AI Agents Face a Real World Hurdle, Not Just a Promise

Most firms want AI agents to run entire workflows in three years, but three-quarters can't today.

People want the future of work to be smoother, faster, and mostly autonomous. The hitch, according to the latest briefing on enterprise agentic AI, is that ambition is racing ahead of organizational design. A survey shows 85 percent of organizations aim to be agentic within three years, yet 76 percent say their current operations and infrastructure can’t support that shift. The gap isn’t just tech debt; it’s the entire operating model. PwC UK Consulting’s Prasun Shah argues the problem isn’t about bolting smart agents onto old processes. It’s about rewiring the work itself. “They’re embedding AI employees into what is a human operating model,” he says, likening the move to sticking tape on a structure that’s already breaking.

The promise is bold. When deployed at scale, AI agents could accelerate business processes by 30 percent to 50 percent and cut low-value work time by 25 percent to 40 percent in the early proving grounds of customer service, HR, and sales. But that potential comes with a caveat: the more capable the agents become, the more complex the orchestration behind them has to get. The industry has started talking in terms of agentic business transformation, or ABT, a term co-promoted by an agentic AI platform called Ema alongside HFS Research. ABT signals more than a pilot project; it signals a full redesign of how work gets done, who signs off on decisions, and how success is measured.

Two big realities frame the current moment. First, the value of agentic AI lives not in standalone automations but in end-to-end workflows that agents can coordinate, adapt to changing conditions, and iterate on with limited human input. That makes the organizational changes more sweeping than most pilots anticipate. Second, readiness is uneven. The same group talking about “agentic” transformations struggles with people, processes, and data flows that are not yet aligned. In practice, many firms are layering new AI agents onto existing, sometimes brittle, workflows instead of rebuilding the underlying operating system. It is a classic case of the sticky tape problem, patching symptoms without fixing the core architecture.

From a practitioner standpoint, three, four, even five concrete conclusions follow. First, the path to value must start with rethinking the operating model. The strongest gains come when AI agents are trusted to drive complete tasks across functions, rather than serving as smarter assistants to humans who still carry the entire decision loop. Second, readiness isn’t optional. A successful ABT initiative requires clear ownership of decision rights, governance, and data readiness, plus targeted training so people understand how to work with agents rather than around them. Third, complexity is a design choice. It’s tempting to chase broader functionality, but the ROI lives on disciplined scope, fast feedback loops, and robust monitoring for misalignment or escalation needs. Fourth, pilots should be explicit about outcomes. It’s not enough to measure time saved; leaders must quantify how agents improve throughput, accuracy, and customer experience, and where they still fall short.

Analogy helps: you would not swap a turbocharged engine into a car whose chassis and transmission are decades out of date. The car might thunder down the road for a moment, but it will stall, overheat, or misbehave if the rest of the system isn’t engineered to handle the power. That is the current inflection point for enterprise AI, because gains without a coherent operating model risk disappointment rather than disruption.

For products shipping this quarter, the takeaway is clear: buyers should demand ABT-aligned roadmaps, with explicit plans to redesign workflows and governance, not just to deploy more agents. Vendors should frame deployments as systemic transformations, offering ready-made patterns for end-to-end workflows, governance structures, and measurable business outcomes. If you are banking on 30 percent to 50 percent process acceleration, you need to show you’ve redesigned the operating model to actually let agents drive those outcomes, not just parceled them into a patchwork of improvements.