Agent Washing: Why 95% of "AI Agents" Aren't Actually Agents
Cutting through the buzzwords to help you spot the real deal
If you’ve attended a vendor presentation, conference, or even opened Youtube or LinkedIn in the past six months, you’ve almost certainly heard the term “agentic AI.” It’s everywhere. Every software company suddenly has an “AI agent” to sell you.
Leading AI researcher Andrew Ng pointed out in a recent talk, once the term “Agentic” gained traction, marketers immediately “slapped it as a sticker on everything in sight”.
However, according to Gartner, only about 130 of the thousands of vendors claiming to offer agentic AI solutions are actually delivering genuine agentic capabilities. The rest? They’re engaged in what analysts now call “agent washing”, which is rebranding existing chatbots, robotic process automation (RPA), and scripted workflows under a shiny new label.
This isn’t just semantic pedantry. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. If you’re making investment decisions based on vendor claims, you need to know the difference between genuine innovation and marketing spin.
Let’s cut through the noise.
What Actually Makes AI “Agentic”?
The term “agentic” comes from “agency” which is the capacity to act independently and purposefully. But what does that mean in practice? Here are the four characteristics that define a genuine AI agent:
1. Autonomous Goal Pursuit
A real agent doesn’t just respond to individual prompts. It takes a high-level objective and independently breaks it down into steps, executes those steps, and adjusts its approach based on results. The key word is “independently” - without requiring constant human instruction at every stage.
2. Reasoning and Planning
True agents think several moves ahead. They consider the consequences of actions, evaluate multiple approaches, and select strategies based on context. They’re not following a script; they’re solving problems.
3. Tool Use and Environment Interaction
Agents don’t just generate text. They interact with external systems - calling APIs, querying databases, executing code, browsing the web, manipulating files. They take action in the real world (or at least the digital one).
4. Learning and Adaptation
Genuine agents adjust their behaviour based on feedback and outcomes. When something doesn’t work, they try a different approach. They maintain context across extended interactions and learn from the results of their actions.
A useful way to think about a genuine AI agent is to compare it to a someone doing their day job. You don’t give a senior employee a prompt every five minutes. You give them an objective, some constraints, and access to the tools they need. They decide how to approach the task, adapt when something doesn’t work, use different systems as required, ask for help, and keep going until the outcome is achieved. That combination of autonomy, judgement, tool use, and adaptation is what we recognise as competence in people, and it’s exactly what “agentic” is meant to describe in AI.
If a product doesn’t exhibit most of these characteristics, it’s not an agent. Full stop.
The Litmus Test: What Actually Is an "Agent"?
Let’s look at where agentic AI is actually working today.
AI Coding Assistants: The Clearest Example
Software development has emerged as what New York Magazine called “the most definitive use case of AI agents.” Here’s why coding assistants like Claude Code, GitHub Copilot’s, Cursor’s Agent Mode, and Devin represent genuine agentic capability:
They pursue multi-step goals autonomously. Give a coding agent a task like “refactor this authentication system to use OAuth2” and it doesn’t just suggest code snippets. It analyses your codebase, creates a plan, writes the code, runs tests, identifies failures, debugs issues, and iterates until the task is complete - all without you micromanaging each step.
They interact with real tools and environments. These agents execute code, run test suites, access file systems, make Git commits, and even browse documentation. They’re not just generating text; they’re taking action.
They reason and adapt. When a test fails, a coding agent analyses the error, hypothesises about the cause, tries a fix, and repeats until successful. This is fundamentally different from autocomplete that suggests the next line of code.
As one senior engineer at Render put it after extensive testing: “This isn’t your typical AI assistant that just suggests code. The agent literally takes over your development environment.”
The key insight here is that coding is a domain where agent capabilities can be meaningfully measured. Success is objective; either the code works or it doesn’t. This makes it harder to hide behind vague claims of “AI-powered” capability.
Other Genuinely Agentic Applications
Beyond coding, genuine agentic AI is emerging in:
Research and Analysis: Tools like Deep Research that can autonomously gather information from multiple sources, synthesise findings, and produce comprehensive reports without step-by-step prompting.
Complex Customer Service: Systems that can investigate issues across multiple internal systems, make decisions about refunds or credits within defined parameters, and take action - not just route to a human.
Trading and Financial Operations: AI systems that continuously monitor market conditions, make autonomous trading decisions within guardrails, and adapt strategies based on results.
The common thread? These applications involve sustained autonomous operation toward a goal, with genuine decision-making and action-taking along the way.
What Isn’t Agentic AI (Despite the Marketing)
Here’s where it gets uncomfortable for a lot of vendors. These products may be useful, but calling them “agents” is misleading:
Rebranded Chatbots
If your “agent” is fundamentally a question-and-answer system that responds to individual prompts without maintaining goal-directed behaviour across interactions, it’s a chatbot. Adding generative AI to a chatbot makes it a better chatbot. It doesn’t make it an agent.
The tell: Does it wait passively for your next question, or does it proactively work toward a goal? If it’s the former, it’s not an agent.
Scripted Workflow Automation
Many “AI agents” are actually traditional workflow automation with an LLM layer on top. The system follows predetermined steps, calls the LLM to handle certain inputs, and routes based on rules. The workflow is designed by humans; the AI just handles specific nodes. It’s tempting to call these workflows agents because they do carry out tasks, sometimes quite complex ones, but their behaviour is ultimately scripted rather than autonomous.
