# The Dangers of AI Agentwashing: How AI Agents Can Fail Us

> Agentwashing: why 8 out of 10 marketed AI agents are just chatbots in disguise, the anti-patterns that sink projects, and how to build a real agent program in 2026.

- Date : 2025-07-17
- Lecture : 15 min
- Catégorie : genai
- Tags : Generative AI, AI Agents, Tech Ethics, Risk
- URL : https://www.adservio.fr/en/insights/articles/les-dangers-de-l-agentwashing-ia-comment-les-agents-ia

## TL;DR

- Agentwashing refers to the misuse of the term "AI agent" for systems that lack its defining characteristics: roughly 80% of "agents" marketed in 2024-2025 were arguably glorified chatbots.
- A real agent is autonomous, goal-driven, context-aware, and interactive, distinct from agentic AI, the broader system architecture that surrounds it (orchestration, evals, observability, governance).
- Seven technical anti-patterns explain most failures: misapplying probabilistic tools to deterministic problems, compound uncertainty in agent chains, automating dysfunctional processes, vanity amplification, useless agent loops, cognitive overload, and solutionism.
- Chaining agents mathematically degrades accuracy: with 5 agents at 95% accuracy each, end-to-end accuracy drops to roughly 77%, and to 60% with 10 agents.
- In 2026, a real AI agent program is judged by its instrumentation, multi-layer evals, continuous observability, governance aligned with the AI Act, far more than by the sophistication of its initial prompt.

## Agentwashing: a definitional crisis with very real consequences

Agentwashing refers to the misuse of the term "AI agent" to describe systems that lack the fundamental architectural characteristics of an agent. This marketing trend dilutes the technical definition and creates unrealistic expectations, to the point where the term has come to cover almost anything that contains a single call to a large language model.

### A confusion that costs organizations dearly

In the 2026 enterprise AI landscape, terminological clarity has become a strategic imperative rather than an academic debate. Organizations that conflate chatbots with agents risk misdirected investments and major operational disappointments. At Adservio, we regularly see projects delayed by six to twelve months because of this initial confusion between a simple conversational assistant and a genuinely autonomous system.

The underlying problem: roughly 80% of systems marketed as "AI agents" in 2024-2025 were, in reality, glorified chatbots, and the trend has only partially corrected since. This semantic confusion leads to failed implementations and wasted budgets. Adservio estimates that this confusion costs large enterprises between 15% and 40% of their AI innovation budgets, mostly through late-stage rework and disappointed business expectations.

### Anatomy of fake agents

What is NOT an agent: a stateless chatbot, following a simple request-response pattern with no contextual continuity and unable to maintain a coherent conversational state; an automation script, whose deterministic if/else logic neither learns nor adapts to contextual variation; a basic LLM wrapper, a thin envelope around a language model with no state management or persistent memory; a passive conversational UI, limited to reactive replies with no capacity for action or initiative of its own. By 2026, a new and widespread variant has emerged: the RAG pipeline rebranded as an "agent", when it merely retrieves and rephrases information without any decision loop or action on the outside world.

In practice, the differences are stark. A real AI agent runs in an event-driven, autonomous mode, relies on persistent and contextual memory, makes adaptive, goal-oriented decisions, and triggers proactive actions across multiple systems, at a development cost of three to six months given the architectural complexity. A fake agent, in reality a chatbot, stays in reactive request-response mode, with stateless or session-limited memory, scripted and deterministic decision-making, purely conversational actions, and can be shipped in two to four weeks via a simple API integration.

## What is an AI agent, really?

This is not a debate about definitions. We have no intention of getting into semantic arguments over "AI agent" versus "agentic AI". Defining the concept perfectly will not stop your AI initiatives from failing if the architectural fundamentals aren't in place from the start.

### The four core traits of an effective agent

Regardless of UX form factor, a chat interface, a background process, or an embedded system, an effective AI agent is built around a focused task and demonstrates four core traits. It is autonomous: able to operate without constant human prompting, autonomy is clearly distinct from automation and does not mean operating unsupervised, but rather without micromanagement. It is goal-driven: designed to pursue and dynamically adapt goals, it transcends simple reactivity to inputs. It is context-aware: able to reason sophisticatedly over state, memory, and feedback, with persistent, structured memory usable for inference. It is interactive: able to take action, collaborate, negotiate, and delegate within a multi-actor ecosystem, beyond merely answering questions.

