What is agent washing?
Agent washing is the practice of relabelling existing software as an "AI agent" when nothing underneath has actually changed. Gartner defines it as the rebranding of existing products, such as AI assistants, robotic process automation and chatbots, without substantial agentic capabilities. The label changes. The product does not.
The word matters because the price usually goes up with it. A chatbot renamed an agent is often quoted at agent money.
Gartner put a number on how common this is. In its 25 June 2025 press release, it estimated that only about 130 of the thousands of agentic AI vendors are real.
Read that again. Thousands of companies sell agents. Roughly 130 of them build them.
How does agent washing actually happen?
Agent washing happens at the marketing layer, not the engineering layer. A vendor takes a product that already works, swaps the noun on the website, and reprices it. Three older product categories get relabelled most often.
The three products most commonly renamed
---
Chatbots. A scripted question-and-answer flow becomes a "customer service agent". It still follows the same decision tree it followed in 2023.
Robotic process automation. A macro that copies rows between two systems becomes an "operations agent". It still breaks the moment a column moves.
AI assistants. A chat window that answers questions about your data becomes a "finance agent". It still cannot do anything unless you type a question first.
---
None of those three products is bad. All three are useful. The problem is only the label, and the invoice attached to the label.
There is a local tell worth knowing. In Hong Kong, a relabelled product usually arrives with a demo recorded somewhere else. The screen shows a US or Australian business, the sample invoices are in a foreign currency, and the sample customer messages are in English only. That is not proof of agent washing on its own, but it tells you the product has not been tested against the way your customers actually write to you, which is a mix of Chinese and English in the same sentence.
The second tell is a demo that never fails. A vendor who has built something real will show you where it breaks, because they have watched it break for two years. A vendor who has renamed something will show you a recording.
Gartner returned to this in a 20 May 2026 press release on supply chain planning software, warning that relabelling conventional automation as agentic increases the risk of misaligned investments and long-term lock-in. Lock-in is the part small businesses feel most. A twelve-month contract signed on the wrong understanding is twelve months of paying for a category you did not need.
Assistant, automation or agent: what is the actual difference?
The three categories differ by who decides the next step. An assistant waits for you to ask. An automation follows a route you drew in advance. An agent is given a goal and works out its own route, then adapts when the route fails.
Gartner's own guidance to buyers is a clean triage rule: use agents when decisions are needed, automation for routine workflows, and assistants for simple retrieval.
Apply that to a Hong Kong shop with a single trading account.
---
Simple retrieval. "How much did I spend on packaging last month?" That is an assistant. You ask, it answers, it stops.
Routine workflow. Every time a payment lands, file the receipt in the right folder and update the sheet. That is automation. The steps never change, so nothing needs to think.
A decision. A supplier is thirty days late, three orders are affected, and someone has to decide which customer gets told what, in which order, and whether to switch supplier. That is agent territory, because the correct sequence of steps is not knowable in advance.
---
Here is the sentence that saves the most money. If the sequence of steps is knowable in advance, you do not need an agent, and paying agent prices for it is waste. Anushree Verma, the Gartner analyst behind the June 2025 research, said it plainly: many use cases positioned as agentic today do not require agentic implementations.
If you want the fuller picture of what falls into the middle category, our guide to AI workflow automation covers it.
Why does agent washing matter to a Hong Kong small business?
It matters because Hong Kong SMEs are buying right now, and buying fast, which is exactly the market condition agent washing feeds on. Vendors sell hardest into the group with the least time to check.
HKT's 2026 survey, reported by Telecom Review Asia, found that 67% of respondents have implemented, are piloting, or plan to use AI, split as 79% among large enterprises, 65% among mid-market businesses and 49% among SMEs.
Roughly half of Hong Kong's small businesses are therefore in the market. Most of them are buying their first AI product, which means they have no previous purchase to compare a quote against.
The failure rate is already visible at the top end. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Those are enterprises with procurement teams. A five-person office has none.
