However smart, self-learning or labour-saving Agentic AI becomes, AI never stands on its own. The integration of human-in-the-loop verification and domain knowledge is essential for the effectiveness of AI agents.
That was the message at the Voka panel 'AI as the engine of innovation', where we shared the three most important lessons from our experience building Agentic AI solutions in production.
First: the potential of Agentic AI is enormous. In many cases it makes people up to ten times more productive. Every company that wants to optimise processes should be working on this today, not next year.
Second: AI only really works when you start from a clear bottleneck. Without a sharp problem there is no meaningful solution, let alone impact. The most successful projects we have delivered all began with a process that visibly hurt: too slow, too expensive, too error-prone. The technology came second.
Third: domain knowledge is the key to effectiveness. Without substantive context, an AI agent may know how to do something, but not what matters. An agent that drafts replies without knowing your policies, or plans interventions without understanding your operational constraints, produces output that looks plausible and lands wrong.
The pattern behind all three lessons is the same: Agentic AI rewards companies that know their own processes deeply, and punishes those hoping the technology will figure it out for them.
That is also the work: not the model, but everything around it.