### Agentic AI: the system architecture around it

Agentic AI is the broader system architecture that enables and governs these agents: it includes multiple agents collaborating in a distributed decision network, AI evaluation (evals) frameworks for continuous monitoring, communication protocols that orchestrate decisions across agents and systems, the Model Context Protocol has largely become the interoperability standard in 2025-2026,observability layers to trace and debug agent behavior in production, and governance and security mechanisms that guarantee ethical alignment and regulatory compliance.

This architecture rests on five components whose business criticality varies: multi-agent orchestration and communication protocols, which coordinate distributed workflows and guarantee interoperability, are critical; the evaluation framework and observability layers, which ensure quality monitoring and enable debugging, are of high criticality; governance and security remain critical for compliance and risk management, all the more so as the European AI Act's full application to high-risk systems, in summer 2026, makes these mechanisms non-negotiable for many sectors.

Any use case that combines agents with tools such as AI evals, observability, or communication protocols qualifies as an AI agent or agentic AI initiative, whether it's a research assistant for drug discovery or a customer-success platform for selling farm equipment.

## Three anti-patterns that betray a poor tool-problem fit

In the analysis of AI agent deployment failures we conduct at Adservio, three technical anti-patterns systematically top the list of root causes. They share a common thread: using a probabilistic tool where a deterministic tool, or a redesigned process, would have sufficed.

### Using probabilistic tools for deterministic problems

This first pattern alone accounts for 35% of the inefficiency cases we detect in our audits. Using an LLM for tasks with a deterministic, well-defined output is a fundamental architectural mistake that generates hidden costs. An input like "create an S3 bucket with versioning and KMS encryption" must produce strictly deterministic Terraform code, not a different variant on every run. Yet an LLM generates varying code on every call once temperature is above zero, requiring extensive validation (syntax checks, security linting, policy compliance) and adding roughly 500ms of overhead versus 10ms for a template-based approach, a 50x performance degradation, at a cost of around $0.002 per generation versus near-zero for templating, up to $20,000 a year for 10 million generations. LLMs remain justified when requirements are ambiguous and need contextual interpretation, when facing pattern combinations impossible to enumerate exhaustively, or for generating tests from natural-language specifications. The rule is simple: if you're adding guardrails to force deterministic behavior, you're using the wrong tool, favor template engines (Jinja2, Handlebars), business DSLs, or rule engines instead.

### Compound uncertainty in agent chains

Chaining probabilistic agents without intermediate validation creates an exponential accumulation of errors. If each agent has 95% accuracy, a chain of N agents has an accuracy of 0.95 raised to the power N: with a single agent, accuracy stays at 95.0%; with 3 chained agents, it drops to 85.7%; with 5 agents, to 77.4%; with 10 agents, to just 59.9%. A typical ticket-processing pipeline illustrates this: a request passes through an intent-classification agent, an entity-extraction agent, a routing agent, and then a response-generation agent, each at 95% individual accuracy, for an end-to-end result that doesn't exceed 81% accuracy. Across a volume of 1,000 operations, a single agent generates about 50 failures, a chain of 5 agents generates 226, and a chain of 10 agents generates 401. Technical mitigations exist: human or deterministic validation checkpoints after each critical agent, confidence thresholds blocking the chain below 0.85, fallback paths for high uncertainty, and audit trails of confidence at every step.

### Automating a process that should have been redesigned

In his seminal 1990 HBR article, "Reengineering Work: Don't Automate, Obliterate", Michael Hammer warned that we should stop paving cow paths, rather than embedding obsolete processes into silicon and software. The quote remains just as relevant in the agentic AI era: after the automation wave of the 2000s and the cloud wave of the 2010s, the AI wave of the 2020s is repeating the same structural mistakes. Using AI agents to automate a broken process doesn't make it better; it makes the dysfunction faster, harder to detect, and harder to audit, especially behind an AI layer that creates a black-box effect. At Adservio, we estimate that 40% of AI agent projects fail precisely because they automate processes that should have been redesigned or eliminated. The numbers speak for themselves: automating the status quo without redesigning it produces only 15-20% gains, with high technical debt and a negative three-year ROI; optimizing the process before automating it raises gains to 40-60%; fully reinventing the process with AI enables gains of 70-150%, with low technical debt and a very positive ROI. We consistently recommend a "process archaeology" phase before any automation initiative: understand why the process exists in its current form before deciding whether to accelerate it or remove it.

