The conversational software market is going through a relabeling wave. Platforms that until yesterday presented themselves as "chatbots with generative AI" now call themselves "autonomous AI Agents," often without having changed a single line of code. Gartner has given the phenomenon a name, Agent washing, and predicts that over 40% of Agentic AI projects will be canceled by the end of 2027, driven by rising costs, uncertain business value, and weak risk controls. According to the same analysis's estimates, of the thousands of vendors calling themselves 'agentic,' only around 130 offer real agentic capabilities.
The mechanism fueling Agent washing is simple. Frontier language models are getting more and more alike and, on a generic task, the perceived difference between their answers has narrowed. Any product built on top of these models therefore looks intelligent in a fifteen-minute demo. The difference between a real AI Agent and a well-built interface doesn't show there. It shows later, in production, when the cost of the mistake is not a disappointing demo but a project that ends without measurable results, after months of implementation and a budget invested in the wrong promise.
This article gathers five concrete criteria to recognize the difference as early as the selection phase.
What Agent washing is
Agent washing is the practice of presenting as an AI Agent a system that, in essence, remains a chatbot. It understands natural language and generates relevant answers, but it doesn't act on the client's systems, doesn't maintain the state of a request across multiple steps, and doesn't change its behavior based on the corrections it receives. It is conversational, not agentic. Two different qualities, often used as synonyms.
This is not widespread bad faith, but a market incentive. Relabeling an existing chatbot takes a week of work on the website and the sales materials. Building the capabilities of a real AI Agent - multi-step orchestration, write integrations into the client's systems, a tracked learning loop - takes months of engineering and an architecture designed for auditability, not just for conversation. Those who invested in the second path have every interest in making the difference visible. Those who stopped at the first have every interest in hiding it behind shared vocabulary. Hence, the need for recognition criteria that don't depend on the vendor's vocabulary.
What makes an AI Agent truly one
An AI Agent understands a request, decides on a course of action, acts on the client's systems with real read and write operations, maintains the state of the conversation across multiple steps, and incorporates corrections over time. A chatbot answers. An AI Agent answers, decides, and does.
Broken down into its components, a real Agent shows five distinctive behaviors.
- It understands. It interprets intent beyond the single sentence, including the context accumulated throughout the conversation.
- It decides. It chooses among multiple possible paths based on rules, available data, and configured priorities, without following a rigid question-and-answer decision tree.
- It acts on systems. It reads and writes on CRMs, ticketing systems, e-commerce platforms, and booking systems. The action produces a verifiable effect in the target system, not a sentence describing it.
- It maintains the state. A case opened three days ago and resumed today keeps its context, the decisions made, and the data already collected. The user doesn't have to repeat themselves.
- It learns. Human corrections flow back into the system through a traceable process, instead of vanishing at the end of the session.
If even one of these five behaviors is missing, it's not an AI Agent. It's a well-built conversational interface - useful, but different from what is being sold.
The five recognition criteria
None of these criteria requires technical expertise. It's enough to observe how the system behaves beyond the surface of the conversation.
1. Action, not just answers
Many systems describe the action instead of performing it. The AI Agent says it will open a ticket, update an address, or issue a refund, but the actual step is completed by an operator behind the scenes or by a separate integration, set up ad hoc for the occasion. A real AI Agent performs the operation in real time and leaves a verifiable confirmation in the target system - a case ID, a timestamp, an updated record that can be checked directly. The distance between describing and doing is the first signal, and also the easiest to observe.
2. State and orchestration
Real requests rarely open and close within the same session. A case can start in chat on Monday and continue over the phone on Thursday. A chatbot without true state management starts from scratch and asks again for information already provided. An AI Agent recognizes where the conversation left off, remembers what was decided in the meantime, and picks up without making the user repeat anything. Context persistence across channels and over time is an architectural capability, and it can't be improvised.
3. Verifiable learning
A precise distinction is needed here. One thing is architectural memory, that is, a structured process in which conversations are labeled, corrections are validated by a human team, and the Agent's behavior changes in a traceable way. Another is the generic promise of "continuous fine-tuning" or "self-improving AI," often with no verifiable mechanism behind it. The first model, with the team in the loop as the guarantor of quality, is the one worth building an enterprise implementation on. In projects set up this way, improvement becomes measurable over time, because the share of analyzed and labeled conversations grows and most of the proposed corrections are confirmed as correct by the team validating them. A serious vendor can show this loop in action, with the same level of detail - who validates the corrections, how often, and with what measurable evidence.