As InfoWorld noted, many vendors are selling “deterministic workflows plus LLM calls” while marketing them as “autonomous, goal-seeking agents.”
The tell: Can the system deviate from its scripted path to handle unexpected situations? Can it choose a different approach when its planned method fails? If not, it’s automation, not agency.
RPA with AI Garnish
Robotic process automation handles rule-based, repetitive tasks, clicking buttons, filling forms, transferring data. Adding AI-powered text extraction or classification doesn’t transform RPA into agentic AI. The fundamental architecture remains: pre-programmed steps executing in sequence.
The tell: Does the system require humans to define the exact process, or can it figure out how to accomplish a goal on its own?
“AI Assistants” and “Copilots”
Most products labelled as AI assistants or copilots provide suggestions that humans then implement. They augment human capability rather than acting independently. Useful? Absolutely. Agentic? No.
The tell: Is a human still making every decision and taking every action? Then it’s an assistant, not an agent.
Why This Matters in Business
This isn’t just a theoretical distinction. Gartner’s research reveals real consequences:
Misaligned expectations lead to wasted investment. When CIOs implement what they believe is agent-based automation but get “glorified automation instead, they miss out on the transformative potential of true agents while still paying the premium,” notes Payam Shayan, founder of a management consultancy. A recent poll found that 31% of organisations are taking a “wait and see” approach, and given the confusion in the market, that caution is justified.
Maintenance overhead. Labelling a scripted or workflow-based system as “agentic” sets the wrong expectations. These systems don’t adapt; they depend on every step continuing to work exactly as designed. When upstream systems change, as they inevitably do, the workflow degrades, breaks, or silently produces incorrect results. The result is ongoing maintenance, hidden operational cost, and frustration when a product sold as “autonomous” behaves like fragile automation.
The wrong tool for the job creates technical debt. If you’re trying to automate complex, variable processes with tools that only handle scripted workflows, you’ll end up building elaborate workarounds that become maintenance nightmares.
Security and risk profiles differ dramatically. A true autonomous agent that can take actions in your systems has a fundamentally different risk profile than a chatbot that generates suggestions. If you’re evaluating them as equivalent, your risk controls are probably inadequate.
How to Evaluate “Agentic” Claims
Before your next vendor conversation, arm yourself with these questions:
1. “Can you show me the system handling an unexpected situation?”
Real agents adapt. Ask for a demonstration where the system encounters something it wasn’t explicitly designed to handle. Does it reason through the problem, or does it fail or default to human escalation?
2. “What decisions can this system make autonomously?”
Get specific. Not “it can handle customer queries” but “it can issue refunds up to £500 without approval” or “it can modify production code after passing tests.” If they can’t give you concrete examples of autonomous decision-making, be sceptical.
3. “How does it interact with external systems?”
Real agents need to take action. What APIs does it call? What databases can it query? What actions can it perform? If the answer is essentially “it generates text that humans then act on,” it’s an assistant, not an agent.
4. “What happens when its first approach fails?”
This reveals whether you’re looking at genuine reasoning or scripted responses. An agent should be able to explain its reasoning, identify why an approach failed, and try something different.
5. “Can I see the architecture diagram?”
InfoWorld’s advice is blunt: “If a vendor cannot explain, in clear technical language, how their agents decide what to do next - if they talk vaguely about ‘reasoning’ and ‘autonomy’ but when pressed, everything trickles down to prompt templates and orchestration scripts - that’s agent washing.”
The Path Forward
None of this means agentic AI isn’t real or valuable - it is. Equally it is important to be clear that AI-powered workflows and automations are extremely useful in their own right. The risk comes when organisations believe they are buying autonomous agents, when in reality they are purchasing well-designed, but ultimately scripted systems that still require human oversight and maintenance.
Gartner predicts that 33% of enterprise software applications will include genuine agentic AI by 2028, and 15% of day-to-day work decisions will be made autonomously by that point. The technology is advancing rapidly.
But we’re in a period where hype has massively outpaced reality. As Anushree Verma, Senior Director Analyst at Gartner, puts it: “Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.”
The executives who navigate this period successfully will be those who:
Start with clear use cases where genuine agency adds value, rather than chasing the buzzword
Evaluate critically rather than accepting vendor claims at face value
Match the tool to the job - sometimes a well-designed workflow is exactly what you need
Build internal capability to assess these technologies independently
The agent washing will eventually fade as the market matures. Until then, stay sceptical, ask hard questions, and remember: if it’s not autonomous, goal-directed, reasoning, and taking action in the real world, it’s not an agent.
It might still be useful. Just don’t pay agent prices for it.
Recommended Viewing:
For a deeper dive into how Agentic workflows are increasing speed and changing how startups are built, this talk is essential viewing. Andrew Ng explains why “Agentic AI” is the most significant trend of the year and how it differs from the static prompting we have grown used to.
Andrew Ng: Building Faster with AI
This video is relevant because it provides a credible, technical breakdown of the “Agentic” concept from one of the leading minds in AI, specifically contrasting linear workflows with iterative, agentic loops.