A newer figure sharpens it. On 1 September 2026 Gartner published survey work covering 1,303 respondents at organisations with at least US$50 million in FY2025 revenue, finding that only 22% had successfully scaled AI across multiple business units while 85% of functional leaders planned to increase AI spending in 2026. Spending is rising faster than success.
The practical risk for an SME boss is narrower and more concrete than any of those statistics. It is signing a year of software for a job a HK$0 automation could have done, and discovering it in month four.
How can you tell a real AI agent from a rebranded one?
You test it, using your own messy data, in the free trial. Five questions separate a real agent from a relabelled one, and none of them requires a technical background to ask. A vendor who cannot answer all five is selling you the label.
The five questions to ask before you sign
---
1. Show me the trail. Ask the vendor to display what the product did, step by step: which tool it opened, what came back, and why it chose the next step. A real agent can produce that trail because it generated one. A relabelled chatbot cannot, because there was never a decision to record.
2. What happens when the input is wrong? Hand it a misspelled supplier name, a photo of a receipt taken at an angle, or an instruction with a step missing. Demo-grade products only survive clean inputs. Your business does not produce clean inputs.
3. Does it stop when it is unsure? A product that never asks a question is not confident, it is unsupervised. Ask to see what happens when it hits something ambiguous, and who gets told.
4. What can it actually reach? An agent that cannot open your accounting system, your inbox and your order sheet cannot complete a real task in your business. Ask for the exact list of systems it connects to in Hong Kong, not the global list.
5. What does month thirteen cost? Ask for the price after the introductory discount ends, the notice period for cancelling, and whether an annual prepayment is refundable. Agent washing and aggressive contract terms travel together.
---
One more test costs nothing. Ask the vendor to name a customer of a similar size in Hong Kong. Not a global logo. A shop, a clinic, a trading firm with a headcount close to yours.
Once a product is inside your business, the question shifts from marketing to measurement, which is the subject of our guide to AI evals.
What do buyers most often get wrong about AI agents?
Three misconceptions do most of the damage. Each one is understandable, and each one leads to overpaying for the wrong category of software.
Misconception one: agents are simply better, so buy the agent. They are not better, they are different. An agent is the right tool only when the steps cannot be written down in advance. For everything else it is a more expensive way to do what a rule already does.
Misconception two: real agents run without people. No serious product on the market today runs a business function unattended. A vendor claiming full autonomy is describing an ambition, not a shipping feature. Gartner's own forecast is that roughly 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. That is a small share, four years out.
Misconception three: if it uses AI, it is agentic. Almost all business software now contains some machine learning. Categorising a bank transaction using a model is AI. It is not an agent. The presence of AI in a product tells you nothing about whether it can pursue a goal.
There is a fourth misconception that only bites later. Buyers assume the label is stable. It is not. Gartner expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024, which means much of what is honestly called an agent in two years is being called an assistant today by a more careful vendor.
The bottom line for a Hong Kong SME boss
Agent washing is not a technology problem. It is a vocabulary problem that turns into an invoice problem.
The defence is not technical literacy. It is one habit: describe the job before you look at the product. Write down, in one sentence, the task you want done and whether its steps are knowable in advance. If they are, you need automation or an assistant, and you should refuse to pay agent prices. If they are not, you are in genuine agent territory, and the five questions above will tell you whether the vendor has actually built one.
Not knowing the difference is not a failing. Nobody handed Hong Kong business owners a glossary before the sales calls started. We understand AI. UD stands with you. Twenty-eight years of sitting on the buyer's side of the table is what turns a confusing category into a decision you can make in an afternoon.
Reviewed by the UD AI team, Hong Kong, 2 September 2026.
Ready to work out what you actually need?
Before you compare vendors, it helps to know which jobs in your business are genuinely worth automating and which are not. UD's free AI Ready Check takes about three minutes, and if you want to go further, we will walk you through every step from assessing your needs to going live.
