Agent Washing: The Definition — and a Scorecard to Catch It

Agent washing — Gartner’s term for rebranding chatbots, RPA, and AI assistants as “agentic” without substantial agentic capability — has become the defining diligence problem in enterprise software buying. Gartner estimated in June 2025 that of the thousands of vendors claiming agentic AI, only about 130 are the real thing. Every other vendor deck on your desk is, statistically, selling a relabel.

The stakes are not abstract. The same Gartner release predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls — and a mislabeled product is a head start on all three. Meanwhile the term itself has jumped from analyst vocabulary into vendor marketing, legal risk memos, and a second Gartner warning in May 2026.

This guide does three things. It pins the definition to its primary sources with the dates most coverage gets wrong. It grounds the pattern in two regulator-confirmed enforcement cases that predate the buzzword. And it gives you the Digital Applied Agent-Washing Test: six axes, scored 0-5, a 30-point maximum, and bands that tell you whether you’re buying an agent or a rebrand.

The term comes from Gartner, introduced in mid-2025. In its June 25, 2025 press release — the one predicting the 40%-plus cancellation rate — the firm named the practice and sized it. Vendors were, in Gartner’s words, engaging in “agent washing”: the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities. The deliberate echo of greenwashing is the point — it’s a claims problem, not a technology problem.

What made the release land was the number attached to it. Gartner estimated that of the thousands of vendors marketing agentic AI products, only about 130 were genuinely agentic. That is not a rounding-error minority — it implies the overwhelming majority of "agentic" pitches a buyer receives are repackaged automation wearing this year’s label.

Gartner analyst Anushree Verma’s framing in that release is worth internalizing rather than quoting back at vendors: most agentic projects today are early-stage experiments driven by hype and often misapplied, which blinds organizations to the real cost and complexity of deploying agents at scale. Her sharper observation cuts both ways for buyers — many use cases positioned as agentic today don’t require agentic implementations at all. Sometimes the washed product isn’t just mislabeled; it’s answering a question nobody in your workflow asked.

Here is the detail nearly all secondary coverage gets wrong. Gartner has issued two separate, dated agent-washing warnings, and they don’t say the same thing. The ~130-vendors statistic belongs exclusively to the June 25, 2025 release. On May 20, 2026, Gartner warned again — this time about agent washing in the supply chain planning technology market — without restating that figure. Articles that attribute “130 vendors” to the 2026 release are conflating two distinct statements made nearly eleven months apart.

The 2026 release matters for a different reason: it shows the warning aging into specific markets rather than fading. Per the release, Snoeckx warned that agent washing obscures the real differences between planning tools "by relabeling conventional automation as agentic, increasing the risk of misaligned investments and long-term lock-in.” And it sets a concrete autonomy bar: “True autonomous planning would require the automatic generation of plans, automatic selection of the optimal plan and seamless execution without human intervention. Most current solutions have not reached that level of end-to-end autonomy, and vendors claiming end-to-end autonomous supply chain planning before 2027 are overstating what is possible in the near term.” Swap “planning” for your own domain and that sentence is a usable diligence question.

Agent washing isn’t a con that succeeds against buyers' better judgment — it succeeds because of a structural gap between enthusiasm and experience. Gartner’s 2026 CIO survey, cited in its April 15, 2026 Hype Cycle for Agentic AI, found that only 17% of organizations have deployed AI agents to date, yet more than 60% expect to do so within the next two years — described as the most aggressive adoption curve among the emerging technologies measured. When intent runs more than three times ahead of deployment, most buyers evaluating an “agent” have never operated one. That inexperience is the surface agent washing sticks to.

The money tells the same story. A January 2025 Gartner poll of 3,412 webinar attendees found the largest bloc — 42% — making only conservative investments in agentic AI, with 19% investing significantly, 31% waiting or unsure, and 8% not investing at all. Commitment is cautious even while the vocabulary is everywhere.

None of this means the category is fake — the same June 2025 Gartner release forecasts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI (up from 0% in 2024) and 33% of enterprise software applications will include agentic AI (up from under 1% in 2024). The washing problem and the genuine trend coexist; that’s precisely what makes the label so profitable to misuse. It also compounds a separate, well-documented problem: even honestly built agent projects struggle to ship — most AI agent projects never reach production for execution reasons that have nothing to do with labeling. A washed product stacks a definitional failure on top of an already unfavorable base rate.

The clearest evidence the term has escaped analyst-land: vendors now use it against each other on stage.

An important precision: no major current AI-agent vendor has been publicly accused of “agent washing” by that name, and neither of the two regulatory cases below uses the term. What both cases document — with the evidentiary weight of federal enforcement — is the exact claims-versus-reality pattern Gartner describes: products marketed as autonomous that, in practice, ran on undisclosed human labor or untested capability claims. Treat them as the pattern’s case law, not its name-and-shame list.

