Let me share a recent conversation I had with a CIO at a global manufacturing company. He was excited. Really excited. His team had built an AI agent that could automatically triage IT incidents, route them to the right teams, and even resolve common issues without human intervention. The demo was slick. The boardroom was impressed. The pilot was declared a success.
Six months later, that project is sitting in what I call the “pilot graveyard.” Cancelled. The agent never made it to production.
Here is the thing. This story is not unusual. It is the norm.
And if you are leading an AI initiative right now, the statistics should make you uncomfortable.
The Numbers Nobody Wants to Talk About
We are in the middle of what everyone is calling “the year of the AI agent.” LinkedIn is full of announcements. Vendors are competing to launch agentic capabilities. Every enterprise software company has suddenly realised that its product was always designed to include autonomous agents.
But here is what the research actually shows.
According to Deloitte’s 2025 Emerging Technology Trends study, while 30% of organisations are exploring agentic AI options and 38% are piloting solutions, only 14% have solutions ready to deploy. Only 11% are actively using these systems in production.
Let that sink in. Eleven percent.
It gets worse. MIT’s NANDA initiative published research this year showing that only about 5% of AI pilot programs achieve rapid revenue acceleration. The vast majority of stalls deliver little to no measurable impact on profit and loss. Their research covered 150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments.
S&P Global’s data is equally sobering. They found that 42% of companies scrapped most of their AI initiatives in 2025, up dramatically from just 17% the year before. The average organisation abandoned 46% of AI proofs of concept before reaching production.
And Gartner? They predict that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls.
We have a situation where nearly everyone is experimenting with AI agents, but almost no one is deploying them successfully. The demos are impressive. The production deployments are rare.
Why?
The Three Killers
After three decades in this industry, working with Fortune 500 companies on digital transformation, I have observed patterns. The pilots who die share common characteristics. Let me walk you through the three killers.
Killer One: The Legacy System Problem
Here is something that seems obvious but is often overlooked. Traditional enterprise systems were never designed for agentic interactions.
Most AI agents today still rely on APIs and conventional data pipelines to access enterprise systems. This creates bottlenecks and severely limits what agents can actually do autonomously. You build a brilliant agent, but it cannot talk to your ERP. Or it can talk to your ERP, but not your CRM. Or it can access both, but not in real time.
Gartner predicts that over 40% of agentic AI projects will fail by 2027, specifically because legacy systems cannot support modern AI execution demands. The agent is not the problem. The plumbing is.
I see this regularly in my work. A company will invest heavily in an AI agent for customer service, only to discover that the agent cannot access the order management system without going through three middleware layers that add latency and break half the time.
Killer Two: The Data Architecture Mess
UiPath’s research found that lack of interoperability is the second most cited reason for pilot failures, right after data quality issues. In the same study, 63% of executives cited “platform sprawl” as a growing concern.
What does platform sprawl look like in practice? It seems like twelve different AI tools across six departments, none of which talk to each other, all generating insights that nobody can consolidate, and all requiring separate governance frameworks that nobody has time to maintain.
The Bain 2025 Technology Report puts it well. While AI investment is up, returns often lag behind expectations. They attribute this gap to fragmented workflows, insufficient integration, and misalignment between AI capabilities and business processes. In many companies, AI tools operate in silos, producing insights or drafts but failing to drive end-to-end outcomes.
Your agent is only as good as the data it can access. And if your data is scattered across seventeen systems with inconsistent formats and no unified governance, your agent will fail. Not because the AI is bad. Because the foundation is broken.
Killer Three: Treating Agents Like Software Instead of Employees
This one is perhaps the most important, and the most overlooked.
McKinsey recently published insights from over 50 agentic AI builds they have delivered. One of their key findings resonated deeply with my experience. They found that companies often hear users complain about “AI slop” or low-quality outputs. Users quickly lose trust in the agents, and adoption levels collapse.
Their conclusion? “Onboarding agents is more like hiring a new employee versus deploying software.”
Think about that for a moment. When you hire a new employee, you do not just hand them a laptop and expect them to figure everything out. You give them a clear job description. You onboard them. You provide training. You give them feedback. You help them understand the culture, the processes, the unwritten rules.
But when companies deploy AI agents, they treat it like installing a new application. Configure it, switch it on, expect miracles.
Agents need clear job descriptions. They need to be onboarded into workflows. They need ongoing feedback to become more effective over time. Developing effective agents is challenging work that requires harnessing individual expertise, creating proper evaluations, and codifying best practices with sufficient granularity.
