Of the thousands of vendors selling “agentic AI” right now, approximately 130 actually have the technology to back up the claim. That number comes from Gartner’s 2025 analysis, which also projects that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The rest of the market? Chatbots in trench coats. RPA with a new logo. Workflow builders that got a rename and a 10x price hike.
The industry has a name for this: agent washing. It is the AI equivalent of greenwashing, where companies slap the word “agentic” onto existing products without adding any of the capabilities that make agents genuinely useful, like autonomous goal pursuit, dynamic planning, or error recovery without human intervention.
What Agent Washing Looks Like in Practice
Agent washing is not always obvious. The vendors doing it are not necessarily lying; they are stretching definitions until the term “agent” means nothing. Here are the most common patterns, drawn from Gartner’s market analysis and real-world audits documented by Particula.
The Chatbot Rebrand
A customer service chatbot that follows a decision tree, escalates to a human when stuck, and can only answer questions it has been explicitly programmed for gets relabeled as a “customer success agent.” The interface might look modern. It might use an LLM for language generation. But it cannot set goals, plan multi-step solutions, or act on the customer’s behalf across systems. It waits for input, generates a response, and stops. That is a chatbot, regardless of what the marketing page says.
The RPA Costume Change
A robotic process automation tool that executes a fixed sequence of clicks, form fills, and data transfers gets rebranded as an “intelligent process agent.” The workflows are still hard-coded. The tool still breaks when a website changes its layout. It cannot adapt its approach, learn from failures, or handle scenarios outside its script. Mindflow’s analysis of the agentic automation market found that many “agentic RPA” products add an LLM wrapper for natural language configuration but keep the same brittle, deterministic execution underneath.
The Workflow Automation Upsell
Marketing automation platforms that orchestrate email sequences, CRM tools that trigger actions based on if/then rules, and project management software that auto-assigns tasks are increasingly sold as “agentic systems.” They execute predefined logic. They do not reason about whether the logic makes sense for a specific situation. The distinction matters because a real agent would notice that a customer who just filed a complaint probably should not receive a promotional email three minutes later, even if the workflow rule says to send it.
Why 40% of Agentic AI Projects Will Be Canceled
Gartner’s prediction is not pessimism. It is pattern recognition. Based on a poll of 3,412 webinar attendees in January 2025, 19% of organizations had made significant investments in agentic AI, while 42% made conservative investments. But most of that spending was going to vendors who could not deliver real autonomy.
The failure drivers stack up in a predictable sequence.
Costs escalate faster than value. Real agentic systems require persistent state management, orchestration infrastructure, guardrail layers, and continuous monitoring. When organizations buy an “agent” that is actually a chatbot, they end up bolting on all that infrastructure themselves, often spending 3-5x the original purchase price on integration work. XMPRO’s post-mortem analysis of failed agentic projects found that most teams underestimated infrastructure needs by 60-80%.
Business value stays unclear. A genuine agent that autonomously handles expense report approvals, vendor negotiations, or incident triage creates measurable value: fewer human hours, faster resolution, lower error rates. A rebranded chatbot that still requires human oversight at every decision point creates overhead, not value. The Outreach.io guide to agent washing documents cases where companies paid enterprise pricing for “AI sales agents” that turned out to be glorified lead scoring tools with a chat interface.
Risk controls are missing. Real agents need real governance: audit trails, rollback capabilities, permission boundaries, output validation. Agent-washed products rarely include these because the underlying product was never designed for autonomous operation. When something goes wrong (and it will), there is no mechanism to contain the damage.
The Five Traits That Separate Real Agents from Fakes
After reviewing Gartner’s framework, McKinsey/QuantumBlack’s evaluation criteria, and Amazon’s real-world agent deployment lessons, five capabilities consistently distinguish genuine agentic systems from rebranded automation.
1. Goal Decomposition, Not Script Execution
A real agent takes a high-level objective (“reduce our accounts receivable cycle from 45 days to 30 days”) and breaks it into subtasks autonomously. It decides which invoices to follow up on, drafts appropriate messages, escalates stuck accounts, and adjusts its strategy based on results. A fake agent executes a predetermined sequence: send reminder on day 30, send second reminder on day 37, flag for human on day 42.
Vendor test: Ask the vendor what happens when the agent encounters a scenario that was not in the training data. If the answer involves “configuration” or “custom rules,” you are looking at automation, not an agent.
2. Tool Use and Environment Interaction
Genuine agents select and use tools dynamically. They can query a database, call an API, read a document, and compose findings into an action, all within a single task and without a human choosing which tools to invoke. Anthropic’s evaluation framework specifically measures whether an agent picks the right tool for the task, not just whether it can use tools at all.
Vendor test: Ask to see the agent’s tool invocation logs. If every run uses the same tools in the same order, it is a pipeline, not an agent.
