Toyota let factory staff build their own AI. They made 10,000 models

Toyota skipped a quote of over ¥100m and let plant staff build inspection AI themselves. By 2024 they made 10,000 models a year across 10 factories.

Toyota Motor Corporation let its own quality and production staff build the AI that inspects its parts, instead of buying it. The first three models replaced an outside quote of more than 100 million yen and were running within six months, according to Toyota's own publication, Toyota Times. By 2024 staff on the floor of Toyota's 10 car and unit factories were creating 10,000 machine-learning models a year on an in-house platform, with nearly 1,200 active users and over 10,000 man-hours a year saved, Toyota says. It has not published what the platform cost.

The idea doing the rounds this week is that AI makes the people closest to a problem want to build their own tools, and that locked-down systems start to feel like a cage. That is an argument usually made about phones. Toyota has been running the factory version of the experiment since 2018, and it has written down what happened.

How it started: one quote, two quality engineers

In 2018 a supervisor in the Quality Control Division at Toyota's Miyoshi Plant asked two staff, Masaya Furutani and Manabu Okuyama, to try something with AI. Toyota Times describes the brief as vague. Neither was a data scientist. Furutani had spent more than ten years in production engineering; both taught themselves machine learning and went on outside courses.

The first obstacle was not the technology. Company PCs would not allow a programming environment, and installing software needed approval. Okuyama recalls exchanging more than 100 emails just for approvals. In 2019 the pair persuaded the plant general manager to give them part of a meeting room with its own network line, separate from company systems, so that nothing they broke could touch production IT.

Their target was three AI models working on the line. An outside firm quoted more than 100 million yen. They built them instead, on cheap single-board computers and USB cameras, and hit the target of three on-site implementations within six months. Okuyama's point, as Toyota Times reports it, is the one most plant managers learn the hard way: building an inspection model on a PC is the easy part; wiring it into equipment that is already running is the job. They managed it because one of them knew the machines and the other could write the glue code.

The model is the cheap part. Fitting it to a line that is already running is the work.

What it grew into

Two things came out of that room, and they are worth separating because a smaller manufacturer can copy one without the other.

The first is D-ROOM, a staff community. It started as an open chat of about 500 members, moved to Teams channels in 2020, and now has 18 on-site hubs in plants including Miyoshi, Motomachi and Tahara. Volunteers run each hub; there were no dedicated staff. People borrow cameras, lenses and Raspberry Pi kits and post problems to the channel. Toyota Times cites one request for a development setup answered in nine minutes, and a connection problem that drew 79 replies. There is even a channel where people post their failures. In 2022 D-ROOM was moved into the Digital Transformation Promotion Division, which gave it a budget to buy and lend equipment.

The second is the AI Platform, which Toyota's Production Digital Transformation Office set out to build in 2022 so that factory staff could train models without AI expertise. Kohdai Gotoh of Toyota's AI Group wrote in a Google Cloud blog post that six developers built it in a year and a half. Staff upload images or sensor data through a web app; training takes 10 to 15 minutes at best and up to 10 hours at worst. It runs on-site most of the time and borrows cloud GPUs at peaks, so each plant does not need its own GPU servers.

What changed on the line

Toyota's figures, all from that Toyota-bylined post: the platform is in all 10 of its car and unit factories; models created rose from 8,000 in 2023 to 10,000 in 2024; nearly 1,200 people use it; and more than 400 employees go through the in-house training each year. At the Takaoka plant it inspects finished parts, checks the adhesive bead that holds the glass on rear doors, and flags abnormal behaviour in the injection moulding machines that make bumpers.

The time figure is stated two ways in the same post. Early on it says the platform would be able to save as many as 10,000 hours of work a year; later it says adoption has produced a reduction of over 10,000 man-hours a year in the manufacturing process. Treat it as Toyota's estimate, not an audited number.

What did it cost?

Toyota has not said. The only money in the record is the outside quote of more than 100 million yen for the first three models, which the in-house team avoided. The founders say they hoped to build each inspection rig for under 10,000 yen; Toyota Times reports that as an early ambition, not a measured unit cost. Six developers for a year and a half, cloud GPUs and a training programme for 400 people a year are real costs that are not priced anywhere.

What the numbers do not tell you

Ten thousand models a year is a measure of activity, not of value. Toyota does not say how many are in production, how many were retrained versions of the same model, or what happened to defect escapes. And 10,000 hours a year across ten factories is modest. Our arithmetic, not Toyota's: at roughly 1,900 working hours per person per year, that is the time of about five people, spread over ten factories. The bigger return Toyota describes is cultural: staff who now try things themselves, and a pipeline of people trained to do it.

Does this apply to a 200-person plant?

Partly. You do not need a platform, a cloud contract or 1,200 users. What transfers is the first phase: a quality engineer and a maintenance technician, a cheap camera and computer, a network segment kept apart from production IT, and one inspection job with a clear pass or fail. Toyota got three working systems in six months that way, without paying the outside quote.

What does not transfer is the scale. The platform made sense because Toyota had ten factories and a thousand people who wanted to train models. A single plant should stop at a handful of well-chosen inspections and maintain them properly.

What to do on Monday

Ask your quality lead for the one visual check that costs the most in operator time or escaped defects. Ask IT for a sandbox: a separate network line and a laptop with admin rights, isolated from production systems. Budget for a camera, a small computer and training time for two people. Write down the before numbers now — inspection minutes per shift, defects caught, defects escaped — so that in six months you can say what changed, which is more than Toyota has published.