AI agents are changing not just how organizations work, but how they staff. Anthropic engineers now use its AI tools to write practically 100% of their code, and Gartner reports that by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention.
Both tech and hiring budgets are being re-evaluated as a result. McKinsey CEO Bob Sternfels now counts AI agents as part of their workforce, with his latest tally putting his headcount at 60,000, consisting of 40,000 humans and 20,000 AI agents.
As companies pull AI virtual agents out of the tech budget and into the hiring budget, leaders face a tricky challenge: deploying AI in ways that protect service quality while preserving the human connection and seamless experience customers still expect.
What CX leaders need to consider before they deploy
The productivity gains from AI are real. Researchers find AI produces faster output and 60% greater overall productivity, allowing for up to 36% more time for higher-order work. The problem is that many organizations are rushing to hit those numbers, treating AI agents as straightforward headcount replacements, and running into trouble on three fronts.
First, cost. The budget math is harder than it looks. Full replacement of human workers with AI agents is proving more expensive than a balanced approach.
Tech investor Jason Calacanis recently noted that AI agent costs quickly rose to $300 a day while using the Claude API at one of his organizations, while those agents were replacing only a fraction of what a human employee handles. Among large corporations, AI budgets are skyrocketing. A recent Wall Street Journal assessment finds that Microsoft, Alphabet, Meta and Amazon spent a combined $410 billion on capital expenditures last year and are expected to spend more than $670 billion in 2026 as they invest in more chips, servers and data centers to support growing AI usage.
That investment only pays off when agents are deployed on the right problems at sufficient volume. Spread across a portfolio of underdeveloped pilots, the run costs compound fast.
Second, operational drag. Without well-defined success criteria, AI deployments face a garbage-in, garbage-out problem. Even a well-built agent will underperform if it’s pulling from bad data, overextended or focused on problems it isn’t equipped to handle. The discipline of defining what’s broken before deciding what to automate directly impacts efficiency.
Third, a worse customer experience. Capacity’s Closure Index research found that only 42% of customers say their issue feels fully resolved after a support interaction, and one in three stops using a brand after a single unresolved experience. Poorly deployed AI directly harms the customer experience and their ability to achieve closure by abandoning customers in automation loops, losing context across channels and leaving them with lingering concerns.
Capturing the real ROI of agentic AI means CX leaders have to approach this differently: It starts with selecting the right problems before solving them.
How to decide what belongs to a virtual agent today, tomorrow or never
CX leaders tend to be great problem solvers. But with AI deployment, the much harder task is problem selection: determining which tasks to automate first, which can wait and what needs to stay in human hands.
- Start with high-volume tasks with predictable inputs and clear resolution paths. AI excels in interactions where the customer needs a clear, quick answer. Tier-1 support interactions, like requesting an order status, password resets or appointment rescheduling, consume significant time and have the least strategic and relational value. Deploying virtual agents there deflects volume in a meaningful way, takes pressure off the human team and frees agents to focus on the interactions where they can make a real difference.
- From there, move from volume to complexity. The next tier of automation candidates are issues that follow a defined process but require access to multiple systems or data sources to resolve. These take longer to deploy well because the AI is only as good as the data it can access. Integration with underlying systems of record isn’t optional.
- Keep humans involved. What stays human, at least for now, are the interactions where the outcome depends on judgment, tone or emotional reassurance. According to the Closure Index, 85% of customers say the ability to seamlessly move from AI to a human when needed is important for feeling true closure after a support interaction.
Customers can tell when a situation exceeds what an AI agent can handle, and they need to know a person is reachable. Building that escalation path as a first-class feature, not an afterthought, is what keeps automation from becoming a liability.
How well-scoped AI deployments can generate ROI
Successful deployment is possible. Forward-thinking leaders are embedding AI agents into simplified, well-scoped processes so automation gradually replaces work rather than stacking infrastructure that never fully delivers.
Airbnb built a custom AI agent that now handles roughly a third of its customer support issues in North America. With global expansion underway, the company expects more than 30% of total support tickets to be handled by AI voice and chat across all languages where it also employs human agents.
PacSun took a similar approach, using virtual agents to handle 85% of all customer inquiries. Previously, more than half of PacSun’s customer chat interactions happened outside working hours. With AI handling around-the-clock volume, customers are never left without support, and human agents can focus on the interactions that actually need them.
The future of hiring in an AI-driven world
Hiring and tech budgets will both stay in flux as this shift plays out. For the investment in agentic AI to pay off, CX teams need to act with intention, selecting the right problems first, building clean escalation paths and maintaining the human touch where it matters most.
The organizations that bring that discipline to their deployments will be the ones that build lasting competitive advantage. The ones that don’t will have a lot of abandoned infrastructure to show for it.
