Daniel Dines knows office automation like few others. The UiPath founder built a European software giant selling software “robots” that replicate repetitive tasks. Today, in the midst of the generative AI wave, his voice calls for caution: do not cut jobs too fast.

The plea does not come from an idealist. Dines is a billionaire who made automation his business. Yet, as LLMs threaten to swallow tasks that until recently seemed safe, he admits to feeling anxiety himself. That unease mirrors a widespread nervousness in boardrooms and among engineers evaluating on-premise deployments of language models.

Why the human factor remains an unresolved trade-off

The tension is old, but new workloads make it tangible. Those who choose to run inference within their own data centers – often on GPU servers with hundreds of GB of VRAM – do so for data sovereignty or to contain TCO over time. But the speed at which an LLM agent can replace processes previously handled by people raises questions that no quantization or pipeline architecture can ignore. Dines does not talk about hardware, yet his message is equally technical: efficiency does not automatically equal the right decision.

The on-premise automation paradox

Self-hosted LLMs promise control: data stays in-house, latency is predictable, vendor lock-in is avoided. Yet, the very ability to fine-tune local models amplifies the potential for disintermediation. A company adopting an on-premise AI assistant can redesign entire workflows, often without having mapped the consequences for staff. This is the same scenario that made UiPath a giant, but now Dines urges not to rush.

Reading the automation anxiety between chips and contracts

For those watching the AI infrastructure market, the UiPath founder’s confession signals that even the most seasoned vendors recognize the limits of pure substitution. Accelerator cards, data prep pipelines, inference libraries – the whole on-premise ecosystem – delivers performance measurable in tokens per second, but cannot measure the impact on corporate culture. AI-RADAR constantly explores these trade-offs, without prescribing solutions, but offering frameworks to evaluate LLM deployment in real-world contexts.

Ultimately, Dines’ caution is a reminder: intelligent automation deserves to be deployed with the same care you would put into designing a GPU cluster. Maybe even more.