Agent Washing: The Four-Question Test Before You Buy Another 'AI Agent'

TL;DR

“Agent washing” is Gartner’s name for the industry’s newest trick: taking a chatbot, a robotic-process-automation script, or a basic assistant, and relabelling it an “AI agent” with no meaningful change to what it actually does. Of the thousands of vendors now marketing agentic AI, Gartner reckons only around 130 meet its own definition of the real thing - and it predicts over 40% of agentic AI projects will be cancelled by the end of 2027, mostly because the technology underneath them was never agentic to begin with. This matters for anyone in the mid-market currently being pitched an “AI agent” for their CRM, their support desk, or their marketing stack. Genuine agentic AI is already producing real results elsewhere on this blog. The job now is telling it apart from the imitation before it’s on your invoice.

What Is “Agent Washing,” and Why Has Gartner Given It a Name?

Because the label had stopped meaning anything, and Gartner’s analysts got tired of watching buyers pay for the change of name rather than the change of capability.

Agent washing is the rebranding of an existing chatbot, RPA workflow, or AI assistant as an “AI agent”, with no substantial increase in autonomy underneath the new label. The interface looks the same. The demo looks the same. Only the word on the pricing page has changed, usually alongside the price.

Gartner’s own estimate, published alongside its 40% cancellation prediction, is stark: of the thousands of vendors now marketing agentic AI, only around 130 meet Gartner’s definition of a genuine agent. Everyone else is, to some degree, washing.

That’s not a rounding error. That’s most of the market.

Why Will Over 40% of Agentic AI Projects Be Cancelled by 2027?

Not because agentic AI doesn’t work. Because most of what got bought under that name was never agentic in the first place.

Gartner’s prediction attributes the coming wave of cancellations to escalating costs, unclear business value, and inadequate risk controls. Read between the lines and a pattern emerges: a business bought a product marketed as an autonomous agent, discovered it still needed a human to babysit every step, couldn’t point to a metric that improved, and eventually pulled the plug.

This is the sequel to a stat we’ve covered before. We wrote in November about the 45% of martech leaders who said vendor AI agents were already failing to meet expectations - before “agent washing” even had a name. The pattern hasn’t changed. The label has just got more specific about what’s actually going wrong.

And the peg for this is fresh. Gartner’s inaugural 2026 Hype Cycle for Agentic AI, published this spring, places the category squarely at the Peak of Inflated Expectations - the point where vendor claims and market attention have overtaken what the technology reliably delivers in production. The same Gartner CIO survey behind that placement found 17% of organisations have actually deployed AI agents, against over 60% who expect to within two years. That gap between “deployed” and “expects to deploy” is exactly the gap agent washing exploits. Vendors selling into “expects to deploy” don’t need a working agent. They need a plausible demo.

McKinsey’s parallel finding makes the stakes concrete: only around 6% of organisations qualify as AI “high performers”, defined as attributing more than 5% of EBIT to AI, even as 88% report using AI somewhere in the business. Adoption and value have quietly become two different numbers. Agent washing is one of the reasons they’ve drifted apart.

How Do You Tell a Real Agent From a Renamed Workflow Trigger?

You ask it to show its working, not its marketing.

The genuine article and the washed imitation are actually easy to tell apart once you know where to look, because the underlying mechanics are different in kind, not degree.

That table is the theory. In practice, four questions do the job.

First, ask for a decision trace on a task the agent completed without a human scripting every step in advance. If the vendor can only show you a transcript of prompts and replies, you’re looking at a chatbot with a rebrand.

Second, ask what percentage of production outputs still require human correction - and ask them to put that number in the contract. Vendors happy to quote a real figure have usually got one worth quoting. Vendors who deflect to “it’s improving all the time” usually don’t.

Third, ask what happens when a tool call fails mid-task. A genuine agent retries, reroutes, or escalates with context. A washed product typically just stops, because there was never any planning logic to fall back on - only a script that assumed nothing would go wrong.

Fourth, ask for an audit log you can read yourself, not a dashboard curated by the vendor. Tool-equipped AI needs governance, not avoidance - and the same logic applies to procurement. If you can’t see what the agent actually did, you can’t tell whether it did anything agentic at all.

Run those four and most agent-washed products fail within the first two.

What Does a Genuine Agentic Deployment Actually Look Like?

Unglamorous, specific, and already working elsewhere in this blog’s back catalogue - which is exactly the point.

We covered the shift from chatbot to working colleague back in May: genuine agents handling lead qualification, order processing, and contract review with defined stopping points and measurable baselines. None of that required an agent that operates with zero human oversight. It required an agent that could plan, use tools, and adapt - the three things a washed product can’t do, however confidently it’s marketed.

The tell isn’t autonomy for its own sake. It’s whether the system can handle the step nobody scripted for it.

