Thousands of vendors now market themselves as sellers of “AI agents.” According to Gartner, only approximately 130 of them offer genuine agentic capability. Gartner named the gap between the marketing and the reality “agent washing” in a press release dated June 25, 2025. The term echoes greenwashing: a label substitutes for the underlying substance.
The distinction is not academic. A rebranded chatbot or a scripted robotic-process-automation (RPA) flow, sold under the word “agent,” does not do what an agent is supposed to do: plan a sequence of steps, decide which tools to call, and act on the result without a human scripting each step in advance. A business that signs a contract without checking which one it is buying has no way to verify what it purchased until the product fails to perform.
What is agent-washing?
Agent-washing is the practice of marketing a product as an “AI agent” when the system underneath does not perform genuine agentic work. Gartner’s June 2025 press release put a number on how widespread the practice has become: of the thousands of vendors that describe themselves as agentic-AI companies, only approximately 130 offer genuine agentic capability (Gartner press release, June 25, 2025). The rest range from conventional chatbots with a new label to rules-based RPA repackaged as “autonomous.”
The term names a buyer problem, not a market trend. A vendor's homepage, sales deck, and demo cannot be trusted at face value to establish what category of product is being sold. The label “agent” has become detached from a specific technical claim.
What separates a real agent from a scripted flow?
A genuine agentic system plans a course of action toward a goal, selects and calls tools or APIs to carry it out, and adapts that plan as new information arrives, without a human pre-defining every branch. A scripted flow, by contrast, follows a fixed sequence of steps written in advance: if condition A, do step B, then step C. Both can look identical in a sales demo, because a demo is built to walk a single happy path. The difference only shows up when a real case falls outside that path, and the scripted flow has nowhere to go.
This is the practical test a buyer needs, independent of what a vendor's marketing calls the product: does the system decide its own next step, or was that step decided by an engineer before the buyer ever saw it?
What did Gartner actually say?
Gartner's press release of June 25, 2025 makes two connected claims. First, on agent washing: only approximately 130 of the thousands of self-described agentic-AI vendors offer genuine agentic capability. Second, on what happens to projects built on that hype: over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value, or inadequate risk controls.
Gartner analyst Anushree Verma, quoted in the same release, explains why: “Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype.” The two figures describe the same underlying problem from two angles: one counts how many vendors are selling something other than what they claim, the other counts how many of the resulting projects do not survive contact with a real budget cycle.
Why do vendors rebrand RPA and chatbots as agents?
The commercial incentive is straightforward. “Agentic AI” is the term commanding attention and premium pricing in 2025 and 2026, and a vendor with an existing chatbot or RPA product can reach that category with a website update rather than a rebuild. Rebranding is faster and cheaper than building genuine planning, tool-use, and autonomous decision-making into a product. In a fast-moving market, a vendor that waits to build the real thing risks losing the deal to a competitor that did not wait.
Buyer-side incentives compound the problem. A business under pressure to show an AI initiative is often evaluating on the strength of a demo and a sales narrative, not a technical audit of what the system does under load or edge cases. That asymmetry is what agent-washing exploits.
What does it cost a business to buy into the hype?
A business that signs with an agent-washed vendor pays for more than a chatbot with a new name. It buys into a population of projects that Gartner forecasts will see over 40% canceled by the end of 2027, for the same underlying reasons that produce agent-washing in the first place: costs escalate past what a scripted or shallow system can justify, business value was never clearly defined before the contract was signed, and risk controls were never built because the product was never designed to need them.
A separate, related data point shows how a business acquires agentic capability in the first place. Research from MIT NANDA found that internally built AI systems succeed at approximately 33%, compared with approximately 67% for purchased tools (Fortune, August 18, 2025). Buying is directionally the safer path than building from scratch, but only if the thing being bought is a genuine agent, which is exactly the fact agent-washing obscures.
How do you tell a real agent from a scripted flow before you sign?
Three questions, asked before a contract is signed, do most of the diligence work:
- Does it plan and act, or does it follow a script? Ask the vendor to show the system handling a case that falls outside the demo path: a missing field, a contradictory input, a tool call that fails. A genuine agent adapts. A scripted flow either breaks or hands the case to a human with no visible reasoning in between.
- What is the monthly cost at real volume, not demo volume? A demo runs a handful of cases. Production runs thousands. Ask for the per-transaction cost at the volume the business actually processes, not the volume the vendor used to build the pitch.
- Who is accountable when the agent is wrong? A genuine agentic deployment needs a named answer to this question before it goes live, not a general assurance that the system is “highly accurate.” If no one can name who reviews a wrong output and what happens next, the risk control Gartner cites as a cancellation cause has not been built.
A vendor that cannot give a specific, demonstrable answer to any of these three questions is not necessarily lying about what it sells. The buyer still has no basis to know it is buying a genuine agent rather than a rebranded one.
What kill thresholds should you set before you sign?
The diligence above answers what a business is buying. A separate step answers what happens if the answer turns out to be wrong after deployment: agreeing on kill thresholds before the contract is signed, not after the invoices start arriving.
A kill threshold is a specific, pre-agreed condition under which the project is shut down rather than carried forward on the strength of money already spent. Set before anyone has a stake in the outcome, a threshold turns a difficult judgment call into the execution of a decision that was already made. Useful thresholds are measurable and dated in advance: a cost-per-transaction ceiling the system must stay under by a fixed review date; a minimum rate of cases it must resolve without human correction; a maximum error rate on the categories of case that matter most to the business. If the system misses the threshold at the scheduled review, the project ends at that review instead of continuing through another quarter of hoping it improves.
This step separates a business that can walk away from a failed pilot from one of the projects inside Gartner’s 40% cancellation forecast. Both may fail, but only one fails on a plan decided in advance rather than argued over after the fact.
Where Mirai360 fits
Mirai360 AI provides the stack to build, run, and govern AI agents, including the evals and guardrails that make the diligence questions above answerable with evidence rather than a vendor's assurance. Whether a business is evaluating a third-party vendor or building its own agentic workflow, the same standard applies: the system should be able to show its reasoning on a case outside the happy path, its cost at real volume, and a named owner for the moments it gets something wrong.
A 30-minute call is the place to run that diligence against a specific vendor, a specific workflow, or a specific proof of concept already underway.
Talk to us
Book a free 30-minute call to walk through vendor diligence for an agentic AI purchase already on the table, or to scope what a genuine agent would look like for a specific workflow: https://calendly.com/shivang-mirai360/30min
Sources
- Gartner, press release, June 25, 2025 (agent washing, ~130 genuine vendors, 40%+ cancellation forecast, Anushree Verma quote)
- Fortune, August 18, 2025 (MIT NANDA, internal-build vs. purchased-tool success rates)
FAQ
- What is agent-washing?
- Marketing a product as an AI agent when the system underneath does not perform genuine agentic work. Gartner estimates only ~130 of thousands of self-described agentic-AI vendors offer genuine agentic capability (Gartner, June 25, 2025).
- What separates a real agent from a scripted flow?
- A real agent plans, selects tools, and adapts without a human pre-scripting every branch. A scripted flow follows a fixed sequence and breaks outside its demo path.
- How do you tell a real agent from a scripted flow before you sign?
- Ask three questions: does it plan and act or follow a script, what is the monthly cost at real volume, and who is accountable when the agent is wrong.
- What kill thresholds should you set before you sign?
- Pre-agreed, dated, measurable conditions (cost ceiling, resolution rate, error rate) under which the project is shut down rather than carried forward on sunk cost.