In the fast-paced world of enterprise technology, history has a tendency to repeat itself. Anand Naidu, a seasoned expert in technology strategy and AI risk management, joins us to discuss a troubling new trend he calls “agentwashing.” Drawing parallels to the “cloudwashing” era that left many companies mired in renamed technical debt, Anand warns that the stakes are even higher with artificial intelligence. In our conversation, we’ll explore the critical distinction between true AI agents and cleverly marketed automations, the serious strategic risks of mislabeling this technology, and the rigorous governance required to separate transformative potential from expensive hype. Anand provides a clear-eyed guide for leaders on how to demand technical evidence, write smarter contracts, and ultimately make sure they are buying genuine capability, not just a compelling story.
You draw a powerful parallel between today’s “agentwashing” and the “cloudwashing” era. Beyond misspent funds, what are the most damaging strategic consequences you’ve seen when a company mistakenly adopts a souped-up chatbot, believing it to be a truly autonomous agent?
The financial waste is what gets the headlines, but the real damage is the loss of time and trust. It’s a deeply frustrating experience. I’ve seen leadership teams that genuinely believed they were making a strategic leap, investing millions in what was pitched as a transformative agentic platform. A year later, they find themselves with nothing more than a brittle, hardcoded workflow that breaks at the first sign of an unexpected input. The strategic consequence is that they’ve lost a year they can never get back, while their competitors who were more discerning are now lapping them. It also erodes the board’s confidence in the IT organization’s ability to vet new technology. They end up with a high-maintenance system, operational overhead that wasn’t in the plan, and a profound sense of having been misled, all while their core architecture hasn’t modernized one bit.
The content outlines four key characteristics of a true agent, including autonomous goal pursuit and adaptability. When evaluating vendors, what specific technical evidence and failure-mode documentation should leaders demand to distinguish a real agent from a system that just uses prompt templates and scripts?
This is where you have to get beyond the slick demos. A polished demo is designed to work perfectly; I always ask to see what happens when things go wrong. First, demand detailed architecture diagrams. If a vendor can’t explain precisely how their system plans a sequence of actions, executes them, and adjusts to feedback without defaulting to a human, that’s a major red flag. Second, ask for their evaluation methods and failure-mode documentation. Don’t just ask “Does it work?”—ask “How does it fail, and what happens when it does?” A true agentic system will have a documented process for adapting or recovering from unexpected conditions. A scripted system just fails. I want to see the logs and the limitations, not just the marketing slides. A vendor who is honest about their system’s boundaries is far more trustworthy than one who claims it can do everything autonomously.
It’s mentioned that some vendors describe a single LLM call as a “society of cooperating agents.” Can you walk us through a step-by-step due diligence process to uncover this misrepresentation? What are some common red flags you’ve seen in architecture diagrams or vendor demos?
Absolutely. The “society of agents” line is one of the most common and misleading phrases out there right now. My due diligence process is simple: I ask them to trace the data and decision-making flow for a complex, multistep task. I’ll say, “Show me exactly, step-by-step, how the ‘agents’ delegate, reason, and adapt when the first API call fails.” A huge red flag is when their explanation gets vague, falling back on words like “reasoning” or “autonomy” without technical substance. If you strip away the branding and their architecture diagram boils down to a single LLM call with some simple glue code around it, you’ve found agentwashing. Often, what they call “cooperating agents” is just a well-written prompt template that asks the LLM to structure its output in a certain way. It’s a clever trick, but it’s not a dynamic, adaptable system.
The text advises treating agentwashing as a “fraud-level governance problem.” What specific, measurable success criteria should be written into contracts to hold vendors accountable for claims of autonomy, and how can risk teams audit these systems to ensure they’re not just brittle workflows?
You have to move from vague promises to concrete, measurable outcomes in your contracts. Instead of a contract that says you’re buying an “autonomous agent platform,” it should specify, for example, “The system will process X type of invoice with a 95% success rate without human intervention, even when formats vary by up to 20%.” You tie the vendor’s claims directly to quantifiable improvements and explicit autonomy levels. For auditing, risk teams need to go on the offensive. This means stress-testing the system with edge cases and unexpected inputs, not just the standard examples from the demo. If the system fails outright instead of adapting or gracefully requesting help, you’ve proven it’s not truly agentic. By treating it like a financial representation, you apply the same rigor—you demand proof and refuse to fund it until you see it.
What is your forecast for the agentic AI market? How will this “agentwashing” trend likely play out, and what will ultimately separate the truly transformative platforms from the marketing hype over the next few years?
I believe we’re in for a period of disillusionment, much like we saw after the initial cloud hype. Many of the companies currently engaging in agentwashing will either pivot or fail once customers realize the platforms are brittle and don’t deliver on their promises of autonomy. The survivors—the truly transformative platforms—will be the ones who are honest and precise from the start. They won’t promise full autonomy on day one. Instead, they will offer powerful, supervised automation with very clear guardrails and use cases. They will build trust by being transparent about their systems’ limitations and providing a realistic road map toward greater autonomy. Ultimately, the market will be won by those who deliver tangible, reliable results, not by those who spin the most compelling narrative. The enterprises that succeed will be the ones that insisted on that technical honesty from the very beginning.
