Agentic marketing without agent washing: how to test whether a tool is a real agent
Five hands-on probes expose relabeled chatbots before you sign.

Agentic marketing is worth buying only when the tool can do the work its label promises. The way to distinguish a real agent from a relabeled chatbot is to test its actions, planning, logs, recovery and memory before signing. These five probes turn a vague claim about autonomy into evidence a technical marketing buyer can judge.
Why "agentic AI" became a label anyone can apply without proof
The category has no enforced definition. There is no certification body, required specification or equivalent of a UL sticker that a product must earn before calling itself an agent. A vendor can market an FAQ bot as an “autonomous support agent” without changing its architecture. What, then, gives the buyer grounds to challenge that description?
Gartner coined the term agent washing in June 2025 to describe the rebranding of AI assistants, chatbots and RPA tools without substantial agentic capabilities. The term names a pattern. Anushree Verma, Senior Director Analyst at Gartner, said current models “don't have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time”. Her criticism concerns the gap between propositions and capabilities.
Commercial incentives help explain that gap. The agent label can draw money and strategic buyers that a pitch for a chatbot or automation might not reach. A vendor may also describe a roadmap in the present tense, selling today’s copilot as tomorrow’s autonomous system. Without a shared standard, each vendor can draw the boundary just below its own product.
Gartner separately forecasts a sharp increase in enterprise applications with task-specific agents by the end of 2026, while estimating that fewer than one in twenty agentic propositions delivers what it claims. Whether the claim is a lie or a roadmap narrated in the present tense, the buyer deploys the product and inherits its shortfall.
What separates a real agent from a relabeled chatbot or RPA script
The useful distinction is who decides the next step. A chatbot answers questions or follows conversation trees with paths that a designer set ahead of time. RPA and workflow automation execute fixed sequences when triggered. A copilot suggests, drafts and surfaces options, but a human drives each action. An agent chooses its path at runtime in response to what it finds.
Anthropic’s Building Effective Agents draws the same boundary. Workflows run language models through predefined code paths, while agents dynamically direct their own process and tool use. Better prose does not change that distinction. A fluent response can sit on top of a fixed script, so model quality alone tells a buyer little about autonomy.
Four hallmarks make the distinction concrete. Perception means reading live state, such as a CRM record or another tool’s output. Reasoning and decision-making mean choosing the next action from the goal and context. Autonomous action means using APIs, databases, browsers or code execution to change something outside the conversation. A feedback loop then lets the system observe the result and adapt.
These capabilities do not require unlimited freedom. At a minimum, the system must plan and carry out routine steps while sending high-stakes actions through a human approval gate. Below that threshold, a copilot may still be the right purchase. The mismatch begins when it is priced and scoped as an agent expected to finish a multi-step task without someone driving every step.
The financial and operational cost of buying the label instead of the capability
That mismatch costs money when the tool takes on work the demo left out. By month three, the supposed agent may stall on the very task it was bought to do, completing a multi-step task without a human guiding it all the time. Someone must bridge the gap. The buyer pays for autonomy while staffing the supervision it expected to reduce.
The cost difference can also sit inside the architecture. Real agents can use smaller, cheaper models for routine steps and save frontier models for complex orchestration. Agent-washed products tend to use expensive models for all their work without cutting costs in that way. As the amount of work grows, real agents can need less human oversight, while rebranded automation needs more people to oversee it in proportion to the work. A demo that needs little human help does not prove the tool will cost little to run.
The FTC alleged that Air AI sold conversational AI as a tool that could replace human customer service representatives even though some customers never got access to it. The FTC also alleged that, when available, the technology was faulty and often could not perform basic functions. This is a vendor-specific case. The settlement included a substantial monetary judgment, largely suspended because the defendants could not pay, with only a nominal sum actually paid for consumer relief.
Buyers should test what a tool can do before they commit to using it. A contract cannot supply an architecture the product lacks. Nor does a smooth conversation prove that the system can finish the task it is being bought to do. The next step is to make that claim falsifiable through work the buyer can observe.
Five tests that reveal whether a tool is a genuine agent before you sign
The first test is API reachability. Ask the vendor to connect to an actual tool, such as your CRM, calendar, email or database, and complete an end-to-end action using your data. The product should update a record, create a quote or send a message. Drafting is not the same as acting. If a human must execute the output, the product fails this test of autonomous action.
