Most software that uses a language model buys the tokens on your behalf and sells them back to you inside a subscription. That is convenient, and there is a markup, and the markup is not usually the interesting part.
Three things you get back
- Price transparency. You pay the model provider directly, at their published rate, and you can see exactly what the work cost rather than inferring it from a bundled line item.
- Model choice. When a better or cheaper model ships, you switch on your own schedule instead of waiting for a vendor roadmap.
- Continuity. If you stop paying us, your keys keep working. They were never ours. Everything you built on them, the enrichment, the classifications and the matching history, remains something you can run.
The part procurement cares about
There is a second argument that lands harder with security teams than with finance. If the model calls are made with your key, under your account, then your provider relationship governs them: your data processing agreement, your retention settings, your region.
That turns a long questionnaire about sub-processors into a much shorter conversation. Combined with an on-premises deployment, it removes the vendor from the data path entirely, which is usually the only answer that satisfies a review in a regulated industry.
What it costs you
Honesty requires naming the downside: you now hold a credential, and you have to rotate it, budget for it and notice when it expires. That is real operational work, and for a small team it may not be worth it.
For anyone with a security review to pass, or a catalogue large enough that model spend is a line somebody asks about, it is worth it several times over.