In the first half of 2026, a confluence of signals — slowing monthly hiring numbers, a wave of tech-sector headcount reductions, and the rapid productionisation of large language models — has pushed AI’s impact on the labor market from abstract concern to concrete policy debate. Against that backdrop, Goldman Sachs has now attached a specific number to the disruption: 15 million US workers.
What Happened
Joseph Briggs, who leads global economics research at Goldman Sachs Research, appeared on the bank’s Exchanges podcast to lay out his current read on AI-driven labor market dynamics. His headline figure: approximately 9% of the US workforce — translating to around 15 million workers — will be displaced as AI adoption broadens across the economy.
“9% of workers being displaced by AI would correspond to 15 million workers,” Briggs said, drawing an explicit parallel to the tech-driven labor upheaval of the late 1990s and early 2000s. In that framing, displacement means workers must exit their current roles and seek employment elsewhere — not that the jobs simply vanish from the economy permanently.
Briggs also offered a more immediate data point. In sectors where AI tooling is already embedded — technology, management consulting, and graphic design — he estimates the technology is suppressing monthly payroll growth by somewhere between 10,000 and 15,000 jobs. That is a measurable drag on headline employment figures, happening right now, not at some future inflection point.
Also appearing on the same episode was MIT’s Neil Thompson, who took a notably more tempered view. Thompson argued that raw AI capability is only the first barrier to mass displacement. An AI system also requires access to the appropriate data — a non-trivial obstacle in fields like medicine where HIPAA-style privacy regulation constrains training pipelines — and it must be economically viable to deploy at scale. In Thompson’s assessment, those friction points mean real-world adoption will lag well behind benchmark-level capability for years.
Why It Matters
The timing of Briggs’s comments is not incidental. The June 2026 US jobs report, released on July 3rd, showed the economy added just 57,000 jobs — roughly half of consensus economist expectations. April and May figures were also revised down by a combined 74,000. The unemployment rate fell to 4.2%, but largely because workers exited the labor force rather than because hiring accelerated. Taken together, those numbers gave Briggs’s podcast remarks an immediate empirical backdrop that would have been absent six months ago.
What makes the Goldman analysis structurally interesting is the tension it reveals between two different timescales. Briggs’s historical argument — that 85% of job growth over the past 80 years has come from technology creating entirely new categories of work — is compelling as a long-run equilibrium claim. But Thompson’s friction-point argument operates on the medium-run adoption curve, where data access, regulatory compliance costs, and inference economics all slow deployment. The June payroll miss sits in a third, shorter timeframe: the immediate labor market signal. The risk is that analysts and policymakers conflate these three timescales, using long-run optimism to dismiss near-term pain or using near-term softness to catastrophize about permanent displacement. Neither move is analytically defensible.
Briggs himself resists the catastrophist read. The US labor market, he noted on the podcast, creates and destroys roughly 30 million jobs annually. A 5% acceleration in job creation — well within historical variation — would, in his model, be sufficient to reabsorb the estimated 15 million displaced workers. That is a meaningful benchmark for tracking whether the optimistic scenario is materializing.
The sector-specific suppression data Briggs cited is worth dwelling on for engineers and architects who work in or adjacent to the affected industries. A 10,000–15,000 per-month drag on tech, consulting, and design payrolls is already showing up in the aggregate. This aligns with the broader pattern of large enterprise AI spending being financed partly through headcount reduction, rather than representing purely additive investment. The arithmetic is straightforward: if AI tooling raises individual developer or analyst productivity materially, firms face pressure to demonstrate margin improvement — and headcount is the most legible lever.
Thompson’s GPS-versus-taxi-driver analogy is analytically useful here. When turn-by-turn navigation automated the core cognitive skill that taxi drivers monetized — route knowledge — individual driver wages fell. But the total number of drivers increased substantially because lower friction made the market larger. AI may follow a similar pattern in knowledge work: individual compensation for certain task types may compress, while the total number of people performing AI-augmented versions of those tasks grows. Whether that net outcome is positive depends heavily on how quickly new task categories emerge and whether they are accessible to workers displaced from the automated ones.
For software teams specifically, the implication is architectural as much as economic. The firms building AI-native workflows are making bets about which tasks sit below the automation threshold — where the cost of inference, the quality of available training data, and the regulatory surface area make automation economically rational. Thompson’s point about data access is particularly salient: in domains with rich, structured, publicly accessible data (code generation, document summarisation, data analytics), automation is already viable. In domains with fragmented, privacy-constrained, or poorly labelled data (clinical diagnostics, legal discovery, novel research), the timeline extends materially.
