UK workers expect to hand over nearly a third of their workload to artificial intelligence agents within the next year, but most still regard the technology as unreliable and say responsibility for mistakes remains unclear, according to new research.
A survey of more than 1,000 UK knowledge workers, published in Asana’s Global State of AI at Work 2025 report, found that 64 per cent considered AI agents unreliable, while more than half said the systems either ignored feedback or produced inaccurate information with confidence. Despite this, 74% of employees already use AI agents in some capacity, and three-quarters see them as a fundamental shift in how work will be organised.
The research highlights a sharp disconnect between ambition and preparedness. On average, workers expect to delegate 32% of their tasks to AI within the next 12 months, rising to 41% within three years. Yet only a quarter believe they are ready to do so today.
Accountability is emerging as a critical gap. When asked who should take responsibility for an AI agent’s mistakes, 20% of respondents pointed to the end user, 18% to IT teams, and just 9% to the system’s creator. But 39% either assigned blame to no one or admitted they did not know. Only a tenth of organisations surveyed reported having ethical frameworks or clear deployment processes for AI, and fewer than one in five said they measured agent errors.
This lack of oversight risks compounding what Asana described as “AI debt,” the accumulation of costs from unreliable systems, poor data quality, and weak governance. Nearly four-fifths of organisations expect such burdens to grow.
The survey also points to a training gap. While 82% of employees said proper instruction was essential to use AI effectively, only 32% reported receiving any. More than half called on their companies to establish clearer boundaries between human and machine responsibilities.
Mark Hoffman, Work Innovation Lead at Asana’s research unit, said that while adoption was spreading rapidly, trust had not kept pace. “When no one is responsible for mistakes, employees hesitate to hand over meaningful tasks, even though they’re eager to delegate,” he said. “To succeed, organisations must treat agents like teammates — by providing context, defining responsibilities, embedding feedback loops, and training employees to use them effectively.”
Without such measures, the report warns, AI adoption risks remaining confined to administrative tasks such as organising and locating documents or scheduling meetings, rather than extending to more complex work that could deliver significant productivity gains.
