Agent Washing: The Billion-Dollar Lie Destroying Enterprise AI
Is it true that only 130 of thousands of "agentic AI" vendors are genuine?
The gap between the promise of agentic AI and the reality on the ground has never been wider.
While venture capital poured a staggering 265% more funding into agentic AI between Q4 2024 and Q1 2025, and 82% of companies claim they’ll integrate AI agents by 2027, the sobering truth is that Gartner predicts over 40% of these projects will be cancelled by the end of 2027.
Even more damning, Andrej Karpathy, the OpenAI cofounder who helped build some of the most advanced AI systems on the planet, recently stated that functional AI agents are still about a decade away.
This isn’t just a minor disconnect; it’s a chasm that threatens to waste billions in capital and erode trust in a technology that, properly applied, could genuinely transform how we work.
The question is not whether AI agents have potential, but whether the industry can get its act together before the hype cycle collapses under the weight of its own delusions.
Is the state of the AI agent market in 2025 really a house of cards?
Let’s find out.
Is the Market Built on Sand?
The issues plaguing agentic AI deployment are not subtle.
Escalating costs
Unclear business value - this is at epidemic levels right now.
Inadequate risk controls
But these are symptoms of a deeper malaise.
The fundamental problem is a catastrophic misalignment between what the technology can actually do and what vendors, consultants, and overeager executives claim it can do.
Gartner estimates that of the thousands of vendors claiming to offer “agentic AI,” only about 130 are genuinely delivering autonomous systems.
The rest are engaged in what the industry has started calling “agent washing”—rebranding chatbots, robotic process automation tools, and glorified workflow engines as cutting-edge agentic systems.
This isn’t just marketing spin; it’s actively misleading customers into deploying solutions that cannot possibly deliver the promised outcomes.
When a company invests millions into what they believe is an autonomous agent capable of complex decision-making, only to discover they’ve bought an expensive chatbot with a conversational interface, the result is predictable: the project gets cancelled, trust evaporates, and the next genuinely useful AI initiative faces an uphill battle for funding.
The technical reality, as Karpathy points out, is that current AI agents are “cognitively lacking.”
I wrote extensively about this a few months back. 👇
Most agent systems in the market today cannot be given something and expected to remember it in a meaningful way.
These aren’t minor limitations; they’re fundamental constraints that make the autonomous, goal-directed behaviour promised by the “agentic” label largely unattainable with current technology.
The mathematics are equally unforgiving.
ScaleAI’s analysis shows that with a 20% error rate per action—a generous estimate for current LLMs—an agent attempting a five-step task has only a 32% chance of getting everything right.
This compounds exponentially as tasks become more complex, making truly autonomous operation a statistical impossibility for anything beyond trivial workflows.
The integration challenges are equally problematic.
Legacy systems, data silos, and the need for extensive process redesign mean that even when the technology works as advertised, the cost and complexity of deployment often exceed initial estimates by orders of magnitude.
Organisations underestimate the sophistication required to deploy genuine AI agents, leading to budget overruns, missed timelines, and ultimately, project cancellation.
Are the Critics Pessimistic or Realistic?
Karpathy’s decade-long timeline isn’t plucked from thin air. It reflects a sober assessment of the technical challenges that need to be solved before AI agents can reliably operate autonomously in complex, real-world environments.
The industry’s current trajectory—overshooting the tooling relative to present capability—creates a dangerous illusion of progress. We’re building orchestration frameworks, deployment platforms, and governance structures for a level of AI autonomy that simply doesn’t exist yet.
The “agent washing” phenomenon is particularly insidious because it creates a self-reinforcing cycle of failure.
Vendors overpromise, customers over-invest, projects under-deliver, and the resulting disappointment taints the entire category. This isn’t just bad for individual companies; it’s bad for the entire AI ecosystem.
When 40% of projects fail, the lesson learned isn’t “we need better agents,” it’s “AI agents don’t work.” That perception, once it takes hold, is extraordinarily difficult to shift, even when the underlying technology eventually matures.
Finally, it’s important to call out that the level of transformation with any generative AI ambition is expensive, time-consuming, and organisationally disruptive. Many companies simply aren’t prepared for it, and no amount of hype can compensate for that lack of readiness.
What Can Be Done - A Disciplined Approach
The path forward requires discipline, realism, and a willingness to resist the hype.
The first step is honest use-case selection.
Not every problem requires agentic AI, and in many cases, traditional automation, machine learning, or even well-designed software will deliver better outcomes at a fraction of the cost.
