Why ServiceNow Built Its Own LLM Instead of Just Renting OpenAI's

There's a decision every enterprise software company has had to make in the last two years, and watching how they answer it tells you everything about their strategy. The question is brutally simple: do you build your own model, or do you just call somebody else's?

ServiceNow built its own, NowLLM, and I think the reasoning is worth unpacking, because it's a case study every technical leader can steal from.

The default, lazy answer is "just call a frontier model." And for a chatbot demo, sure. But run that math at enterprise scale and the picture changes fast. When you're processing millions of workflow tasks a day, classify this ticket, extract these fields, route this case, you are not having deep philosophical conversations. You're doing high-volume, repetitive, narrow inference. And for that, a giant general-purpose model is like commuting to work in a Formula 1 car. Overpowered, overpriced, and oddly fragile for the job.

A domain-tuned model wins on the things that actually matter at scale. Cost, because you're not paying frontier-model rates for routine classification. Latency, because a smaller specialized model answers faster, and when it's in the critical path of a workflow, milliseconds compound. Data residency and governance, because a managed model inside the platform's trust boundary is a very different compliance conversation than shipping your enterprise data to a third party. And predictability, because a model you control doesn't change its behavior the week a vendor ships an update you didn't ask for.

But here's the part I respect most: they didn't slam the door. ServiceNow kept a generic LLM connector so you can bring your own model, a third-party frontier model or something custom, when the use case genuinely calls for it. That's the mature answer. NowLLM is the default because it's right for most workflow tasks, not because they're forcing it on you.

The transferable lesson is this. The build-vs-rent question isn't ideological. It's a function of volume and specificity. If your AI workload is narrow, repetitive, and high-volume, owning a tuned model usually wins. If it's open-ended and occasional, rent the big brain. Most serious enterprises will end up doing both, and the smart ones, like ServiceNow here, architect for exactly that from day one.