Read the SEC’s Presto order and the FTC’s DoNotPay release with a buyer’s eye and two diligence questions fall straight out of them. From Presto: what percentage of this product’s output currently requires human intervention, and who exactly operates the AI? From DoNotPay: what testing supports the capability claim, and can we see it? Both questions reappear as scored axes in the test below.

This post won’t re-litigate the full definition of agentic AI — our agentic AI glossary and AI agent glossary cover the vocabulary in depth. What matters here is the dividing line the washing claim blurs. IBM’s reference definition draws it cleanly: an AI agent autonomously performs tasks by designing workflows with available tools — deciding, acting, and interacting with external systems beyond its training data. A chatbot is a modality; agency is a capability. An LLM-powered chat interface with no tools, no memory, and no reasoning loop is non-agentic no matter how fluent it sounds, and traditional automation that follows predefined rules doesn’t become an agent by being renamed.

AWS’s prescriptive-guidance comparison makes the same cut as a spectrum, scoring traditional AI, software agents, and agentic AI across characteristics like autonomy, proactivity, and agency — from “tools for humans” through “operates independently within predefined bounds” to “operates with purpose, goals, and self-direction.” Its strategic guide sketches the autonomy ladder most washed products quietly sit at the bottom of:

Buyers aren’t short of definitions — they’re short of a way to score a specific product against them. Frameworks exist, but they’re fragmented: AWS’s comparison is taxonomic rather than evaluative; enterprise vendor WRITER published a four-level autonomy framework (noting most misrepresentation involves marketing level-1 tools as level-3 or level-4 — though WRITER sells its own agent product, so read it as positioned analysis); and other vendor blogs offer question checklists without scoring. Nothing in that pile gives a procurement team a number. That’s the gap the next section fills.

The test scores a product on six axes, 0-5 each, for a maximum of 30 points. The axes are a Digital Applied synthesis — informed by IBM’s agent framing, AWS’s autonomy spectrum, and the public Gartner warnings above, but the anchors, weights, and bands are our editorial methodology. Score against what you observe in a live demonstration on your own scenario, never against the deck. Anchor descriptions are written for 0, 3, and 5; score 1-2 or 4 when the product sits between anchors.

Two scoring notes. First, the sixth axis is deliberately about honesty, not capability: a vendor with explicit, measured human checkpoints scores a 5 there even if the product is only partially autonomous — while a product hiding its humans behind an “autonomous” claim scores 0 on that axis regardless of how good the demo looks. That is the Presto lesson, encoded. Second, the bands measure labeling accuracy, not product quality — an 11-20 assisted agent can be an excellent purchase at an assisted-agent price. What the test catches is paying agentic prices, and accepting agentic risk assumptions, for rung-1 automation. For the price side of that equation, pair this test with what a real agent actually costs to build and run.

Scoring takes one structured session. Insist on a live run against a scenario you supply — not a recorded demo — and ask three questions the enforcement cases taught us to ask: What percentage of production outputs currently involve human intervention, and will you contractually report that number? Which steps in the flow we just watched were pre-scripted versus planned by the system? What testing evidence backs the headline capability claim? A vendor’s comfort with those questions is itself a signal — the genuinely agentic minority tends to answer them eagerly, because the answers are their moat.

Gartner’s May 2026 release offers three buyer missteps to avoid that map cleanly onto the test: don’t mistake vendor positioning for true autonomy, avoid monolithic transformations and legacy retrofits, and don’t pursue high-risk autonomous use cases too early. Then act on the band you scored:

The scorecard also sharpens two adjacent decisions. If you’re evaluating the big CRM platforms’ agent offerings, run the test axis by axis against our guide to CRM AI agents from Salesforce, HubSpot, and Zoho — the same six questions expose exactly where each platform’s "agent" sits on the autonomy ladder. And a low score reframes the build-vs-buy decision: if the branded “agent” scores 8, a purpose-built workflow you own — scoped to rung 2 or 3 honestly — often beats renting the label. That comparative evaluation is the first step of our AI transformation engagements, where we score shortlisted vendors against live demos before any budget is committed.

Looking forward, we expect the test’s sixth axis to become the whole game. With securities lawyers now framing overstated agent claims as disclosure risk, and Gartner tracking governance and FinOps profiles for agentic AI on its 2026 Hype Cycle, the direction of travel is toward measured, reported autonomy — the intervention rate as a contract term, not a secret. Vendors who can publish that number will take the premium; vendors who can’t will quietly drop the label. Buyers who score today are simply ahead of where procurement standards are headed by 2027.