How to Tell a Real AI Agent From an Agent-Washed One

October 21, 2026 · Josh Agusti, Founder, Aiven Consulting Group

Almost every vendor pitch now uses the word agent. Chatbots with a slightly longer context window get called agents. Workflow automation tools that chain together a handful of API calls get called agents.

Dashboards that summarize data on a schedule get called agents. The word has stretched so far that it has stopped doing useful work, and that is a real problem for the executives who have to decide what to actually buy, build, and govern this year.

What "Agent" Is Supposed to Mean

A genuine agent does three things that a scripted workflow does not. It reasons over an ambiguous goal rather than following a fixed sequence of steps. It takes multiple actions in a row, using tools, without a human approving each one along the way. And it adapts its next move based on what the previous action returned, rather than running the same path every time regardless of outcome. A monthly report generator checks none of those boxes, no matter how its vendor describes it. A system that can read an incoming contract, flag the clauses that deviate from standard terms, draft a redline, and route it to the right reviewer without a human re-prompting it at each step does.

Ask What Happens When the System Is Wrong

The more useful diagnostic question is not what the system can do when everything goes right. It is what happens when it is wrong, and who finds out. A real agent operating with any autonomy needs defined boundaries: what actions it can take without approval, what it must escalate, and how a bad decision gets caught before it compounds across a dozen downstream steps. Vendors rarely volunteer this detail unprompted, which is exactly why it belongs on the short list of questions a buying committee asks before a pilot, not after one. Enterprises that have already mapped their own acceptable-use policies and approval thresholds have a ready template to hold a vendor's answer against, instead of improvising the standard in the room.

The Gap Is Bigger Than Most Teams Assume

This is not a hypothetical concern. Sinequa's 2026 survey of 740 senior leaders at companies with $1 billion to $20 billion-plus in revenue found that 51 percent say they have AI agents in live production. Only 24 percent met the survey's stricter definition of a true agent, and just 10 percent had deployed systems with real autonomous, multi-step behavior. Eighty-four percent of the leaders surveyed said they had encountered an agent-washed product, one marketed as agentic that did not meet that bar, during a vendor evaluation in the past year. The distance between what gets claimed on a slide and what actually ships is not a rounding error. It is the median outcome.

Build the Internal Fluency Before You Buy

The fastest way to stop being agent-washed is to get specific internally about what you are actually trying to automate before you start taking vendor meetings. Teams that have worked through a concrete library of agentic AI use cases tend to ask sharper, more falsifiable questions in a demo, because they already know what a real version of the capability looks like for their own workflows.

That internal fluency is also what separates organizations that scale agent pilots from the ones that stall. IBM's research on enterprise readiness to scale AI agents found that a CEO mandate to scale is common, and the same pattern of ambition outrunning readiness shows up in its findings on the scale readiness gap. A clear internal definition of what counts as a real agent, built before the first vendor call, is the cheapest way to close that gap.

Key Takeaways

A real AI agent reasons over an ambiguous goal, takes multiple actions in a row without per-step approval, and adapts its next move based on what the previous action returned; scripted automation missing all three isn't an agent no matter how a vendor brands it.
Sinequa's 2026 survey of 740 senior leaders found that while 51% claim AI agents are in live production, only 24% met a stricter definition of a true agent and just 10% had deployed real autonomous, multi-step systems, with 84% reporting they'd encountered an agent-washed product during evaluation.
The most useful question to ask a vendor isn't what the system does when it works, but what happens when it's wrong and who catches it; mapping your own governance boundaries and use-case fluency before the first demo is the cheapest defense against agent-washing.