Agentic AI excels in complex, dynamic environments where tasks require ongoing adaptation, decision-making under uncertainty, and the ability to handle exceptions.
Supply-chain optimisation, cybersecurity threat response, and dynamic credit approvals are good candidates.
Simple order tracking, basic code generation, or repetitive data entry are not.
The key is to focus on scenarios where the unique capabilities of agentic AI—the ability to plan, adapt, and act autonomously across multiple steps—create measurable business value that cannot be achieved through simpler means.
This requires a rigorous cost-benefit analysis that accounts not just for the direct costs of the technology, but also the indirect costs of integration, change management, and ongoing support. If the business case doesn’t clearly justify the investment, don’t do it.
Second, organisations must adopt a “composite AI” approach. Rather than betting everything on a single agentic system, combine multiple AI techniques—machine learning, symbolic reasoning, traditional automation—to address specific pain points.
This increases accuracy, transparency, and performance while reducing costs and data requirements. It also provides a more graceful degradation path: if the agentic component underperforms, the rest of the system can still deliver value.
Third, organisations need to invest in readiness.
This means ensuring data is AI-ready, workflows are sufficiently mature and documented, and risk controls are in place before deployment. Attempting to deploy agentic AI into a chaotic, undocumented environment is a recipe for failure. The technology cannot compensate for organisational dysfunction.
Finally, organisations must focus on enterprise productivity, not just user productivity. The greatest value from agentic AI comes from orchestrating actions across siloed applications and business units, not simply improving individual user experiences.
This requires thinking beyond departmental boundaries and reimagining workflows at an enterprise level. It’s harder, but it’s where the real value lies.
Four Key Lessons
Lesson One: Resist the Hype, Demand Evidence
The agentic AI market is rife with exaggeration. Before committing to any solution, demand concrete evidence of autonomous capability.
Ask for demonstrations of multi-step workflows, error handling, and adaptation to unexpected conditions. If the vendor can’t show you the agent genuinely planning and acting autonomously - it’s not really an agent.
Insist on proof of concept deployments with measurable outcomes before scaling. The burden of proof is on the vendor, not on you.
Lesson Two: Start Small, Measure Ruthlessly
The temptation to deploy agentic AI across the enterprise is strong, but it’s almost always a mistake.
Start with a single, well-defined use case where the business value is clear and measurable. Deploy, measure, iterate, and only then consider scaling. Establish clear KPIs upfront—cost savings, time reduction, error rates, customer satisfaction—and track them religiously.
If the pilot doesn’t deliver measurable value, don’t scale.
This sounds obvious, but the number of organisations that skip this step and go straight to enterprise-wide deployment is staggering.
Lesson Three: Invest in Foundations, Not Just Tools
Agentic AI is not a silver bullet that can be dropped into a dysfunctional organisation and expected to work. It requires solid foundations: clean, accessible data; well-documented processes; clear governance structures; and organisational readiness for change.
Investing in these foundations is less exciting than buying the latest AI platform, but it’s far more important. Without them, even the best agentic AI will fail.
This means investing in data infrastructure, process documentation, change management, and training before you invest in the AI itself.
Lesson Four: Embrace Human-AI Collaboration, Not Replacement
Karpathy’s vision of humans and AI collaborating, rather than AI rendering humans obsolete, is not just more humane; it’s more practical.
The most successful AI deployments are those that augment human capabilities, not replace them.
Design your agentic systems to work alongside humans, not instead of them. Build in checkpoints for human review, create interfaces that make the agent’s reasoning transparent, and ensure humans can intervene when necessary. This not only improves outcomes; it also builds trust and makes the technology more acceptable to the workforce.
The goal is not to eliminate human judgment, but to extend it.
The agentic AI reality gap is real, and it’s not going away anytime soon.
The technology is still maturing, the market is confused, and the hype is out of control. But this doesn’t mean AI agents have no value.
It means that success requires discipline, realism, and a willingness to do the hard work of proper use-case selection, rigorous measurement, and organisational readiness.
The 40% of projects that will fail are those that chase the hype, ignore the fundamentals, and expect the technology to compensate for organisational dysfunction.
The 60% that succeed will be those that approach agentic AI with clear eyes, realistic expectations, and a commitment to building the foundations required for genuine autonomous operation.
The choice is yours.
Until the next one,
Chris
ScaleAI’s analysis shows that with a 20% error rate per action—a generous estimate for current LLMs—an agent attempting a five-step task has only a 32% chance of getting everything right.
Chris, may I have the link to ScaleAI? I could not find it at https://scale.com/leaderboard/tool_use_enterprise