The next enterprise AI failure may not be a hallucination or a rogue agent. It could be much simpler: a company spends millions on AI, celebrates hundreds of pilots and still cannot explain what changed in the business.
That tension ran through Charlie Perreau’s conversation with H CEO Gautier Cloix at HumanX Amsterdam. The session was called Leapfrogging from Chatbots to Autonomous Workflows, but Cloix’s argument was less about giving AI more autonomy and more about raising the standard for what companies expect once software starts taking action. HumanX framed the discussion around a future in which AI no longer waits inside a chat window for instructions, but begins initiating and completing workflows itself.
Perreau opened with the anxiety surrounding increasingly powerful AI systems. Was the fear justified, he asked, or had fear itself become a marketing strategy? Cloix’s answer deliberately left room for both. Some companies, he argued, benefit from presenting their models as extraordinarily powerful, but the underlying security challenge is real because AI can now interact with software systems humans have built over decades and potentially expose vulnerabilities inside them. His response was pragmatic rather than apocalyptic: “AI, it’s still software. You can still put guardrails and block it.”
That became more concrete when Perreau asked what happens when an agent gains access to APIs, databases or payment rails. Cloix described H’s work deploying AI into large organisations with legacy software and argued that computer use, where an agent operates through an existing graphical interface, can sometimes inherit useful constraints already built for human users. His broader rule was even simpler: if an employee has a certain permission, the AI can inherit it, but it should not automatically receive more.
The context matters because autonomy cannot carry the same risk everywhere. Missing the cheapest procurement option may be undesirable; making an error in healthcare can be unacceptable. Cloix’s point was that there is no universal level of autonomy. The workflow, consequence of failure and existing controls have to determine how much freedom an agent receives.
Perreau then moved to the question employees inside those companies are likely to ask first: Is this system here to replace me?
Cloix reframed the problem around the work people currently spend their time doing. He said H has found that employees frequently devote large portions of their day to tasks below their skill level: opening applications, logging in, copying information between windows, completing forms, requesting summaries and sending the result somewhere else. He described workers as becoming, in effect, “slaves of software.” His healthcare example made the argument tangible: automate the administrative screen work and a nurse can spend more time with patients instead.
But the sharpest exchange came when Perreau asked how large companies should escape the cycle of endless AI pilots.
Cloix’s answer was unequivocal: “Stop doing pilots.”
For the first phase of generative AI, he said, companies could reasonably measure adoption, experimentation or the number of use cases deployed. That period, in his view, is over. The metric now has to move toward economic results: What did the company spend, and what did it receive in return?
He pushed the distinction even further: “Efficiency, productivity is not an ROI.” If an AI vendor claims it can reduce procurement costs by 5%, the customer should be able to see that reduction rather than substitute a softer story about employees feeling more productive. Cloix even suggested companies should be willing to remove an AI provider after six or seven weeks if a promised result fails to materialise.
His closing advice to CTOs and AI leaders followed naturally: start with the systems you already have, choose a business problem with an observable outcome and shorten the feedback loop. Do not wait months to rebuild an entire software stack for APIs or commission an extended consulting exercise simply to identify possible AI use cases. Cloix offered a hypothetical bank taking 45 days to complete KYC: reducing that to four days creates a concrete target whose business impact can actually be measured.
That may be the real leap from chatbots to autonomous workflows.
A chatbot proves AI can answer.
An autonomous workflow has to prove something harder: that it can act within the right permissions, remove work that should not require a human and create a result the business can actually measure.
For enterprises entering the agentic era, Cloix’s message from HumanX was difficult to misread: stop counting how much AI you have deployed and start counting what it has changed.
About the HumanX Session
“Leapfrogging from Chatbots to Autonomous Workflows” took place at HumanX Amsterdam on September 24, 2026, from 10:25 to 10:40. HumanX lists Gautier Cloix, CEO of H, and Charlie Perreau, Head of Tech, Media and Start-ups at Les Echos, with the session focused on the workforce and organisational impact of AI systems capable of taking initiative rather than remaining static chatbots.