## Four anti-patterns that betray poor product design

The following four anti-patterns are no longer about a poor technical fit, but about a drift in product design and the user experience of agentic systems.

### The vanity amplifier

Using AI to decorate trivial outputs won't make them meaningful. Agents that amplify the byproducts of a process, rather than its outcomes, simply scale noise. Generating flashy dashboards, auto-polishing slide decks, or writing process reports nobody reads contributes nothing to core value; in our audits, 25-35% of AI compute budget goes toward generating content that is never consulted. A sound product strategy must serve outcomes, not dress up the process around them: don't automate the output if the output didn't matter in the first place. This simple rule eliminates up to 40% of the AI use cases proposed during our strategic-scoping workshops.

### The useless agent loop

The vanity amplifier often paves the way for an even worse failure. It starts when a first agent generates or amplifies an in-process artifact, a meeting summary, a dashboard, a performance report, but the output is messy or unclear, prompting the introduction of a second agent to simplify or reformat it. You then end up with two agents in a feedback loop, one creating complexity, the other trying to clean it up, all around an artifact that never needed to exist. We've documented cases where up to five agents are chained to recursively fix the failing outputs of previous agents. If the loop exists to manage its own byproducts, it hasn't solved anything, it has automated the mess. Early detection requires dedicated metrics: input/output ratio, reformatting rate, correction cycles.

### Cognitive overload

AI agents are supposed to reduce cognitive load, not increase it. Yet their creative, guessing nature often leads users into long, polite but unstructured conversations, where the path feels productive until the user realizes they've completely lost the thread. We can extend an AI's context window to hundreds of thousands of tokens, but human working memory remains limited to roughly 7±2 items, per Miller's classic finding. Product teams need to consider new UX patterns: context anchoring, which shows users where they are in the flow, cuts cognitive load by 40%; short-term memory aids, which recap decisions made, cut it by 35%; escape hatches, allowing users to reset or go back, cut it by 25%; visual breadcrumbs cut it by 30%. At Adservio, these patterns, documented in our agentic design framework, have demonstrated a 60% improvement in user satisfaction scores.

### Solutionism

The classic solutionism trap is alive and well in the generative AI era: technical capability does not equal business value. An LLM can indeed tell you what to cook with what's left in your fridge or order a pizza on your behalf, that doesn't mean these are legitimate problems to solve. Customers don't buy AI; they buy outcomes, just as they did with the smartphone, the internet, or the computer. At Adservio, we estimate that 45% of AI agent POCs presented to investment committees are pure solutionism, with no validated problem-solution fit. The history of technology cycles confirms it: in the mobile era of the 2010s, the dynamic flipped toward 60% market-pull, yielding a 45% success rate; the AI of the 2020s has swung back to a 65/35 technology-push dynamic, with an estimated success rate of just 20%.

## Building a real AI agent program in 2026: evaluation, observability, and governance

Avoiding the seven anti-patterns isn't enough: you also need to instrument the agent's journey from prototype to production, with the same rigor as any other critical distributed system.

### Evaluate before, during, and after deployment

A mature AI agent program in 2026 rests on three layers of evaluation: offline evals against reference datasets before every production release, online evals (shadow mode, canary) against real traffic before full rollout, and continuous evals in production to catch performance drift. Evaluation frameworks have standardized around task metrics (completion rate, action accuracy), safety metrics (appropriate refusal rate, prompt-injection resistance), and experience metrics (perceived latency, human-intervention rate). Without this instrumentation, an agent chain running at 77% accuracy can stay invisible for months before a customer incident exposes the problem.

> Related read: [How to Evaluate an LLM System](https://www.adservio.fr/en/insights/articles/comment-evaluer-un-systeme-llm): Evaluating an LLM system in production: quality metrics, eval datasets, LLM-as-judge, prompt regression testing and continuous drift monitoring.

### Observe and govern continuously

Agent observability goes beyond simple application logging: it traces every decision, every tool call, and every reasoning step to enable post-incident debugging and continuous improvement. The Model Context Protocol has accelerated the standardization of integrations between agents and enterprise systems, but it also shifts the attack surface and demands access governance as strict as for any critical API. On the regulatory front, organizations operating in the European Union must now build the AI Act's traceability and human-oversight requirements directly into the architecture of their high-risk agents, rather than bolting them on afterward.

