What it takes to build an AI-ready workforce

A KPMG study examined roughly 100,000 conversations and approximately 1.4 million prompts from about 2,500 employees across eight months spanning January through August 2025, covering back-office roles in accounting, finance, communications, and HR. Despite achieving around 90 percent AI adoption, only approximately 5 percent of employees were using AI in a way KPMG considered sophisticated or genuinely value-creating. That gap prompted KPMG to partner with UT Austin to define and measure effective use. Anthropic's Kristin Swanson described a parallel AI Fluency Index built on a 4D framework covering delegation, description, discernment, and diligence, developed with academics Rick DeCon and Joe Feller. The fluency index study covered a one-week window with approximately 10,000 conversations, and despite the difference in scale and time horizon, both studies reached the same conclusions, providing mutual triangulation and validation.

Sophisticated users were distinguished by several concrete behaviors: they switch between AI models depending on the task, use casual but specific language reflecting genuine comfort with the system, treat AI as a reasoning partner by giving bounded instructions and pushing back on outputs, and question the model's reasoning process even when it is not visible in the interface. Unsophisticated users, by contrast, tend to interact with AI the way they would use a Google search, accepting outputs without iteration or challenge. Swanson argued that iteration is the gateway to sophisticated AI use, because continued refinement leads to outputs users are more comfortable acting on, and that behaviors making someone effective when collaborating with humans transfer directly to effective AI collaboration.

Jamie Schmidt noted that AI training produced statistically significant behavioral improvement in the month it was taken, but that improvement began to diminish shortly afterward, consistent with the broader observation that training effects wear off within roughly a month. KPMG has responded by shifting toward local, hands-on, team-based training where groups working in the same industry tackle targeted AI problems together rather than receiving generic instruction. Swanson noted that demand for AI education is at a volume not previously experienced, and that helping people understand foundational concepts such as what a context window is, and providing tools like MCPs to help manage it, are treated as important steps in building genuine AI readiness.

As AI use shifts toward agentic surfaces, both speakers identified new demands on users. Setting a clear goal or north star for an agent emerged as a uniquely important early behavior in agentic contexts. Schmidt described the current moment as a judgment frontier, arguing that the judgment dimension becomes even more critical when autonomous agents are directing other agents, which KPMG is already doing with orchestrator agents. Within both studies, it was rare for participants to set acceptance criteria or ask the AI to identify weaknesses in its own output, and very few pushed back on AI outputs at all. Both speakers flagged this pattern as a serious risk as agentic systems become more independent.

Coding was cited as a strong early domain for agentic tools partly because software development has an established culture of testing. A renaissance of testing across all knowledge work was predicted as agentic systems grow more capable. One accounting firm example illustrated the stakes: the firm ran 5,000 iterations to determine acceptable performance thresholds before deploying an agentic tool to its workforce, and firms are also grappling with how to respond when context changes and a previously validated tool may no longer perform adequately.

The 4D framework centers delegation as its first principle, helping learners decide which tasks are appropriate for AI versus humans. Constant relearning is treated as a core feature of the curriculum given the pace of change, and Swanson described her remit at Anthropic as exploring model-agnostic and product-agnostic collaboration behaviors to support a safe transition to AGI. The combined picture from both studies is that high adoption rates are a poor proxy for value creation, that the behaviors separating effective from ineffective AI users are learnable but require sustained reinforcement, and that the stakes of those behaviors rise sharply as agentic systems take on more autonomous decision-making.