PwC finds daily AI use rising in UK offices, but 44% of users say workloads have grown

Adoption remains sharply uneven by seniority. The study, reported by TechRadar, found that 85% of senior executives and 73% of managers have used AI, against 35% of non-managers. PwC argues that even if more knowledge workers adopted the technology as part of their daily routines, AI is not by itself making work easier.

Employees who use the tools do report gains: 70% say AI has improved the quality of their work, 63% say it helps them use more of their skills and 58% say it brings value to the workplace. Even so, 44% say their workloads have increased since they began using AI, and 45% worry that their jobs have become more complex.

Claire Reid, PwC UK's chief innovation and technology officer, said employers and employees need to rethink how work gets done rather than simply pushing for more output or higher quality. "The real prize isn't just doing more work; it's redesigning work so that people and AI together can deliver better outcomes," she said.

PwC attributes part of the gap between managers and other staff to fear of negative consequences from experimenting with AI, confusion about employer policy and a need for more training, suggesting a clearer strategy and better communication could widen access to the technology's benefits. Reid added that organisations need to "rethink work around the right problems, give people the skills and confidence to use AI well, and put the right safeguards around it."

Cost is emerging as a second constraint on workplace AI. The shift from small-scale experimentation to deploying AI agents at scale has brought bigger bills, with Uber saying it spent its entire 2026 AI budget in four months, according to the TechRadar report. That has pushed many companies toward model routing and cheaper open-source options, including Chinese models such as Kimi K3, which offers cut-price access to near-frontier capability.

Alteryx's chief product officer, writing on TechRadar, argues that targeting which models handle which work is a legitimate way to control spending but is too narrow on its own. Organisations that never define when a large language model should be used, and where ordinary data processing makes more sense, will generate unnecessary token consumption, he wrote.

The problem, in his account, is that LLMs constantly rebuild context to answer questions. Tasks such as file reconciliation, applying business rules, compliance checks and interpreting source documents can be completed by a model, but not cost-efficiently, and many of the requests put to LLMs are repeatable, so there is no reason for a model to start from scratch each time.

He points to a business logic layer holding pre-built analytics workflows that encode organisation-specific rules and definitions. Asked by an employee to calculate a team's current margin performance, a standalone model might pull data from many sources, fill a large context window and still return the wrong figure if it ignores internal definitions of margin; a business logic layer supplies workflows that calculate the core variables daily, and the model works from that output. Such layers can also carry compliance guardrails and connect to cloud data platforms, limiting duplication of compute costs. "The cheapest token is the one you never generate," he wrote.

The argument carries particular weight for tax, compliance and finance queries, which the author describes as deterministic questions requiring deterministic answers, and as agent deployments spread. In an enterprise setting, he wrote, 1,000 agents across a workforce cannot produce 1,000 different answers to every question and still be a force for good.