Why Doesn’t the Fix Start With Buying a Better Agent?

Because a genuine agent bolted onto a broken foundation still fails - just for a different reason than a washed one does.

We made this case in detail in our last pillar post: AI agents are only as reliable as the data layer they can reach. Half of martech leaders already say the vendor AI agents they’ve piloted don’t meet expectations, and a good share of that isn’t agent washing at all - it’s a genuine agent given messy, siloed, half-trusted data to work from. Vetting the vendor is step one. Getting your CRM and marketing stack properly connected is step two, and skipping it undoes step one.

Buy a real agent and point it at bad data, and you’ll cancel that project too - just eighteen months later, and for a more expensive reason.

What Should You Do Before You Sign Anything Called an “Agent”?

Three steps, in order, before budget gets committed.

1. Run the four-question test on the vendor, not the sales deck. Decision trace, correction rate, failure handling, readable audit log. If any answer is vague, price that vagueness into the risk, not into your optimism.
2. Check the data the agent will actually touch. A genuine agent given unreliable inputs produces unreliable outputs with more confidence, not less. Confirm the foundation before you confirm the contract.
3. Define the cancellation criteria before the deployment criteria. Decide now what “not working” looks like and by when you’ll know. Projects that get quietly extended forever are the ones that end up in next year’s 40%.

None of this requires distrust of agentic AI as a category. It requires the same scepticism you’d apply to any vendor claim that arrived this fast, backed by this much marketing spend, in a category this new.

The Bottom Line

Agentic AI is real, it’s already working in production for the businesses that vetted it properly, and it isn’t going away. Agent washing is real too, and it’s currently the more common product on the market by a wide margin - roughly 130 genuine agents against thousands of imitators wearing the same word. The businesses that end up in Gartner’s 40% aren’t the ones who bet on agentic AI. They’re the ones who bought a chatbot that borrowed its name.

Ask the four questions before the contract, not after the cancellation.

Frequently Asked Questions

What is agent washing?

Agent washing is Gartner’s term for rebranding an existing chatbot, robotic process automation script, or AI assistant as an “AI agent” with no meaningful increase in autonomy. The product looks the same. Only the marketing copy has changed.

How many AI agent vendors are actually genuine, according to Gartner?

Gartner estimated in June 2025 that of the thousands of vendors marketing “agentic AI”, only around 130 met its definition of a genuine agent. The rest were, to varying degrees, agent washing.

Why will 40% of agentic AI projects be cancelled by 2027?

Gartner attributes the coming cancellations to escalating costs, unclear business value, and inadequate risk controls - not to the technology failing outright. Most of the cancelled projects were pilots built on a washed product, an undefined success metric, or both.

What is the difference between an AI agent, a chatbot, and RPA?

RPA executes a fixed sequence of steps and cannot deviate from it. A chatbot responds to a prompt and waits for the next one. An agent is given a goal, then plans its own route to it - selecting tools, adapting when a step fails, and completing multi-step work with a human checkpoint rather than a human keystroke at every stage.

How do I test whether an AI agent is genuine before I buy it?

Ask the vendor to show you a decision trace for a task the agent completed without a human writing every step in advance, ask what percentage of production outputs still require human correction, ask what happens when a tool call fails mid-task, and ask for an audit log you can read yourself rather than a dashboard they curate for you.

Does this mean businesses should avoid agentic AI altogether?

No. The technology behind genuine agents is real and already delivering measurable results in customer service, revenue operations, and finance. The problem isn’t agentic AI - it’s the volume of non-agentic products wearing its name. Vet the vendor, not the trend.

References

• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 - Gartner, 25 June 2025 (source for the 40% cancellation prediction, the ~130-genuine-vendor estimate, and the “agent washing” term)
• 2026 Hype Cycle for Agentic AI - Gartner, spring 2026 (agentic AI placed at the Peak of Inflated Expectations; 17% deployed vs 60%+ expecting to deploy within two years, per the 2026 Gartner CIO and Technology Executive Survey)
• Gartner Survey Finds 45% of Martech Leaders Say Vendor-Offered AI Agents Fail to Meet Expectations - Gartner, October 2025
• The state of AI in 2025: Agents, innovation, and transformation - McKinsey, 2025 (6% of organisations as AI “high performers”, 88% overall adoption)
• Why 40% of Agentic AI Projects May Be Canceled By 2027 - Forbes, 7 July 2026

Marc Price is the founder of Aandai, a B2B automation and AI consultancy helping mid-market businesses achieve more with less. With 24+ years in B2B technology marketing and web development, Marc specialises in connecting legacy systems, eliminating manual processes, and implementing practical AI solutions that deliver measurable ROI. Aandai runs its own agentic stack on OpenClaw to automate parts of its consultancy delivery - including the research that informed this article, and the vetting questions in it.