To check how those connections work, ask whether the vendor supports new interoperability standards such as Anthropic’s Model Context Protocol, or MCP. Closed, proprietary integrations that cannot expose tools through standardized interfaces can signal rebranded automation. The decisive evidence is still the live action. A standards claim cannot substitute for a tool call that changes the intended system.
The second test is planning evidence. Ask the vendor to whiteboard how the product handles a request. Then give it a goal with two or three plausible execution paths. Can every path be specified before it runs? If so, the demonstration is showing a workflow. A real agent chooses a route from intermediate results and reroutes when an obstacle makes that necessary, while a washed product follows its script or fails.
The third test is tool-call logs and traceability. Immediately after a task finishes, ask for the action trace, including tool calls, decision points and a rationale for each step. A polished explanation without an auditable action record does not establish what the system actually did. Ask about the governance layer too, including how policy violations are monitored, because autonomous action needs boundaries that can be inspected.
The fourth test is exception handling. Ask the vendor to run the task with one input deliberately broken or one downstream tool unavailable. Failure is the point of this probe. A real agent detects the problem, chooses another path or escalates through a defined governance gate, then completes the task or records a structured exception. A washed product may fail without warning, loop or need a person to restart it, and it fails the test if it cannot handle and log the exception without someone stepping in to help.
The fifth test is state persistence. Interrupt a multi-step task midway, resume in a new session and ask the product to continue. It should resume from its last committed state. Starting over, losing prior context or asking a person to reconstruct the task exposes the missing capability.
Passing all five makes the agent label architecturally defensible, though it does not establish suitability for every job. The results also show which work still needs a human. A tool that passes some tests may be a useful copilot or workflow, with its scope and price set to match what it can do. Failure on API reachability or action logs is especially decisive, because a product that cannot act or show its actions offers little basis for trusting an autonomy claim.
How the same logic applies when the "agent" is running your content and brand presence in AI answers
People increasingly discover brands through direct answers from ChatGPT, Claude, Gemini and Perplexity. Brands must now look at how those engines describe and cite them as well as how their pages rank. A tool publishing at scale without planning, recovery or persistent context can destabilize the very presence it is meant to improve.
The engines cite largely different sources, so coverage of one platform can create a false sense of completeness. A system unable to adapt its strategy by engine cannot address that landscape. Sentiment also shifts more frequently than whether a brand is mentioned at all. Counting mentions alone can miss changes in how the brand is framed.
This makes editorial governance part of the capability test. A publishing pipeline without an action log gives reviewers no auditable route through a systematic framing error as it spreads across pages. Without human review, scale can amplify that error. Planning and state persistence are therefore not merely ways to produce more content. They help distinguish a campaign with retained editorial context from a succession of disconnected drafts.
For a marketing buyer, the same probes become concrete questions about the content pipeline. Can the tool call your CMS, analytics and brand guidelines database, or does it only draft? Can it explain a topic or angle choice rather than follow a fixed editorial calendar, and does it log each publishing action? What happens when a downstream system rejects a piece? When the campaign resumes, does the tool retain its editorial context?
The measurement layer completes the feedback loop. A brand-presence agent needs to detect whether the engines actually name and cite the brand before it can adapt to those results. Producing content is only one side of that cycle.
Running the checklist in practice: what to ask, what to watch, and when to walk away
To put the tests into practice, run them before signing, in the vendor’s demo environment, using your data or a reasonable facsimile. Start with API reachability, the fastest falsifier. If the product cannot take a real action, there is no need to continue evaluating it as an agent. If it can, proceed through planning, logs, a deliberately broken step and resumption across a session boundary.
Watch how the vendor responds to those conditions. A vendor with genuine capability will run the live tests, while an evasive response may involve a curated sandbox, a separate technical call or a pivot to another feature. That response is a signal. A request for time to “pull the logs” is a yellow flag, while an inability to produce them fails the traceability test outright.
Before signing, also require a documented stop mechanism and defined action limits. Asking how to stop the system tests whether the vendor has accounted for its ability to act. A clear answer should describe controls already built into the architecture.
Define success upfront through a measurable criterion tied to a specific task-completion threshold, and put it in the contract. Resistance to that criterion exposes a gap in accountability. Treat failure on real tool calls or action logs as disqualifying for an agent purchase.
The goal is a contract whose scope, price and required human oversight match the product’s actual level of autonomy. A well-understood copilot can serve a team better than a misrepresented agent. With these five tests, the buyer has a practical way to reject an unsupported marketing claim and make the purchase on evidence.


