Training for the Wrong Job

A startup called BAND appeared in public consciousness last week with $17 million in seed funding and a product most people outside enterprise IT have never thought about: governing what multi-agent AI systems do. The company builds the routing, permissioning, and audit layer that lets autonomous AI agents work together without colliding or burning through compute budgets. Think of it as recursive AI for the work output side rather than the tech development side. This alters workforce policy calculus. For two years, the dominant question has been which jobs or skills AI will eliminate. But how about taking a look at the work AI is creating, and what kind of worker will be needed to do it? The BAND example points to an answer: work in developed economies is moving up the abstraction ladder, less execution and more oversight. New analysis from Gad Levanon at the Burning Glass Institute shows three technology-exposed sectors—white-collar services, retail trade, and advanced manufacturing—posting 3.2 to 3.9 percent annualized labor productivity growth. The rest of the private economy is at 0.1 percent. The information sector alone has seen real output more than triple since 2006 while hours worked are down six percent from their 2022 peak. Multiple forces are at work; AI deployment is among them. A study by Brynjolfsson, Li, and Raymond , published in the Quarterly Journal of Economics in 2025, illuminates the productivity picture in an instructive way. The authors found that generative AI raised customer support productivity by 14 percent on average, with the largest gains accruing to less experienced workers. Read narrowly, this looks like good news for entry-level workers. Read against the Levanon data, it may tell a different story. AI is raising the performance of less-prepared workers toward a median that itself is being compressed. Though not certain, the AI flywheel may be cutting off the professional advancement routes it appeared to be opening.  The pattern across these studies is consistent. AI is compressing the value of task execution and raising the value of work that operates one level above it. This type of work can be identified and specified. Companies like BAND are building infrastructure around exactly these functions. Specification: translating organizational intent into AI-legible guardrails. Governance: setting permissions, escalation pathways, and risk thresholds. Audit: interpreting system outputs and identifying failures. The emerging knowledge-work profile requires a mix of domain knowledge, on-the-job experience, systems thinking, and accountability. These demands are in basic tension with much of workforce development policy. The Workforce Innovation and Opportunity Act , the Carl D. Perkins program , and most state and local training systems are designed to certify competencies and the ability to execute tasks in defined jobs. The economy increasingly needs something else: the capacity to design, govern, and judge work that AI performs. Particularly as it relates to judgment this creates a real challenge: these types of capacities are not usually learned in a classroom and are often too diffuse to certify via traditional credentialing methodologies. There is an opportunity inside this misalignment. The field already has some evidence-based models that build the competencies the AI economy needs: apprenticeship , sector-based training , and work-based learning. Each helps grow the abstract capacities the economy increasingly wants through extended exposure to real work. (Note: they are also notoriously difficult to scale.) The shortage of programs that focus on formation of abstract skills is further aggravated by the way AI automates the entry-level jobs that provide the required exposure. Accounting firms no longer need junior analysts who learn to interpret a balance sheet by reading hundreds of them. Law firms no longer need associates who develop legal judgment by drafting routine motions. Customer support centers no longer need agents who learn to manage hard conversations by handling easy ones first. As these entry points contract, the on-the-job training they provide goes with them. This is the bottleneck the workforce field will be solving for the next decade as it tries to push workers up the abstraction scale. The knowledge sector is already showing what the upper end of the abstraction shift looks like. AI capability will continue reaching into even higher levels of abstraction, and the pace of that extension is likely to accelerate. Meanwhile, human development proceeds at its normal, analog rate. Aligning human formation with technological change requires deliberate redesign to focus on internal human development, long time horizons, paid experience, and social patience. Workforce development is not being displaced by AI. It is being asked to solve one of the defining problems of the next decade: how to grow judgment when the experiences that produce it are increasingly pressured by automation. That is a harder, and more important, job than the one the field was built for. The post Training for the Wrong Job appeared first on American Enterprise Institute - AEI .