Agent washing is selling a chatbot or script as an AI agent. Gartner estimates only about 130 of thousands of agentic AI vendors offer real agentic features. Seven on-screen tests, the FTC and SEC cases so far, Massachusetts law, and what a year and a half of running agents taught us.
Agent washing is a vendor renaming a chatbot, an assistant or an old automation script as an "AI agent" without giving it any real ability to act. Gartner, which named the trend, estimates that only around 130 of the thousands of vendors claiming agentic solutions offer real agentic features. The rest is packaging, and agent washing can be spotted before you sign.
On October 23, Brian Gagne ran a LinkedIn Live on what an agentic business operating system is, using the one we built and run every day. In the days after, the lookalikes started showing up in my feed: a chat window with a new name, a thin layer over someone else's agent, a demo with a stopwatch in the corner. Brian called the content version of this "LinkedIn slop" on the webinar. The software version is arriving now.
We have run LTFI, our agentic business operating system, in live production at Kief Studio for a year and a half. That is the lens for this post: what a working agent does on an ordinary Tuesday, and what agent washing hides. The companion piece, questions to ask an agentic AI provider, turns it into a buyer's checklist.
Agent washing is selling a product as an AI agent when it cannot take a job and finish it on its own within limits a person set. Gartner describes it as vendors rebranding existing chatbots and automation tools as agentic AI without delivering genuine autonomous capabilities.
The product underneath may still be fine at what it does. A good chatbot answers questions well. A good automation script moves a file reliably. The problem is the label, because "agent" tells a buyer the software can be handed a job, work through several steps, use tools, stop where it is told to stop, and leave a record of what it did. When those parts are missing, the buyer has paid for staff and received a form.
A working definition helps in a sales meeting. An AI agent, for a business, has five things: its own identity, a job it was assigned, tools it may use, limits on what it may do without a person, and a log. A product missing two or more of those is a candidate for agent washing, however good its demo looks.
Very common, by Gartner's count: only around 130 of the thousands of vendors claiming agentic solutions offer real agentic features. MarTech reported that estimate, alongside Gartner's cancellation forecast, on April 29, 2026.
The pressure to relabel is easy to see in the demand numbers. Gartner's 2026 CIO and Technology Executive Survey, quoted in its first Hype Cycle for Agentic AI in April 2026, found that only 17% of organizations have deployed AI agents so far, but 42% expect to do so in the next 12 months. When that many buyers are about to shop at once, a new label is the cheapest product update a vendor can ship.
The same Gartner forecast predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. Some will be honest experiments that ended. Others will be companies that bought a chat window and expected staff, which is the cost of agent washing landing on the buyer's books.
Regulators have already priced a few cases, and the pattern is consistent: the claim was about work the software did without a person, and the work was being done some other way or not at all.
None of these cases used the phrase agent washing, and each was a different kind of claim. What they share is autonomy that was sold and not delivered. That is the thing a buyer can test, and in my experience the answer shows up in the first hour of a real trial on your own work.
Ask the vendor to show seven things on screen, on a job you chose, and watch which ones they can click to and which ones they describe. Agent washing shows up as a roadmap where a working product shows a screen.
| What the pitch says | What to ask to see | What agent washing usually shows |
|---|---|---|
| "It works on its own" | What it does when nobody is typing | A chat box waiting for a prompt |
| "It acts for you" | The agent's own identity and permissions | Your login, shared with the software |
| "You stay in control" | The list of actions that stop for a person's yes, and who sets it | "You can review anything" |
| "Full transparency" | A log of who asked, which agent acted, which tools and who approved | A chat transcript |
| "Replaces a role" | The share of jobs finished without a person stepping in | No number, or a number nobody can source |
| "Our own AI" | Which model runs each job, and who maintains the layer underneath | A thin wrapper with a logo |
| "Safe by design" | The control that pauses every agent at once | A support ticket |
Two rows deserve more time. The identity row matters because an agent that signs in as you is invisible in every log you own, which is the case made in a non-human identity for an AI agent. The success-rate row matters because the Nate case turned on that number being hidden from the people paying for it.
The "our own AI" row overlaps with the general advice in how to evaluate AI tools without getting sold. For agents it goes one step further. Some products are a thin wrapper around a free open-source agent framework or someone else's chat window, so ask what happens to your workflow when that layer changes its behavior or its terms.
A vendor with a real system answers these quickly, usually by clicking somewhere. The gap between a click and a slide is the same gap described in AI agents that run at 3am versus AI agents that demo well.
It looks quieter than the demos, which is one reason agent washing works: the real thing is less theatrical. Most of what our agents do all day is drafts, research, summaries and keeping records current, and the moments that matter are the ones where an agent stops and waits for one of us.
Brian showed one of those moments live on the webinar. He asked the chief of staff what needed him that day, and it came back with leads that had no next step. He told it to email them. The outreach teammate pulled the context on each contact and drafted the messages, and then the system stopped: "Emails will wait for your approval before they go out." He read what would be sent, approved it, and the emails went. His line afterward was the honest summary of the value: "I approved it. I didn't have to write it."
