New Open AI Speaks Factory Jargon and Answers in 1.5 Seconds
Cuts sourcing answers from 7 seconds to 1.5 — and lifts sales 4%.
Industrial buying is messier than it looks. A factory worker types a shorthand nickname for a metal, a catalog lists a part with almost no detail, and an engineering standard somewhere says the exact specification that must be met. Get it wrong and a machine part can fail. That gap is why generic chatbots struggle in procurement — they can chat, but they can't be trusted with a part number. So a 16-person research team built IndustryLLM, a model designed specifically for that world, and released its weights for anyone to download.
Their twist: train the AI on its own failures instead of just feeding it more internet text. They assembled about 100 billion words — engineering standards, 10 billion words of de-identified real purchase and inquiry records, plus 60 billion words of ordinary text so it doesn't forget how to hold a conversation. Then they rewrote the messy material, translating between eight writing styles, fixing typos, and expanding vague codes. A buyer's '16674' becomes the real standard, GB/T 16674.
The cleverest part is what the AI refuses to do. It uses an 'evidence gate': every claim gets one of three answers — verified, not verified, or unknown. If the product data doesn't prove a spec is met, the AI stays silent rather than guessing. The payoff showed up in numbers: in live A/B tests against real users, sales rose 4.25%, satisfied inquiries rose 8.3%, and response time dropped from 6-7 seconds to about 1.5 seconds.
The catch: this is a 35-billion-parameter model, meaning it needs serious computing hardware — you won't run it on a phone. And its standards are mostly Chinese, so other countries' rules may not carry over. Still, the 'admit what you don't know' design is a template worth stealing for any high-stakes AI, from medicine to law.
- A new open AI model was trained specifically on industrial buying — the messy shorthand, part nicknames, and strict engineering rules that generic chatbots get wrong.
- In live production tests it answered 4-5 times faster (1.5 seconds vs 6-7) and increased sales 4.25%, with 8.3% more customer inquiries fully resolved.
- It's built to say 'unknown' instead of guessing when safety-critical specs aren't proven — a rare safety feature in AI, and the model is free to download.
Why It Matters
Faster, safer sourcing means quicker quotes, fewer costly ordering mistakes, and less time wasted in procurement.