory your own proprietary-language assets — SimTalk, PLC, ladder logic, and the like; (2) prototype a single small model with Devin CLI and measure the actual person-days it takes; (3) decide whether the time saved gets reinvested into running more verification passes.
Counterpoint and risk. As ZDNET Japan's report also notes, a human engineer still has to sign off on the validity of the AI-generated SimTalk code — verification responsibility hasn't dropped to zero. The 78.8% figure is also specific to the small-scale models disclosed this time; it hasn't been validated for larger models or other equipment types. Variance in generated-code quality and dependency on the Plant Simulation version are both worth checking before a production rollout.
There's also a flip side to solving the key-person problem: the review process for AI-generated code needs to be redesigned from scratch. The old model was "the few people who can write it are also the ones who review it." As generation volume rises, review staffing itself could re-emerge as the new bottleneck. It's also worth discounting these numbers as a single case at a single company, not an industry-wide average — another manufacturer running the same setup could see a different reduction rate depending on the size of its existing simulation-model backlog and the skill gap among its own engineers.
The proprietary-language barrier
is no longer measured in headcount.