Two-thirds of organizations running agent AI projects never make it past the pilot stage. These numbers tend to surprise people, until Senthil Muthiah, senior partner at McKinsey & Company, explains why. He argues that stalling is not a technical issue. It’s a matter of strategy. Moreover, most companies make the same two mistakes.
The first is to treat all tasks the same, Muthiah explains.
“Service workflows can range from highly structured, rule-based tasks to tasks that require human judgment. Many companies treat tasks as if they are all the same, layering agent AI across the board with a one-size-fits-all approach. While agent AI moves quickly in structured areas, it tends to slow down when it comes to human decision-making, where change management is key.”
The second mistake is spreading your investments too thinly. Instead of identifying where AI will create the most value and focus efforts there, companies deploy it and hope the results follow. “Each company has a set of economic leverage points that create disproportionate value when AI is applied,” says Muthiah. “Many companies are taking a more organic, blanket approach of applying AI everywhere and there is no clear link to value.”
Both failures are made worse by impatience. He says there is a tendency to apply much more stringent requirements to AI deployments than apply to humans and then evaluate them against inflated standards.
“The goal shouldn’t be AI for everything. AI for the right things, so people are free to focus on high-value work in harmony.”
Related articles from UC Today:
Getting the right handoff
Once an organization knows where to focus, the next challenge is to map out the boundaries between what AI handles and what humans handle.
“The best handoffs happen when both parties are doing what they do best,” he says.
“AI can perform structured, rule-based tasks, and people can step in where nuance, judgment, and real-time decisions are needed. When you design workflows with this in mind, the transition between AI and humans starts to feel much more natural.”
For example, a cash order process may have multiple agents working across invoicing, collections, and dispute resolution before passing exceptions to human operators. Rather than having one team waiting to be sorted, the work progresses from start to finish. But Muthiah is quick to separate the technical and human issues.
“The real challenge we see is not the seamlessness of the handoff itself, but the use and change management required. It has not yet been proven that giving humans seamless AI will necessarily lead to their use of AI.”
This gap between deployment and adoption is something organizations consistently underestimate. A well-designed agent workflow means little if the people it’s built for don’t trust or understand it, or if there’s no reason to change the way they currently work.
What changes will happen to employees?
Much of the conversation about agent AI focuses on what gets automated. Muthiah moves the frame to the cleared one. McKinsey’s research shows that even in a heavily AI-enhanced environment, 70% of human skills are still essential, and when the balance is right, the impact on everyday work can be truly positive.
“AI means people spend less time on repetitive, routine tasks that can be automated with high integrity, allowing them to focus on tasks that actually require judgment and expertise. Over time, this change will make their work more meaningful and focus them on more valuable moments.”
But getting there requires a level of investment that most business cases cannot account for. “Our research shows that for every dollar spent on technology, organizations need to invest approximately $2 in change management, capability building, and adoption to fully realize the benefits.”
A 1:2 ratio significantly reframes the ROI conversation. According to McKinsey’s extensive workplace AI research, the long-term productivity opportunity is worth $4.4 trillion. However, this only applies to organizations that treat the people side of change with the same severity as the technology side. For IT and operations leaders building internal business cases, these numbers are worth keeping in mind.
Governance gap that no one has solved yet
The most honest moment in Muthiah’s assessment is when the conversation turns to governance. Organizations are deploying agents over time without a clear function responsible for managing them, and he’s not making it up.
“Currently, organizations do not have the ability to create, scale, perform performance management, scale, and terminate agents,” he says. “This will be a new organizational capability in the future. At the moment, there is no clear view on who should own it within the organization.”
The comparison to managing people is intentional. Agent governance should work hand-in-hand with workforce planning and performance management, following similar principles, even if the metrics may look different. For UC platforms where agents evolve into a workflow execution layer that triggers actions, routes tasks, and manages escalations across the system, lack of ownership is a real operational risk. According to the McKinsey State of AI 2025 report, the proliferation of agentic AI is already outpacing the governance structures organizations have in place to oversee it.
Where should I start?
For leaders who want to get past the pilot stage, Mutia’s advice is intentionally unappealing. Pick the economic leverage points where AI delivers the most concentrated value, give it appropriate management attention, and start at the rule-based workflow end of the spectrum where initial success is more predictable.
“We are already seeing the impact of this approach,” he said, pointing to the deployment of digital twins that simulate and optimize operations and service transformation programs that have reshaped customer operations around AI-enabled workflows.
The productivity gains from McKinsey projects are significant. But this happens to organizations that treat agent AI as an operational discipline rather than a technological experiment. Muthiah suggests that breaking away from the pilot trap starts with a more honest answer to the simple question: not where AI can be applied, but where it will actually be applied.