Skip this, and your agent will produce mediocre outputs that nobody trusts. Any efficiency gains you achieve through automation will be offset by the loss of trust and decline in quality.
What the 11% Do Differently
So what separates the organisations that succeed from those that end up in the pilot graveyard?
McKinsey’s analysis is instructive here. They found that achieving business value with agentic AI requires changing workflows. Often, organisations focus too much on the agent itself or the agentic tool. This results in attractive agents that do not improve the overall workflow, delivering underwhelming value.
The successful deployments focus on fundamentally reimagining entire workflows. Not just the technology. The steps involve people, processes, and technology together.
Here is another way to think about it. Agentic AI delivers the most outstanding value when used to reengineer business domains rather than simply optimising existing tasks. The 11% are not asking, “How can AI agents make our current process faster?” They are asking, “How should this process work if we designed it from scratch with AI agents as part of the team?”
That is a fundamentally different question. And it leads to fundamentally different outcomes.
A Platform Perspective
Let me share something from my own domain. I work extensively with ServiceNow and have closely followed the platform’s evolution of its agent capabilities.
What strikes me about successful enterprise AI agent deployments is the importance of native integration. ServiceNow’s approach, for example, builds AI agents directly into the platform where workflows already exist. The agents have access to the data, the automation, and the institutional knowledge that customers have built up over the years.
Research from BCG supports this. They found that ServiceNow’s AI agents and Now Assist capabilities are automating IT, HR, and operational processes, reducing manual workloads by up to 60%. Early adopters are seeing 20% to 30% faster workflow cycles.
The difference? These are not bolt-on AI solutions trying to integrate with existing systems through fragile APIs. They are agents that operate natively within the workflow fabric.
This is not about any single vendor. The principle applies broadly. Organisations that succeed with agentic AI choose platforms where the agents are deeply integrated rather than superficially attached.
How to Join the 11%
If you are running an agentic AI initiative or planning one, here is what I would suggest based on what I have seen work.
First, start with workflow redesign, not agent deployment. Before you build or buy any agent, map the entire workflow you want to transform. Understand where humans add value, where they do not, and where handoffs create friction. Design the future state workflow first. Then determine where agents fit.
Second, define measurable outcomes before writing any code. “Improve efficiency” is not a quantifiable outcome. “Reduce invoice processing time from eight days to two days while maintaining 99.5% accuracy” is. If you cannot define success precisely, you cannot achieve it.
Third, conduct a structured review of existing pilots. McKinsey recommends formally closing the exploratory phase by capturing lessons learned and retiring unscalable pilots. Stop throwing resources at experiments that will never reach production. Be honest about what is working and what is not.
Fourth, choose platforms with native integrations. Organisations serious about agentic AI need to prioritise platforms with native integrations, open APIs, and flexible orchestration capabilities. Bolting AI onto legacy systems through middleware rarely works at scale.
Fifth, treat your agents as new team members. Give them clear job descriptions. Onboard them properly. Provide feedback and iteration. Measure their performance and help them improve. Organisations that treat agent deployment as a change-management challenge, not just a technology deployment, are the ones that succeed.
The Real Question
Here is the thing. The technology is no longer the problem. The models are capable. The platforms are maturing. The tools exist.
The problem is how we approach implementation. We still think of AI agents as software to be deployed rather than capabilities to be cultivated. We are still trying to bolt new technology onto old processes instead of reimagining how work should flow.
McKinsey put it well in their recent CEO guidance. The time for exploration is ending. The time for transformation is now.
So let me leave you with a question. Are you deploying agents, or are you transforming how work gets done?
Because only one of those approaches will get you into the 11%.
Enamul Haque is Director of Intelligent Solutions at Wipro, a technology author with over 20 published books, and adjunct professor at Bangladesh Maritime University, where he teaches data science and emerging technologies. His latest book, “AI Horizons: Shaping a Better Future Through Responsible Innovation and Human Collaboration,” is published by Mercury Learning and Information.
References
- Deloitte, “2025 Emerging Technology Trends Study” (2025)
- MIT NANDA Initiative, “The GenAI Divide: State of AI in Business 2025” (2025)
- S&P Global, AI Initiative Research (2025)
- Gartner, “Agentic AI Project Predictions” (June 2025)
- UiPath, “2025 Agentic AI Report” (2025)
- Bain & Company, “2025 Technology Report” (2025)
- McKinsey & Company, “One Year of Agentic AI: Six Lessons from the People Doing the Work” (September 2025)
- McKinsey & Company, “Seizing the Agentic AI Advantage” (June 2025)
- BCG, “How Agentic AI is Transforming Enterprise Platforms” (October 2025)