> Related read: [The Model Context Protocol: Beyond the Hype, the Standard for AI Agents](https://www.adservio.fr/en/insights/articles/le-protocole-model-context-au-dela-de-la-tendance): Architecture, 2026 specification, official registry, security: how the Model Context Protocol went from hype to the production standard for AI agents.

> Related read: [AI agents shouldn't be a security nightmare](https://www.adservio.fr/en/insights/articles/les-agents-ia-ne-doivent-pas-etre-un-cauchemar-de-securite): Prompt injection, data exfiltration, rogue agents: a six-layer framework to deploy AI agents safely in production, from least privilege to the kill switch.

## Focusing on what matters

The technology hype cycle follows predictable patterns, documented since Gartner's original work on the technology maturity curve. A hyped technology is always a mix of excitement, partial understanding, FOMO, and genuine ambition. The rise of AI agents is no exception. In our practice at Adservio, we currently position AI agents past the peak of inflated expectations, deep in the trough of disillusionment, ahead of a gradual climb toward a more realistic plateau of productivity.

Build with clear eyes, Strategic lucidity requires critical distance from the surrounding enthusiasm. We're not writing this to criticize experimentation or dampen innovation, both are vital to any paradigm shift. But if we want to move from hype to real, lasting value, we need to proceed with clear eyes. This transition requires rigorous governance, quantifiable success metrics, and a rational decision framework.

The failure patterns we've described aren't hypothetical; they have real consequences. Some come from our own missteps. They emerge when teams misapply probabilistic tools, amplify what doesn't matter, or automate processes they should have reimagined. These mistakes generate technical debt estimated between €200,000 and €2 million per failed enterprise project.

They reinforce an old truth: technology doesn't fix bad design; it scales it. This law of amplification applies with particular force to AI systems, whose probabilistic nature multiplies the effects of initial architectural flaws.

These principles have a direct impact on ROI: focusing on a real problem delivers up to +150% ROI, with maximum criticality and medium application difficulty; reinventing the process rather than automating it delivers +80%, with high criticality and high application difficulty; ensuring tool-problem fit delivers +60%, with critical criticality but low application difficulty; and favoring architectural simplicity delivers +40%, with high criticality and medium difficulty.

Old lessons still apply, The accumulated wisdom of software engineering remains fully relevant in the era of AI agents. The good news is that old lessons still apply: start with real problems and validate problem-solution fit before any technology investment; don't pave cow paths, question existing processes before automating them; don't solve deterministic problems with probabilistic tools; and always simplify, because complexity is the enemy of reliability.

Yes, AI agents are powerful, but they remain software. And success, as always, depends not on what we call them, but on how we design, integrate, and govern them. At Adservio, we help our clients adopt this pragmatic, structured approach to AI agent adoption, favoring lasting value over cosmetic innovation.

Let's build carefully, and build things that matter.

Note: The statements and opinions expressed in this article are those of the author and do not necessarily reflect Adservio's positions.

## FAQ

### What is agentwashing and why is it a problem?

Agentwashing refers to the misuse of the term "AI agent" to describe systems that lack the fundamental architectural characteristics of an agent, autonomy, goal orientation, context awareness, interactivity. This marketing dilution creates unrealistic expectations: roughly 80% of "agents" marketed in 2024-2025 were arguably glorified chatbots, leading to failed implementations and wasted innovation budgets.

### What distinguishes a real AI agent from a simple chatbot?

A real agent is autonomous (it operates without constant prompting, in event-driven mode), goal-driven (it dynamically adapts its strategy), context-aware (persistent memory and reasoning over history), and interactive (it acts, collaborates, and delegates). A chatbot stays in reactive request-response mode, with no contextual continuity or capacity for autonomous action.

### How do you build a real AI agent program in 2026 instead of a demo that never reaches production?

By instrumenting the agent from the prototype stage with three evaluation layers (offline, online, continuous in production), tracing every decision and tool call through dedicated observability, and building governance and human-oversight requirements, including AI Act requirements for high-risk systems, into the architecture from the start rather than bolting them on after an incident.