4. Action governance
Every action the AI Agent performs on company systems should have an explicit permission, a declared scope, and a consultable log showing who authorized what, when, and with what outcome. This is not a theoretical concern. In our Customer Experience Unlocked 2026 report, compliance of generated answers with company policies emerges among the main challenges (43%) Italian companies face when moving from pilot to production. Be careful not to confuse this criterion with due diligence on the vendor's data practices and certifications, such as hosting, encryption, or ISO 27001. That verification concerns the security of the vendor's infrastructure. What's being observed here is the authenticity of the agentic capability itself - whether the AI Agent can perform tracked, reversible actions, or whether the action is an illusion with no control behind it.
5. Declared boundaries
A serious vendor distinguishes correct escalations - set by policy, risk thresholds, or the client's explicit choice on certain types of cases - from avoidable ones, which stem from an undeclared technical limitation and which a good product reduces over time. A vendor claiming to handle everything is raising a red flag. No system in production handles everything, and whoever claims otherwise is selling an expectation they won't keep. Transparency about boundaries, even before being a matter of commercial honesty, is an indicator of product maturity.
The metrics that matter
Every vendor shows up with a number. The problem is that not all numbers measure the same thing, and some are easy to inflate without technically lying.
Deflection, meaning the share of requests that never reaches a human operator, is the easiest to inflate. It's enough to count as deflected any conversation that closes without a handoff, even when the user solved nothing and simply gave up, perhaps to try again by phone the next day. Automation rate and end-to-end resolution are harder to fake, because they require closing the loop with the system of record. The ticket is actually resolved, the return actually processed, and the case actually closed.
That's why a metric is only worth as much as its definition. What matters is how it's calculated and how it's measured, whether with data self-reported by the platform or verified against actual closure in the client's CRM or ticketing system. This is not a detail for analysts. Again, in our Customer Experience Unlocked 2026 report, measuring ROI and business impact remains an open challenge for 26% of companies moving from pilot to production.
The regulatory framework
The European AI Act adds a lens that is often ignored during the selection phase, and that actually says a lot about a vendor's maturity. The transparency obligation set out in Article 50 of the regulation - users must know they are interacting with an artificial intelligence system - applies to the majority of conversational deployments, which fall under the limited-risk category, and becomes binding on August 2, 2026. Not a distant horizon, but an immediate deadline for anyone starting a project today.
The point is not to ask whether the vendor is compliant, because everyone will answer yes. The point is how the disclosure is implemented. It can be built into the product's architecture, with a welcome message in chat or a voice announcement present by default, or it can be left as a task for the client to configure case by case. A mature vendor clearly distinguishes its own obligations as the provider of the system from those of the client as the deployer using it, and can explain where the line falls without offloading everything onto the other party. A vendor unable to articulate this distinction is signaling the same architectural immaturity that produces Agent washing on the technical front.
The difference is measurable
The difference between an AI Agent and a relabeled chatbot is not an opinion. It's a number that can be verified over time. Real automation autonomously closes the first level of customer care requests - in the most mature deployments, the share exceeds 90%, with reductions in tickets reaching operators of up to 95% - and keeps growing thanks to a tracked correction loop. A static deflection, by contrast, deflates at the first internal audit or the first seasonal shift in volumes.
The five criteria in this article - verifiable action, state that survives, tracked learning, action governance, declared boundaries - are exactly what a relabeled chatbot cannot show when asked to demonstrate them rather than describe them. Recognizing them during the selection phase takes an extra hour of attention. Discovering their absence after go-live costs a project.
FAQ
Is a chatbot connected to GPT an AI Agent?
No, not in itself. Connecting a frontier language model improves the quality of the answers, but on its own, it doesn't add the ability to act on systems, maintain a multi-step state, or learn in a traceable way. It's precisely the five criteria in this article that make the difference, regardless of the underlying model.
Does Agent washing only concern small vendors or younger startups?
No. It concerns any provider, small or large, young or established, that updates its sales vocabulary faster than it updates its architecture. Company size is not an indicator. The product's concrete behavior in production is.
How do you recognize a vendor that doesn't do Agent washing?
By its willingness to show evidence rather than describe it. A real audit log, even anonymized, with actions performed on a production system. The written definition and calculation method of every presented metric. And transparency about what its Agent doesn't handle, along with the reasons for that choice.
