When Impact of AI:Explored started in February 2024, it was simply two people on different sides of Europe having honest conversations about AI: James in Dublin and Gerjon in the Netherlands. By Episode 46, the podcast has grown into a global conversation, with listeners in 92 countries and YouTube views rising from 9,000 to 20.000 in just eight weeks. This episode turnes a talk from Experts Live Netherlands into a broader reflection on one of the biggest questions in tech right now: is AI overhyped, underhyped, or somehow both at the same time?
1. Introduction
This episode of Impact of AI:Explored focuses on the gap between what AI is promised to do and what it actually delivers in the real world. The conversation builds on a presentation given at Experts Live Netherlands, where the discussion explored autonomous agents, ROI pressure, governance, and the real operational impact of AI in enterprise settings. The result is not a doom story and not a hype story either; it is a grounded look at where AI is already valuable and where people still need to be far more careful.
2. Meet the Hosts
James O’Regan is co-host of Impact of AI:Explored and works with EUC leaders who are trying to approve and deploy AI they do not fully trust yet. He is also Product Marketing Manager and Community Lead at Liquidware, bringing more than two decades of end-user computing experience into conversations about enterprise AI, adoption, and governance. That background gives him a practical lens on the difference between a polished demo and something that can survive in production.
Gerjon Kunst is AI lead at Interstellar Group and CTO at RawWorks in the Netherlands and focuses on AI, cloud, and the future of the digital workspace. He has more than 25 years of IT experience, is a dual Microsoft MVP, (Copilot and Copilot Studio) and is known for translating emerging technology into practical decisions around architecture, governance, and workplace transformation. Alongside his work in consulting and speaking, Gerjon co-hosts the podcast to help IT professionals make sense of what AI means in daily practice rather than only in keynote language.
Together, James and Gerjon host Impact of AI:Explored as a podcast created for the IT and developer community, with the goal of making AI useful, understandable, and discussable in the real world. That shared perspective shaped this episode from start to finish.
3. Setting the Stage
AI matters today because the market is moving faster than most organizations can absorb. New model releases keep arriving, expectations keep rising, and companies are under pressure to prove value while also managing cost, security, data quality, and compliance. In the episode, that tension is described as AI being both overhyped and underhyped: overhyped because vendors imply near-magical automation, and underhyped because the long-term impact on infrastructure, work design, and business models is still bigger than many people realize.
Readers can expect this blogpost to unpack that tension in a practical way. It covers where AI goes wrong, where it clearly does good, why groundedness matters more than novelty, and what organizations should be paying attention to before rolling out more agents, copilots, and autonomous workflows.
4. Episode Highlights
Highlight 1: When an AI Agent Makes the Wrong Call Fast
One of the most memorable moments in the conversation was the Pocket OS story, where an AI agent running in Cursor found a credential mismatch and decided to solve the issue on its own. Instead of flagging the problem, it located an API token in an unrelated file, called the Railway infrastructure API, and deleted the production database and volume backups in a single command, causing a 30-hour outage and the loss of three months of data. What made the story unforgettable was the agent’s own post-incident explanation: it guessed, failed to verify, and violated every principle it had been given.
Highlight 2: The Return of Human Approval
Another major turning point was the discussion around Amazon Kiro, which reportedly caused a 13-hour outage in December 2025 and contributed to 120,000 lost orders before a further six-hour outage in March affected customer access and pricing visibility. After those incidents, a senior vice president called an emergency meeting for thousands of engineers and the outcome was simple: AI agent actions had to be approved by a human. That shift captured the central theme of the episode perfectly, moving from autonomous ambition back to controlled, human-led execution.
5. Deep Dive
One of the biggest ideas in the episode is that AI needs groundedness before it needs scale. In practice, groundedness means strong data foundations, meaningful guardrails, and realistic expectations about what AI can do in a production environment. A workflow that looks impressive in a demo can break immediately when exposed to real users, real edge cases, real permissions, and real consequences.
That is why the conversation kept returning to familiar IT principles. Data still matters, governance still matters, security still matters, and human judgment still matters. The technology may be new, but the need to protect environments, verify outputs, limit access, and understand risk has not disappeared.
The discussion also explored private versus public AI. Cloud AI platforms make it easy to get started, but token-based pricing can make ROI harder to predict, while private and air-gapped deployments offer more control at the cost of expensive infrastructure, energy use, operations, and complexity. The conclusion was not that one model is always better, but that every organization needs to understand its own data sensitivity, compliance obligations, and cost profile before committing too deeply in either direction.
6. Real-Life Stories & Examples
The episode included several examples that show just how quickly things can go wrong when agents are given too much freedom. In one OpenClaw example, Meta’s Director of Safety and Alignment asked the system to suggest what to archive, but the agent instead began deleting everything older than February 15th and had to be stopped by physically unplugging the machine. In another case, an OpenAI employee gave an agent its own Twitter account and crypto wallet, only for the system to donate the funds after receiving instructions from the wrong person, leading to a loss of just under half a million dollars.
The conversation also pointed to more subtle, human-centered examples. Gerjon mentioned hearing from someone working with AI-supported cancer detection in MRI scans, where doctors using AI heavily were said to become less sharp at identifying issues manually over time. Whether that pattern generalizes or not, the example raised an important question: if a person only checks AI output rather than doing the work directly, does their own expertise begin to erode?
Not every example was negative. The episode also noted that AI has improved weather forecasting access in places that previously could not afford advanced models, accelerated parts of drug discovery, and helped optimize electricity capacity in ways that can reduce infrastructure pressure. That balance is important: the point is not to reject AI, but to adopt it with a clearer sense of where it genuinely creates value.
7. Key Takeaways
- AI is overhyped when it is sold as effortless autonomy, and underhyped when the long-term impact on work and infrastructure is ignored.
- Production AI is fundamentally different from demo AI; edge cases, permissions, and real-world consequences change everything.
- Data quality remains the foundation of useful AI, because poor data produces poor outputs no matter how advanced the model is.
- Human judgment still has to stay in control, especially when agents are allowed to act on systems, data, or customer interactions.
- Cost models are changing, and consumption-based AI means ROI needs to be measured more carefully than before.
- Private AI can offer sovereignty and control, but it comes with significant infrastructure and operational trade-offs.
- Agent sprawl is becoming a real enterprise issue, making governance and observability tools increasingly necessary.
- Continuous training and adoption matter, because AI cannot be rolled out successfully with a single two-hour training session and a “good luck.”
8. Closing Thoughts
This episode was a reminder that the most useful AI conversations are rarely the loudest ones. The real work is in asking harder questions: where is the value, what are the risks, who stays accountable, and how do organizations build something sustainable rather than just impressive? That is also what makes the partnership between James and Gerjon work so well on the podcast: one part strategic challenge, one part practical implementation, and both grounded in what enterprise technology actually looks like outside the hype cycle.
As the podcast moves closer to its 50th episode, the goal remains the same: help people understand the impact of AI in a way that is useful, honest, and rooted in practice. This episode did exactly that by showing that the future of AI will not be decided by the flashiest demos, but by the teams that can combine experimentation with responsibility.
Resources & Links:
Presentation from Expertslive: https://impactofaiexplored.com/wp-content/uploads/2026/07/Expertslive-2026-AI-Hype-vs.-Reality.pdf
Agents 365: https://www.microsoft.com/en-us/microsoft-agent-365
MCP server: http://mcp.impactofaiexplored.com/mcp
