Exclusive: UiPath's Feiran Hao explains why humans still have the final say in agentic AI

When it comes to orchestrating automation across agents, robots and people, many organisations are struggling to connect the dots.

TechDay sat down with Feiran Hao at the UiPath Agentic Automation Summit in Sydney to explore both the challenges and opportunities in this evolving space.

"The mistake they're making is stitching technologies together loosely - if at all," said Feiran Hao, Vice President of Product Strategy at UiPath.

"They've got a robot checking if there's a new email and another process sending that email, but the connection between the two might not even exist."

Hao believes that while businesses are eager to harness emerging agentic tools, the reality is that many are developing them in silos. "They're prototyping in isolation. So when something breaks, they often don't know where it went wrong," he said.

UiPath's solution aims to bring visibility and traceability to these disjointed processes.

"We look at the whole end-to-end process as one big process - not sub-steps," Hao explained.

"Our logging will say: it failed here because of this reason. That helps them spot root causes much faster."

That kind of clarity is critical in fast-moving business environments where adaptation and reactivity are key. But even as companies embrace new tools, Hao noted that few are ready to hand over decision-making to AI.

"My experience is that businesses don't want agents to make decisions yet. They want recommendations - decisions are still human," he said. "They want AI to go and tell them what to think, but not actually do the thinking."

He used the example of loan processing to illustrate the point. "You don't want AI making the call on a loan application. You want it to gather data - credit score, documents, employment history - then present a recommendation to the human, who can weigh all of that plus any extra context the AI might miss."

That approach balances efficiency with responsibility - especially important in sensitive or regulated industries.

"At the end of the day, in healthcare eligibility, for example, a human still has to look at everything and hit OK," he said.

UiPath is investing heavily in reliability and version control for its agentic tools, Hao added, citing features like traceability and evaluation sets.

"You create a bunch of scenarios and see how the agent behaves across them. If it's being too strict, you can tell it in natural language, 'Be less strict,' and our autopilot will suggest how to adjust the prompt," he said.

That kind of feedback loop is embedded early in the agent creation phase - long before the tool is deployed live. "Because once the agent is already running, it's too late," he added.

Still, for organisations fearful of trial and error, Hao advises caution paired with ambition. "Dream big, but start small. Find little pockets of work you're not afraid to experiment with."

Many businesses, he pointed out, are still playing catch-up with basic automation. "They want to say they're on the cutting edge, but a lot of it is just core automation they haven't gotten around to yet," Hao said.

The path to full autonomy, where AI does everything from instruction to execution, is still a long one - though UiPath is already laying the groundwork.

"We're demonstrating early capabilities," Hao said. "But in high-end enterprise, people aren't ready yet. It's more of a glimpse of what's coming than what they're doing today."

That doesn't mean roles like traditional RPA developers will disappear, Hao noted - they'll simply evolve. "Every technology shift changes what we do. RPA developers will upskill into agentic automation. They're already using the same design interface."

On the interface front, UiPath sees promise in chat - but not just any chat.

"Chat-based tools out there now have a hard time integrating conversation with real action," Hao said.

"Our conversational AI is built on top of automations that already exist. It knows what it can actually go and do."

For example, while a general chatbot might struggle to process a request like submitting an out-of-office form, UiPath's system can execute that task - provided the automation has been built. "If our platform already has an out-of-office automation, then yes, our conversational agent can trigger that."

That orchestration - between data, automation and agents - is also unlocking value in areas like supply chain management.

"One example is pricing optimisation," Hao said.

"Agents can search the web for competitor pricing, automation streams that into our AI model, and the model makes a recommendation on what to quote or how to price. Then the human can act on it, and automation handles the execution."

When asked for a bold prediction about the future of work, Hao's answer was grounded.

"Even if we just get to what we're describing today - small pockets of agent recommendation - it'll already be a massive step," he said.

Despite the hype, he added, many enterprises still need to complete the basics. "The future of agentic AI is exciting, but right now, we're just building the foundations."

And where does he see it all heading?

"One day, you'll say: 'Set my out of office,' and the agent will just know what to do, across all systems - even if that automation hasn't been built yet," Hao said. "But we're not quite there today."