The second thing a year and a half teaches is that agents are eager. Brian put it plainly on the call: agents "are very, very eager to complete their task, you also need to tell them to chill out a little bit every now and then." A product built on agent washing has nothing to tell to chill out, because it never starts anything. A real one needs approval lines, written company rules and a pause, and it needs them before the first job, not after the first mistake.
The third is the log, and it is the clearest line between a working agent and agent washing. I use it more than I expected to. When Brian had shipped a run of updates and I was deep in something else, I asked the system what had happened on that project, and it gave me the summary from the work itself, including the LinkedIn event, which was already in the project's history. The same log is how we tune the work: when a draft misses, the record shows which teammate, which rule and which step to fix. Brian's phrase for it was that the log is "for more than just proof of work."
Because a convincing demo has become cheap, and the governance around an agent has not. Anyone can wire a free open-source agent framework to a chat window in an afternoon; identity, permissions, approval lines, a readable log, a pause and customer-owned data take years.
Open-source frameworks are good work, and plenty of serious products use them. The trouble is a product that stops at the wrapper and sells the result as a governed business system, which is agent washing by another route. That governance layer is the part a lookalike leaves out, which I cover in why AI governance is not a wrapper.
Watch for urgency, too. A launch price that expires Friday says something about the sales plan and nothing about the product, and it works best on buyers who have not yet asked the seven questions above. Why to slow down when a technology deal comes with a deadline covers that pattern.
A demo also hides time. Researchers at Stanford and Berkeley found GPT-4's accuracy on the same prime-number questions fell from 84% in March 2023 to 51% in June 2023 (Chen, Zaharia and Zou, revised October 2023), so ask how a product notices that kind of change. How to catch AI agent drift covers the rest.
Massachusetts has no statute that names agent washing, but the Attorney General has said existing consumer protection law covers false AI claims. For businesses in Worcester and across Central Massachusetts, that gives a state answer alongside the federal one.
Attorney General Andrea Joy Campbell's April 16, 2024 advisory lists among practices that may be unfair or deceptive "Falsely advertising the quality, value or usability of AI systems" and "Misrepresenting the reliability, manner of performance, safety or condition of an AI system."
That makes the paper trail matter when you suspect agent washing. Keep the proposal, the demo recording and the written answers to the questions in the table above. If an "agent" turns out to be a person in a call center or a chat window with a new name, those documents are what a Chapter 93A technology vendor dispute runs on. The full state picture is in Massachusetts AI regulations for business. This is general information, not legal advice.
By answering every row of the table with a screen, because it was built as a whole system rather than a label. LTFI is an agentic business operating system, meticulously built from scratch by Kief Studio: the CRM, projects, content, sites, security and AI staff in one system, on a dedicated server we run for each company, with no shared hosting.
It works with the AI tools a team already uses, including Claude Code, Codex, Cursor, Gemini, GitHub Copilot, Grok and OpenCode, and connects to models from several providers or a company's own local model. Your data stays yours, and if you leave, we export it.
If you are comparing agents right now and want to rule out agent washing, the October 23 webinar replay shows the approval line and the log in a working system, and ltfi.ai has the walkthrough. We set LTFI up remotely for businesses anywhere, the way we have worked with clients around the world for more than 13 years, and we meet teams in Worcester, Shrewsbury and the surrounding towns in person.
AI washing is claiming to use AI when you do not. Agent washing is narrower: the product may use AI, but it is sold as an agent that takes actions and finishes jobs when it is really a chatbot, an assistant or a scripted automation.
Not by itself. A chatbot answers when asked. An agent takes an assigned job, works through steps with tools, acts inside limits a person set and leaves a record of what it did.
Give the product one real, low-risk job from your own work, not the vendor's sample data. Watch what it does without prompting, what stops for your approval, what the log shows afterward, and how you pause it.
Overstated AI claims have led to FTC and SEC actions, and the Massachusetts Attorney General has said false claims about AI systems may be unfair or deceptive under state consumer protection law. Whether a specific claim crosses the line is a question for a lawyer.
No. Open-source frameworks are legitimate building blocks. The concern is a product that is only a thin wrapper, sold as a governed business system without identity, approvals, logging, a pause and data ownership around it.
The proposal, any demo recording, and written answers about identity, approvals, logging, the pause and data export. Those documents are the record of what was promised if the product turns out to be less than advertised.
The questions to ask an agentic AI provider are about control: where it runs, whose keys it holds, what waits for a yes, what is logged, which models work, how you pause it and how you leave. IBM's 2026 research found 92% of organizations with an AI-related breach lacked proper AI access controls.
Only 13% of organizations think they have the right AI agent governance in place, Gartner reported in April 2026. AI governance for agentic AI is six controls built into where agents run: identity, permissions, approvals, audit, data ownership and model choice.
AI agent drift is the slow change in what an agent does after launch. Researchers found one model's prime number accuracy fell from 84% in March 2023 to 51% in June 2023 on the same questions. The four causes, five checks that catch it, and how we handle it in LTFI.
Work With Us
Kief Studio builds, protects, automates, and supports full-stack systems for businesses up to $50M ARR.
Newsletter
Strategy, psychology, AI adoption, and the patterns that actually compound. No spam, easy to leave.
Subscribe
