AI agent adoption hits 95% failure rate in enterprise pilots
An AI deployment specialist argues the failure stems from wrong sequencing, weak training and misplaced control over who builds
Nearly all enterprise artificial intelligence (AI) pilots are failing. Not because the technology does not work, but because organizations are deploying it in the wrong order, with the wrong people in charge and without the training structures needed to make it stick.
The problem is not agentic AI itself. It is the deeply human question of how companies introduce it, and a growing body of evidence from more than a thousand real-world deployments suggests that most are getting the fundamentals badly wrong.
“95% of pilots, where people were trying agentic workflows and building AI agents, are not working. That’s a crazy high number, considering the amount of money and effort going into this technology,” said Faateh Dhillon, AI specialist at Dust, an agentic AI platform. “There is more than a 97% downfall on those hand raises of colleagues using AI agents, and I really want to talk about why this gap exists.”
Dhillon asked the audience two questions by show of hands: how many had personally tried to build an AI agent, and how many could say 90% of their colleagues were doing the same. The first drew many raised hands; the second drew almost none, a gap he described as the central challenge facing enterprise AI today.
He observed a clear shift from late 2025 into early 2026, as industry conversations moved from “what is agentic AI?” to companies actively building agents at scale. The pace of that transition is accelerating, and organizations that have not yet established a coherent adoption framework risk falling significantly behind. Without one, investment in the technology tends to stall at the level of individual experimentation and fails to translate into measurable business outcomes.
The Paris-based Dust has worked with more than 1,000 companies over three years, including e-commerce software firm Clay, code editor Cursor and security compliance platform Vanta, deploying AI agents across all departments, primarily with non-technical employees who do not know how to code or build automation workflows.
That experience has given the company a clear view of where most enterprise rollouts break down and what separates the organizations that reach near-total adoption from those that stall.
Give power to people
Dhillon delivered the session “Getting to 90%: The AI Adoption Playbook for Enterprise” at the AI & Big Data Expo, part of the TechEx conference, in London. The talk drew on data from Dust’s customer deployments to offer a step-by-step framework for achieving organization-wide adoption of AI agents.
The core argument is that most companies approach AI rollout in the wrong order.
“The secret to success lies somewhere in the middle, but in a very particular order. It’s bottom-up first, eventually with the top-down structure layer that follows,” Dhillon said.
Top-down adoption, where C-suite executives mandate AI usage, is a common starting point, but it tends to produce compliance rather than genuine engagement.
“Executives do need to lead the charge, but in a certain way. You need to showcase to people that they have the capacity to experiment and to fail,” he said. “Don’t put it in a box, don’t only give it to three people and then say that you’re doing the right job.”
Bottom-up adoption, giving every employee, regardless of seniority, permission to experiment, is the right first move. With this approach in place, companies typically reach 50–60% internal adoption. But without layering in top-down governance around the 70% mark, further scaling becomes very difficult.
Dust itself fell into this trap: five employees independently built the same agent, leaving colleagues unsure which version to use for routine queries.
On the question of who should build AI agents, IT departments are frequently the wrong answer. Most organizations default to IT controlling who builds on AI platforms, distributing licenses, and owning deployment, a model that concentrates power in the wrong hands and actively harms adoption.
“This kind of technology needs to be put in the hands of every single individual, not a black box in one room. No one should have to come to IT saying, ‘Can you build me this AI agent?’ That salesperson needs to be enabled to do this themselves,” Dhillon said. “IT’s role has to become that of the facilitator, not the one that takes the power.”
He cited a fraud detection team at one of Dust’s largest customers that built an AI agent without any engineering or IT involvement, saving approximately $100,000 in five days.
“The goal is not to show you the monetary value. The goal is who builds this. The builders were not the engineers or the IT team. The builders were the fraud detection team themselves,” he said.
Make training non-optional
A recurring failure point is training, or the lack of it. Organizations that made AI tool training mandatory saw more than 70% of staff complete it, while optional training consistently produced far lower participation and higher dropout. The difference in sustained adoption between the two approaches is stark.
“Optional training sometimes does not have that effect. People experiment, they don’t know what they’re doing, they don’t have anyone to ask, and they simply stop using the technology,” Dhillon said. “If you have to impose one thing on your team, one hour a day of the week, please do the training.”
Mandatory training matters more for AI agents than for previous enterprise tools because agents represent a genuinely new paradigm. Without structured onboarding, employees unfamiliar with prompting and the differences between large language models (LLMs) quickly become overwhelmed and disengage before the technology has a chance to prove its value.
On choosing who leads adoption within teams, Dhillon cautioned against the instinct to pre-select champions.
“The people that really ended up creating interesting use cases were not the ones we were suspecting. Let the champions emerge organically, because that will show who is actually understanding and willing to spend time with this technology, versus someone you picked who might not be the best person,” he said.
AI adoption is fundamentally a cultural challenge, not a technical one. “It’s a cultural shift, a mindset shift, because it’s giving you a lot more processing power than we’ve had before,” he said.
On measuring success, Dhillon drew a clear line between early-stage and mature-stage metrics. At low adoption levels, organizations should track activity: how many agents have been built and what share of staff are using them.
Once adoption crosses roughly 70%, the focus should shift to outcomes.
“If I had a customer success manager handling $2 million of accounts, can they now handle $3 million because of the agents in place? You start connecting the usage of that agent with actual outcomes inside your organization,” he said.
E-commerce platform operator Mirakl reached 55% AI agent adoption before introducing top-down governance. Within 30 days, adoption rose to 90%, with nearly every employee using agents daily. In one Dust customer survey at 80% adoption, 90% of employees said losing access to the platform would seriously harm their work, a sign that the technology had become truly indispensable.