An operations director signs a contract for an “AI agent” to triage orders, convinced the company just bought autonomy: the system will sense the problem, decide what to do, and act on its own. Three months later, the real routine looks different — the “agent” suggests an action, a human approves every suggestion before it becomes execution, and any case outside the script stalls waiting for someone to decide for it. The company didn’t buy an autonomous agent. It bought a recommendation panel with a pricier name.
The phenomenon has a name, and it isn’t a skeptical analyst’s turn of phrase: Gartner calls it “agent washing” — vendors relabeling conventional, rule-based automation as “autonomous agentic AI” to ride the current hype. According to Gartner, out of thousands of vendors now marketing themselves as “agentic AI,” only about 130 offer genuinely agentic capability.
The data behind the label: 40% of projects won’t survive to 2027
In a survey of more than 3,400 organizations, Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027 — not because of a model’s technical limits, but because of a business decision made poorly before the purchase. According to Anushree Verma, senior director analyst at Gartner, “most agentic AI projects today are early-stage experiments or proofs of concept that are mostly hype-driven,” deployed without a clear strategy, without a real grasp of the complexity involved, and without governance capable of sustaining what the sales pitch promised.
This past May, Gartner repeated the warning in sharper terms, analyzing the supply chain planning market specifically: most of the “agentic” capability sold today improves the user experience — query interpretation, recommendations, conversational support — without actually changing decision quality or how the decision gets made. According to Jan Snoeckx, an analyst in Gartner’s supply chain practice, true autonomy would require automatic plan generation, automatic selection of the best plan, and execution without human intervention — a level most solutions haven’t reached yet, regardless of what the sales deck implies.
Automation suggests. An agent decides. The difference isn’t in the interface
That’s the line separating the two products currently fighting over the same market label. Automation with a conversational front end follows a predefined script: it recognizes a pattern, suggests an action from a closed menu of options, and stalls in front of any scenario the script didn’t anticipate — even in the “normal” cases, the final call still gets approved step by step by a human. A genuinely autonomous agent senses the state of its environment, decides the next step without waiting for approval at each stage, and — the point that actually separates the two categories — replans on its own when reality diverges from the plan, instead of freezing and handing the decision back to someone.
From the outside, both look identical: a clean screen, a natural-language response, an “approve” button. The difference only shows up the moment a scenario leaves the script — that’s where disguised automation stalls waiting for a human, and a real agent just keeps deciding.
Why most vendors fail the very test they claim to pass
The root cause isn’t the AI model behind the product — it’s what does or doesn’t exist before that model gets involved. As João Paulo Batistella, innovation executive and former CEO of EISA, argues, AI depends on process formalized as data, not on knowledge that only lives in someone’s head: a routine that was never written down as an explicit rule, trigger, and criterion gives an agent no ground to decide on at all. Hiring an “autonomous agent” to operate on a process the company itself never formalized is asking the vendor’s interface to solve a problem that’s actually the buyer’s homework — and that’s exactly where the promise of autonomy collapses back into “a suggestion a human approves,” because there’s no serious alternative once the basis for deciding was never defined in the first place.
The 3-question test before trusting the next “AI agent”
Before signing the next “agentic AI” contract — or treating something the company already uses as autonomous — three questions separate real autonomy from automation with a new name:
- What does it decide without my approving every step? If every action, even the routine ones, still needs human sign-off before becoming execution, the product is a recommendation panel — not an agent.
- What happens when the scenario leaves the expected path? A real agent replans on its own. Disguised automation stalls and hands the decision back to someone — usually without flagging that it stalled.
- What data does it decide from — is it formalized, or is it a polished prompt sitting on top of a messy spreadsheet? Without an explicit process and criteria behind it, no AI decides reliably, autonomous or not.
If all three answers hold up, the company is looking at a real agent. If any one of them stalls on “it depends” or “we’re still tuning that,” what got purchased is automation as usual — just wearing a name that promises more than it delivers, in a market where, per Gartner’s own research, that’s the rule rather than the exception.
Follow Eleva Tecnologia for more analysis on technology applied to business: follow @ElevaTechno on X or @elevatechnologies on Instagram, or learn about the group at elevatec.net/about.