3. Error Recovery Without Human Intervention
This is the capability that most clearly separates real agents from pretenders. When a real agent’s API call fails, it tries an alternative endpoint. When a document is in an unexpected format, it adapts its parsing strategy. When its initial plan is not working, it replans. A rebranded workflow tool throws an error and waits for a human to fix it.
Vendor test: Deliberately introduce a failure condition during a demo. Change a file format, revoke an API key, give it contradictory instructions. Watch what happens.
4. Persistent Memory and Context
Agents maintain context across sessions and tasks. They remember that the last time they contacted vendor X, the response took 72 hours, so they send the follow-up earlier this time. They track which approaches worked and which did not. Chatbots and RPA tools start fresh every time or rely on manually maintained context databases.
Vendor test: Ask the agent about something it handled last week. If it cannot reference past interactions without a human loading context, it lacks a core agentic capability.
5. Explainable Decision Chains
A real agent can tell you why it made each decision in a multi-step process. Not just what it did, but why it chose that approach over alternatives. This matters for governance, debugging, and trust. Amazon’s deployment experience highlights that the ability to audit decision chains is the first thing enterprise buyers should verify.
Vendor test: Ask for a detailed trace of a completed task. If the vendor can only show inputs and outputs but not the reasoning steps in between, the “agent” is a black box, and a black box is usually not an agent.
The Vendor Evaluation Checklist
Before signing a contract with any vendor claiming agentic capabilities, run through this checklist. It is built from the patterns above and from AppsTek Corp’s enterprise evaluation framework.
| Question | Red Flag Answer |
|---|---|
| What happens when the agent encounters an unknown scenario? | “It escalates to a human” (always) |
| Can the agent modify its own plan mid-execution? | “It follows the configured workflow” |
| How does the agent select which tools to use? | “Tools are configured per workflow” |
| Show me the decision trace for a completed task | Only shows input/output, no reasoning |
| What state does the agent maintain between sessions? | “Each session starts fresh” |
| How do you handle rollbacks when the agent makes a mistake? | No rollback mechanism exists |
| What is the agent’s failure recovery behavior? | “It retries the same action” or “It stops and alerts” |
If three or more of these produce red-flag answers, you are evaluating a rebranded automation tool, not an agent.
What Actually Works Right Now
The 130 vendors Gartner identified as having real agentic capabilities tend to share a few architectural patterns. They use LLMs for reasoning and planning, not just language generation. They implement ReAct (Reason + Act) loops or similar architectures where the agent iterates between thinking and acting. They provide explicit tool-use frameworks, often built on protocols like MCP or A2A. And they include observability infrastructure, because you cannot govern what you cannot see.
Frameworks like LangGraph, CrewAI, and AutoGen give teams the building blocks for real agentic behavior. But even these require careful implementation. The framework does not make your system agentic any more than owning a guitar makes you a musician. The agentic part comes from how you compose the reasoning loops, tool access, memory management, and guardrails.
For enterprise buyers evaluating vendor claims in 2026, the safest approach is to demand a proof-of-concept on your own data, with your own edge cases, and with failure conditions you design. Any vendor confident in their agent’s capabilities will welcome that test. Any vendor who insists on a controlled demo with predetermined inputs is probably selling you a workflow with a new name.
Frequently Asked Questions
What is agent washing in AI?
Agent washing is the practice of rebranding existing chatbots, RPA tools, and workflow automation as “AI agents” or “agentic AI” without adding genuine autonomous capabilities like goal decomposition, dynamic tool use, or error recovery. Gartner estimates only about 130 of thousands of vendors claiming agentic capabilities actually deliver real autonomous agents.
How many agentic AI vendors are real according to Gartner?
Approximately 130 out of thousands. That means roughly 95% of products marketed as AI agents are not agents at all. They are existing products (chatbots, RPA, workflow automation) rebranded to capitalize on the agentic AI hype cycle.
Why are 40% of agentic AI projects being canceled?
Three compounding factors: costs escalate 3-5x beyond initial estimates because agent-washed products lack the infrastructure real agents need; business value stays unclear because rebranded tools still require human oversight at every step; and risk controls are missing because the underlying products were never designed for autonomous operation.
How can you tell if an AI agent is real or fake?
Test for five core capabilities: autonomous goal decomposition, dynamic tool selection, error recovery without human intervention, persistent memory across sessions, and explainable decision chains. The simplest test is to deliberately introduce a failure condition during a vendor demo and watch what happens. Real agents adapt. Fakes crash or escalate.
What is the difference between RPA and real agentic AI?
RPA executes fixed, predetermined sequences and breaks when conditions change. Real agentic AI uses LLM-based reasoning to decompose goals, select tools dynamically, adapt plans mid-execution, and learn from past interactions. Many vendors disguise RPA by adding an LLM layer for natural language configuration while keeping the same rigid execution logic underneath.