Safe, reliable, production-ready AI agents don’t come from the model. They come from everything built around it.
Rasmus Hauch is CTO of Boost.ai, one of the enterprise platforms banks, insurers and public sector organisations use to build customer-facing AI agents. His work sits one layer below every deployment: the orchestration, the guardrails, the testing infrastructure and the model choices that decide whether an agent holds up in front of millions of customers.
In this episode, Rasmus opens that layer, explaining why conversational AI is shifting away from traditional conversation flows towards orchestration, testing and continuous evaluation. He walks through how Boost approaches it in practice: persona-based simulation across thousands of digital and voice conversations, predefined test suites for jailbreaking attempts and financial advice, layered guardrails inside and outside the agent, and a trust loop that feeds production results back into the build.
Rasmus is direct about the hard parts. We touch on why the big LLM providers can’t yet deliver the consistent response times voice demands, which is why Boost runs fine-tuned smaller models on its own infrastructure for latency-critical voice work. Rasmus explains why he advises clients against plugging in their own models.
We also discuss where liability sits under the EU AI Act when an agent gets it wrong, and why a volcano eruption in Iceland is a good example of what no model can be trained for.
Towards the end, Rasmus shares what’s exciting him right now, including adaptive voice, real-time language switching and where agent-to-agent protocols are heading.
Show notes
Discover more about boost.ai
Follow Rasmus on LinkedIn
Follow Kane on LinkedIn
Take our AI Maturity Assessment
Subscribe to VUX World
Subscribe to The AI Ultimatum Substack
