Deployed but unused UiPath models keep consuming AI Units. Why this happens and how to avoid the cost trap in AI Center.
Short answer: yes, it can. A document model deployed as an AI Center ML Skill keeps consuming AI Units at UiPath as long as it is running in "Available" or "Updating" status, regardless of whether any automation ever calls it. Anyone who simply leaves a model deployed after testing pays for pure idle time. That is exactly what a user recently asked in the UiPath Community Forum.
The reason lies in billing by uptime, not by usage. As soon as an ML Skill is deployed and available in AI Center, UiPath reserves compute power for it and bills it by the hour. According to the UiPath documentation a CPU deployment costs around 2 AI Units per replica and hour, a GPU deployment around 20 AI Units per replica and hour. The standard GPU configuration reaches about 40 AI Units per hour, which works out to roughly 30,000 AI Units a month for continuous operation. Billing applies for every hour started, in which the model is "Available" or "Updating", regardless of the number of documents processed.
It's important to distinguish this from the modern, managed Document Understanding models: these cost nothing while idle, since AI Units are only incurred per page processed. The cost trap applies to the classic, self-trained models that run permanently in AI Center as an ML Skill. Many businesses calculate the UiPath license costs for SMEs, but overlook the ongoing AI Units from such deployments.
The most effective measure is mundane: stop or remove a model that isn't currently needed. UiPath explicitly recommends stopping or fully undeploying ML Skills when not in use, to save on hardware costs. Three steps help in everyday practice:
In our automation projects, this is exactly where we start: a document model that, for example, processes invoices automatically, is only deployed when a process actually requests it, and shut down again afterwards. This noticeably lowers ongoing AI Units costs without the automation losing any reliability.
Yes. A model deployed as an ML Skill is billed hourly as long as it is "Available" or "Updating", regardless of the number of documents processed. Only "Stop" or "Undeploy" ends the billing.
A CPU deployment costs around 2 AI Units per replica and hour, a GPU deployment around 20. The standard GPU configuration adds up to roughly 30,000 AI Units a month with continuous operation.
No. For the managed modern models, AI Units are only incurred per page processed, so idle time costs nothing. The trap applies to classic, self-hosted ML Skills.
NordFlux builds digital employees for organisations: automations and AI agents that take over repetitive work. You stay in control.
Agent Units, AI Units, and Platform Units: what UiPath actually charges for agentic automation in 2026, and where the budget tips over.
One process, one internal validation bot: which UiPath license makes sense, why the Community Edition is ruled out, and when attended is enough.
Typing out incoming invoices eats up hours every week. How Document AI reads, checks and posts your documents, and what matters when you put it into practice.
A deployed document model keeps consuming AI Units in AI Center even when nobody works with it anymore, and the cost trap can stay unnoticed for a long time. NordFlux audits your AI Center for exactly this kind of hidden cost and sets up monitoring that flags it immediately from now on.
