如何篩查 AI 智能體項目是否只是 Agent washing

Two guests on Chain of Thought, six weeks apart, described the same market in opposite words. Kelly Vaughn of Spot AI, on episode 15 (March 19, 2025), calls agentic AI a bubble and questions the hype around "agentic AI for x," including the pile of AI notetakers. Siva Surendira of Lyzr, on episode 21 (April 30, 2025), says adoption blew past what he expected and that agents have gone mainstream the way cloud and databases did. The debates archive sets those positions side by side. I would not pick a side and then green-light a project. I would use both claims as a screen for agent washing.

What are the two claims actually saying?

Vaughn is talking about pitches and startups. She argues we are on the hype train of agentic AI for whatever job someone wants to fund. She also argues that noise, weak differentiators, and a focus on being first to market produce a lot of failed companies. She is explicit that agentic AI can still be useful for a specific customer or team problem, and that existing companies can apply it because they already have customers and trust.

Surendira is talking about adoption. He says he always saw agents taking repeated tasks that people do in daily work, and that he did not expect the blow-up that followed earlier hype cycles such as RAG and fine tuning.

I never thought it'll have this level of adoption, but but cannot complain because we were one of the early movers and we are sitting pretty in this whole agent train so called.

Siva Surendira, Lyzr, on Chain of Thought ep 21

Those statements do not cancel each other. A tool category can be in wide use while a wave of named "agent" projects has no substantial agentic capability. Widespread adoption can coexist with overhyped pitches.

Which public figures should I put next to the proposal?

Put two outside sources on the table before you argue from a demo.

In June 2025, Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. It called most current projects early experiments driven by hype. It estimated that only about 130 of the thousands of vendors marketing agentic AI offer real agentic capabilities. It named the failure mode: rebranding existing assistants, robotic process automation, and chatbots without substantial agentic capabilities. That is "agent washing." (Gartner, June 25, 2025.)

Usage data points the other way, toward Surendira's adoption claim. Microsoft and LinkedIn's 2024 Work Trend Index found that 75% of knowledge workers were using generative AI at work. Heavy use of AI can sit beside a wave of agent projects and vendors that do not survive.

Read those numbers as a filter, not as a forecast to quote in a pitch. A predicted cancellation rate above 40% means a proposal needs a reason to be in the surviving set, and broad AI usage is not that reason.

How do I screen one proposal?

I'd start with four questions. Write the answers in plain language. If you need the word "agent" to make the answer sound real, the answer is not ready.

1. Strip the label. Is this a renamed assistant, a chatbot, or an RPA script without substantial agentic capabilities? If yes, Gartner's term applies. Drop it or rename it honestly.
2. Name the agentic part. What decision does it make, and what action does it take, that the old workflow did not? "Real agentic capabilities" is the bar Gartner used when it said only about 130 vendors clear it. A wrapper that calls a model and posts the text does not clear it.
3. Attach the three cancellation causes. What is the business value in one sentence? What cost bound will you stop at? What risk control stops a bad action? Escalating costs, unclear business value, and inadequate risk controls are why Gartner expects so many cancellations.
4. Match the job to a repeated task a person already does, and say what "used correctly" changes. Vaughn argues AI can be a good augmentation tool, with efficiency gains if used correctly, and that it will not replace a team or double its size. Surendira argues the opening is mundane, repeated work. If the proposal claims a doubled team, treat that as a failed screen.

Surendira argues that 95% of all AI work enterprises do today stays a proof of concept and does not move into production. He points to fear of being sued, and to errors such as approving a $100,000 refund when the amount should have been $10, along with hallucination, toxicity, and prompt injection. He also names two bottlenecks: custom applications with no documentation, where someone still has to define the actions an agent may take, and tribal data that is scattered and not ready for a model.

Vaughn's version of the same gate is the customer problem. She argues you should ask what you are solving and whether AI helps. She says plenty of companies do not need AI for their problems. She treats overpromising as the biggest pitfall, and she says you are never going to see 100% accuracy. On a security or safety action, a miss (her example is failing to detect a break-in) is how you lose trust. This sketch only encodes that screen. It is not a scorer anyone on those episodes runs.

# Sketch. Illustrates the screen. Not a production policy.

def screen(p: dict) -> str:
    washed = p["rebranded_assistant_rpa_or_chatbot"]
    agentic = p["chooses_and_takes_a_real_action"]
    value = bool(p["business_value"].strip())
    cost_ok = p["cost_bound"] is not None
    risk_ok = bool(p["risk_control"].strip())
    repeated = p["replaces_a_repeated_human_task"]
    data_ok = p["needs_little_internal_data"] or p["data_and_actions_defined"]
    if washed or not agentic:
        return "drop"  # agent washing, per Gartner's June 2025 wording
    if not (value and cost_ok and risk_ok and repeated and data_ok):
        return "narrow"  # costs, value, or risk controls are the cancellation causes
    return "worth a bounded experiment"

A pass is not a reason to skip user conversations. Vaughn argues you should talk to potential customers before you build, including where data is stored and who can access it.

What should I do with a pass?

Keep the project narrower than the pitch. Vaughn's notetaker riff is the pattern I would watch for: a meeting with a dozen people and several different AI notetakers, and then someone building "the AI notetaker for all AI notetakers." If your differentiator is an agent for a crowded job, you do not have one yet.

If the pass depends on the agent touching money, safety, or a system of record, add the risk control to the design before the demo, not after a pilot. Surendira argues that combining managerial and DAG orchestration still did not solve critical use cases, and that adding code, SQL, machine learning and deep learning agents is what improved determinism. You do not need his stack to use the rule. If you cannot name the allowed action and the check that blocks the wrong one, the project is still a pitch.

For more conversations on how agents get built, deployed, and evaluated, the best AI agent podcast episodes page collects picks from Chain of Thought and other shows.

Checklist

• Write the job with no agent label. If it is still an assistant, a chatbot, or RPA with no substantial agentic capability, call it agent washing and stop.
• Require one sentence of business value, a cost bound, and a named risk control. Those are the cancellation causes in the June 2025 Gartner release.
• Prefer a repeated task a person already does, with an augmentation claim you can level-set. Reject a claim that the team doubles or becomes obsolete.
• If the work needs old internal apps or scattered data, define the actions and the data first, or pick a job that needs little internal data.
• Treat a pass as a bounded experiment with a feedback loop. Do not promise 100% accuracy, and do not read mainstream AI use as proof this project will survive.

The longer explainer, with the episode clips, is on Chain of Thought. It draws on this episode.

Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.

Drafted with AI assistance from the episode transcripts.