Travelers, the insurance company, built its own artificial intelligence model instead of relying only on outside tools like ChatGPT or Claude. The model, called TravelersLLM, was trained on millions of the company's own documents and checked against tens of thousands of real insurance questions.
On insurance-specific work, the company says its home-built model gives better answers than general AI tools and costs far less to run. That second point is the real story here.
Travelers did not throw out the big commercial AI systems. When an employee's software sends a question, a background system decides where it goes. Simple, routine insurance questions go to the cheap in-house model. Harder or more open-ended problems still go to expensive frontier systems from outside vendors. Employees never see this decision happen, they just get an answer.
This matters because AI bills are becoming a real budget problem for large companies, even as the price of using AI keeps falling. Enterprise AI spending has climbed sharply over the past year, not because each question costs more, but because employees and software are asking so many more questions than before. Finance departments that once ignored AI spending are now tracking it the way they track cloud computing bills, because it has grown into a similar-sized cost.
Travelers is not alone in solving this with routing instead of restriction. Snowflake, AWS, and Oracle have all released tools this year that automatically send each task to the cheapest AI model that can still do the job well, rather than defaulting every question to the most powerful and expensive option. Research from AI chipmaker Nvidia has argued that a large share of routine business AI tasks do not need a top-tier model at all, a smaller, cheaper system handles them just fine.
The catch is that most companies cannot copy Travelers directly. Building a model that beats general AI on your own industry questions requires years of clean, organized company data and real subject matter experts feeding it. Travelers has spent more than a decade modernizing its data systems before this project became possible, and still runs some older systems it chose not to touch.
For most businesses, the realistic path is not building a model from scratch. It is asking your existing software vendors, cloud provider, or AI vendor whether they already offer automatic routing between cheap and expensive models. That single setting can cut AI costs without cutting AI use, and it is quickly becoming a standard feature rather than a luxury.
The bigger shift is this: the question companies ask about AI is changing. It used to be which chatbot to buy. Now it is how to manage AI spending as a permanent, growing line item, the same discipline finance teams learned to apply to cloud computing over the past decade.