The Ugly Truth About AI Agents (and How to Spot the Real Ones)

About this video

David explains ‘agent washing’ — companies slapping an AI-agent label on old features — and says Gartner estimated earlier this year that only 130 of the thousands of agentic AI vendors actually have a real agent to offer. He defines an agent, citing ‘Artificial Intelligence: A Modern Approach,’ as something perceiving its environment through sensors and acting on it through actuators, and says major players including Amazon, Anthropic, Google, and Microsoft agree agents have five characteristics: they are autonomous, take observable action, perceive their environment, adapt from feedback, and are proactive. Testing ChatGPT Deep Research against that scorecard, he finds it autonomous and perceptive but not action-taking or proactive, and only somewhat adaptive, concluding what matters is knowing exactly which qualities a so-called agent has. As a fraud example, he says Presto Automation claimed to eliminate human order-taking at drive-thrus, but the SEC filed charges stating the vast majority of drive-thrus with its tech still required humans. Sponsor QuickBooks Online’s accounting agent passes all five tests: it auto-categorizes expenses, adapts when the user recategorizes an expense differently, and works proactively before the user logs in. Microsoft Copilot Studio agents score mostly yes, though adaptivity must be built yourself or via an agent flow with human input; custom GPTs are autonomous and perceptive, can take action with technical work, are somewhat adaptive, but are not proactive. David closes that even a real agent is no guarantee of success because the technology relies on chance and probability while businesses want reliability, and that you may not need an agent at all if the task needs no text or image interpretation or generation and no dynamic workflow.

Key takeaways

• David defines ‘agent washing’ and cites Gartner’s estimate that only 130 of the thousands of agentic AI vendors actually have a real agent to offer.
• The five agreed agent characteristics are: autonomous, takes observable action, perceives its environment, adapts from feedback, and is proactive.
• ChatGPT Deep Research passes autonomy and perception but takes no action and is not proactive, so David advises checking exactly which qualities any product called an agent actually has.
• Presto Automation is given as a fraud case: David says the SEC charged it after claiming to eliminate human order-taking while the vast majority of its drive-thrus still required humans.
• QuickBooks Online’s accounting agent passes all five tests, while Copilot Studio and custom GPTs fall short mainly on built-in adaptivity and proactivity respectively; David adds that reliability, not the agent label, determines usefulness.

Episode details: published October 20, 2025 · 9:09 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Kevin Stratvert’s own claims, shared for education — they are not typical or guaranteed.

← Previous: How to Use Copilot in Microsoft OneDrive · Next: Notion AI Agent — 7 Powerful Things It Can Do for You →

All Kevin Stratvert videos · Kevin Stratvert profile · Kevin Stratvert on YouTube