On September 15, The Conference Board published a new workplace AI report that matters less as a prediction than as a decision tool. In *AI and the Labor Force: Scenarios for Stakeholders*, the organization lays out four plausible paths for the U.S. labor market: gradual augmentation, concentrated gains, massive displacement, and uneven disruption.
That framing is useful because AI adoption is already spreading faster than clean evidence on what it is doing to productivity, employment, and pay. The Conference Board’s accompanying release says that through the end of 2025, about 41% of U.S. workers and 18% of U.S. firms reported using AI. But a usage figure does not tell a manager whether a tool is replacing tasks, lifting output, creating rework, changing skill needs, or merely sitting in pilot mode.
The real question for employers is straightforward: if the labor effects are still modest, uneven, and hard to measure, how do you act now without overreacting? The answer is to treat the report as an early-warning system. Measure task changes first, watch for signals that point toward one of the four scenarios, and tie workforce moves to those signals rather than to a headline about jobs disappearing.
Four futures, one operating problem
The Conference Board is careful not to assign probabilities or present a base case. That matters. These scenarios are not a hidden forecast.
In plain language, gradual augmentation means AI helps more people do their jobs better without creating immediate pressure to cut headcount. Concentrated gains means the benefits show up first in a narrow set of firms, teams, or regions, raising the risk that productivity and pay pull further apart. Massive displacement is the scenario executives fear most: task removal spreads faster than new work, retraining, or demand growth. Uneven disruption sits between those poles, with some occupations losing work while others expand, producing churn and mismatch rather than a single economy-wide outcome.
The report argues that within three years, 60% to 70% of jobs in the cognitive workforce could involve human-AI collaboration, compared with 15% to 25% involving human-only work. That is a projection about how work may be organized in part of the labor market, not a claim that 60% to 70% of all jobs will vanish.
The management distinction is the one too often lost in public debate: task exposure is not the same thing as job loss. A model may automate part of a role, augment a worker, raise output enough to expand demand, or shift hiring toward adjacent skills. Two companies can adopt the same tool and get opposite labor outcomes depending on workflow design, error tolerance, customer demand, capital budgets, and whether productivity gains are used to grow or to trim staff.
The evidence so far says measure the work, not the headlines
The Conference Board’s own evidence points to why broad claims remain premature. It cites studies showing real productivity gains, but gains that vary sharply by task and worker experience. In one customer-support study, issues resolved per hour rose 14%, with a 34% gain among novice and lower-skilled workers and little effect for experienced, highly skilled workers. A writing experiment found AI improved productivity and satisfaction, again with larger gains for weaker baseline writers. Software-development research reported a 26% increase in task completion, with stronger benefits for less-experienced developers.
Those findings are important for managers because they suggest early AI value may come less from eliminating entire roles than from lifting the floor for newer or weaker performers. They are less useful as proof of what happens to total payrolls.
A Federal Reserve Bank of Atlanta working paper, *Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives*, offers a near-term reality check. In its survey of CFO-panel and supplemental respondents, firms reported a mean AI-attributed labor-productivity increase of 1.8% in 2025 and expected 3.0% in 2026. The paper’s implied productivity measure was smaller: about 0.6% in 2025 and 1.8% expected in 2026. Among firms that had not invested, 42% said the technology was too immature, while 36% cited insufficient workforce training and 36% privacy concerns.
That does not undercut the scenario approach. It sharpens it. If realized gains are still modest and barriers remain high, companies should resist making sweeping labor decisions from adoption rates alone. The same is true of the report’s mid-2026 survey finding that 47% of U.S. employees said their organization had adopted AI. Adoption can mean enterprise-scale workflow redesign, or it can mean a scattered set of experiments.
Build a 90-day baseline before you touch headcount
The practical move is to spend the next 90 days building a baseline at the task level. Start with where AI is actually being used, by whom, and for which steps in a workflow. Then track output per labor hour, cycle time, quality, rework, the share of AI outputs that need human correction, staffing by occupation, hiring mix, wage changes, internal mobility, training completion, and worker experience.
This is the missing bridge between technology pilots and labor decisions. If output rises but rework also spikes, the tool may be adding hidden labor. If cycle time falls but only in one team, you may be looking at concentrated gains rather than a companywide transformation. If task removal grows while internal mobility stalls, the risk is not just efficiency but displacement.
Quarterly reviews should then ask a small set of hard questions. Which tasks are being automated, and which are being redesigned around human judgment? Are productivity gains stable after accounting for correction and supervision time? Are managers hiring differently because of AI, or because demand changed? Which workers are benefiting most, and which are losing access to core tasks that once built experience?
Match actions to the scenario you are drifting toward
If the signs point to gradual augmentation — rising output, steady quality, healthy demand, and no broad staffing pressure — the play is deeper workflow redesign and targeted training. Make the skills pathway specific to changed work, not to generic “AI literacy.”
If gains are concentrated in a few teams, business units, or metros, treat that as an inequality warning. The report notes that adoption has been higher in larger firms and knowledge-intensive sectors such as professional services, finance, and insurance, and that only about one-third of U.S. metro areas rank in the top tier of AI workforce readiness under one readiness analysis. That argues for widening tool access, reviewing pay and promotion effects, and investing where readiness is weakest before the gap hardens.
If task removal and hiring freezes start spreading across functions, the right response is redeployment before reduction. Set trigger points in advance: for example, a sustained fall in task volume for a role, a drop in external hiring for related work, and weak internal placement rates. That is when transition support, reskilling budgets tied to real openings, and clearer worker communication should turn on. Scenario planning is not permission to announce layoffs early and sort out the evidence later.
If disruption is uneven — some roles shrinking while others grow — the central problem is matching. Companies need pathways across occupations, not only within them. Workers need to know which tasks are changing first and what adjacent roles are opening. Policymakers, meanwhile, should care less about a single national number than about local readiness, training capacity, and whether support systems can handle churn if adoption accelerates.
The smart reading of The Conference Board’s report is not that the labor market has already turned a corner. It is that companies can no longer afford to manage AI as a software rollout alone. The next advantage will come from knowing, sooner than competitors, whether your organization is augmenting work, concentrating gains, displacing tasks, or reshuffling demand — and having a workforce plan ready for each path.
By
By
By
By