There is a broader market context worth noting. A growing cohort of economists is raising structural concerns about AI’s macroeconomic trajectory, even as enterprise adoption accelerates. The Goldman framing — net-positive over the long run, disruptive in the medium term — represents the centrist consensus among institutional economists, but it is being stress-tested by the current hiring data.
What the Goldman AI Jobs Story Is Missing
The Briggs–Thompson podcast exchange is analytically rigorous by mainstream financial media standards, but it leaves several important questions underweighted or unaddressed.
1. Geographic and occupational heterogeneity. The 15 million displaced workers figure is a national aggregate. It obscures enormous variance by region, education level, and specific occupation. A software engineer in a major metro with transferable AI skills faces a radically different transition path than a paralegal in a mid-sized city whose document-review workflow is being automated. The aggregate reabsorption argument — 5% faster job creation absorbs everyone — assumes labor markets clear efficiently across geography and skill type. That assumption has historically been wrong during technology transitions, as the current surge in AI-skill requirements in job postings illustrates: new positions are being created, but they require a specific capability set that displaced workers may not hold.
2. The inference cost trajectory. Thompson mentions economic viability as a friction point, but neither he nor Briggs quantifies what happens if inference costs continue falling at their current pace. GPU efficiency gains, model distillation, and competition among cloud providers have all contributed to dramatically lower cost-per-token over the past 18 months. If that trend continues, the economic viability threshold that currently protects certain roles could shift faster than either the optimistic or pessimistic labor model anticipates. The podcast treats this as a static constraint; it is a dynamic variable.
3. Policy and institutional response lag. The historical 80-year argument implicitly assumes that institutional structures — retraining programmes, unemployment insurance design, educational curricula — will adapt at a pace sufficient to manage the transition. There is scant evidence that current US labour policy infrastructure is calibrated for a displacement event of this speed and scale. That gap deserves explicit treatment, not a footnote assumption.
What Happens Next
The immediate question is whether the June jobs report is a data anomaly or the leading edge of a structural deceleration. One month of weak payrolls — even with two months of downward revisions — is statistically insufficient to confirm a trend. But if July and August figures show continued softness in the specific sectors Briggs identified as already absorbing AI-related job suppression, the debate will shift from theoretical to urgent.
On the model side, the next meaningful data point will be whether enterprise AI deployment translates into measurable productivity gains in Q2 and Q3 earnings calls. Companies that have cut headcount while increasing AI spend will face analyst scrutiny over revenue-per-employee and margin trajectory. If productivity gains are visible, the Briggs reabsorption thesis gets a data point in its favour. If margins contract despite headcount reductions, the calculus looks considerably more complicated.
For engineers watching this space, the Thompson friction-point framework provides a practical monitoring heuristic: track which domains are seeing rapid cost-of-inference decline alongside improvement in available training data quality. Those are the sectors where the automation timeline is compressing fastest — and where the displacement-versus-creation dynamic will play out first.
The debate between the “rising tide” and the “crashing wave” framing will not be resolved by a single report or a single podcast. But Goldman Sachs has now provided a specific, falsifiable number — 15 million — that will anchor every subsequent analysis. How enterprises actually generate ROI from AI deployment will determine whether that number becomes a manageable transition cost or a political flashpoint.
Three Things to Track
- Monthly payroll data in tech, consulting, and design. Briggs cited these three sectors as already showing a 10,000–15,000 monthly job-growth drag attributable to AI. Watch the Bureau of Labor Statistics monthly establishment survey breakdowns in these subsectors over the next two quarters for confirmation or reversal of that signal.
- Enterprise AI productivity disclosures in earnings calls. As Q2 2026 results are reported, monitor whether companies that have paired AI investment with headcount reductions can demonstrate measurable productivity gains. Specific metrics to watch: revenue per employee, gross margin change, and software engineering output proxies where disclosed.
- AI inference cost benchmarks. Track published cost-per-token and cost-per-task figures from major cloud AI providers. If inference costs fall below key thresholds in data-rich professional domains — legal, financial analysis, mid-market software development — Thompson’s “economic viability” friction point disappears faster than either the optimistic or pessimistic labor models assume.
