Scott Fulton, Chief Product & Technology Officer at BlueCat.getty​Every few years, our industry sells the same promise in new packaging. Client/server was going to retire the mainframe and its operational overhead. Managed runtimes were going to end "DLL hell" and make deploying software as simple as copying a file. The cloud was going to make data centers, and the people who ran them, someone else’s problem. Microservices would replace big, complicated applications with small, simple ones. The serverless framework even gave the dream a name: NoOps.Every one of those shifts was real. And every one delivered the opposite of less operations. Each layer we added to make things simpler became another layer we had to observe, control and understand. The wheel turns, and every turn creates more need for insight into what’s happening.Agentic AI is the newest spin of that wheel. It’s also the most seductive because it promises that software will decide for itself. That is why the way we inject agentic AI into product road maps, marketing and sales pitches is a problem. Gartner calls it “agent washing” and estimates that of the thousands of vendors marketing agentic AI, only about 130 are the real thing.I’ve spent my career building operations software meant to run on its own and taking apart the architecture behind other people’s autonomy claims in due diligence. The gap between what agentic means in a demo and what it means in production remains wide, keeping buyers from taking on the risk or allocating meaningful budget.Why Reasoning Didn't Change The CategorySomething changed this past year. Reasoning models made automation more capable. That progress is easy to mistake for autonomy.Reasoning is a capability. Agency is a control structure. Anthropic has the clearest distinction I’ve seen. You have a workflow and smart automation if a model makes its decisions inside predefined code paths. It’s not an agent until the model directs its own process, choosing what to do next based on feedback from its environment. Reasoning improves the quality of each step. Agency is about who decides the steps. A reasoning layer bolted onto a fixed script is just a better tool.Why Blurring The Line Is ExpensiveMost agentic AI evaluations happen in demos. They are controlled, scripted and built to succeed. Production is none of those things. When Carnegie Mellon tested leading models on real multistep office work, the best models completed about 30% of tasks end to end. And reliability erodes with length. For example, chain 10 steps at 95% reliability each, and you finish below 60% overall. Autonomy that dazzles in a five-minute demo often crumbles across a 20-step process running on your real data.That is why Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. Too many vendors bought into the promise of agentic AI instead of the capability.​How To Separate Agency From AutomationLook at the architecture. Ask your vendors a few questions from the list below and you’ll have enough clues to deduce whether your agentic AI demo is substance or fluff:• What existed before they labeled it AI? If the engine predates the branding, you’re likely looking at a layer, not a leap.• Does it reason, or does it act? Many systems stop at advice. Agency means executing toward a goal and owning the path to get there.• Does it remember, adapt and coordinate? A genuine agent carries context across steps, adjusts when conditions change and orchestrates several tools toward a goal. A tool that starts fresh at every prompt, or fires one function on command, doesn’t clear that bar.• What is it grounded in? An agent is only as good as the record of reality it works from. Most enterprise data is stale, scattered across tools and incomplete. An agent acting confidently on top of that is a liability, not an asset.• What happens when it’s uncertain or wrong? Genuine agents detect failure, recover or hand control back to a human. A demo will never show you this.• Where does the human stand? Credible systems keep a person in the loop for high-impact actions by design, not as an afterthought.Not every system needs all of these. But agency is more than reasoning alone, and no amount of confident prose substitutes for the answers.How Autonomy Without Scrutiny Becomes Your LiabilityThere’s another reason overusing the term agentic AI should matter. You own what your systems do. When an airline’s chatbot invented a refund policy, a tribunal held the airline liable—not the chatbot, and not the vendor. Grant a system more autonomy than it has earned, and you don’t get efficiency; you get the wrong actions, executed with absolute confidence at machine speed.The mature path was never full autonomy or nothing. Systems should earn autonomy in stages, moving from advice to human-assisted action and, eventually, to automation with clear limits. Even then, the model should make recommendations while deterministic, auditable controls decide what happens next.Vendors should be able to show you where their product sits on that path, and what happens when it gets something wrong. The credible ones will have clear answers.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?