AI Can Cut Costs. What Happens to Workers — and Margins?

Chegg’s revenue per employee rose about 71 per cent while the company lost half its revenue and 71 per cent of its staff.

IBM discloses no total employee count anywhere in its 10-K filings across this window.

The US layoffs-and-discharges rate averaged 1.07 per cent over January to July 2026, against 1.21 per cent in 2019. It has stayed inside a 0.9 to 1.2 per cent band for five straight years.

Over the same period the hires rate fell from 3.87 to 3.31 per cent — down about 15 per cent — and the quits rate fell about 15 per cent, while job openings returned to within 3 per cent of their 2019 level. Firms are still advertising and not converting.

In the Information sector the layoff channel did open: the layoffs-and-discharges rate rose from 1.11 per cent in 2023 to 1.84 per cent over January to July 2026, sustained across seven months and about 36 per cent above its 2019 level.

Anthropic cut the list price of its top Opus tier by about two thirds in one step — from $15 and $75 per million input and output tokens to $5 and $25, announced on 24 November 2025 — and has held that price across four subsequent releases.

The same vendor’s pricing documentation states that its newer models use a tokenizer producing approximately 30 per cent more tokens for the same text. The price per unit fell and the unit itself got smaller — and the seller says so, on the same page.

A new tier was introduced above the repriced one. As of 3 September 2026 the same vendor lists models at $10 and $50 per million tokens — twice the price of the tier that was the top of its range a year earlier.

Where the seller had consumer pricing power, the price went the other way. On 16 January 2025 Microsoft bundled Copilot into Microsoft 365 Personal and Family and raised the US price by $3 a month — the first increase since the product’s release.

Microsoft told the SEC that AI is pushing a margin DOWN. In the quarter ended 30 September 2025 Microsoft Cloud gross margin percentage decreased to 68 per cent, driven by scaling AI infrastructure and growing usage of AI features.

Amazon told the SEC that AI is shortening the life of its equipment. Effective 1 January 2025 it cut the estimated useful life of a subset of servers and networking equipment from six years to five, citing the increased pace of technology development, particularly in artificial intelligence and machine learning — reducing 2025 operating income by roughly $0.7 billion.

In the same filings Amazon moved a different useful-life estimate the other way, extending heavy equipment from ten years to thirteen and estimating that this would INCREASE operating income by about $0.9 billion.

Amazon’s operating cash flow rose $23.6 billion to $139.5 billion while its free cash flow FELL 71 per cent to $11.2 billion, because purchases of property and equipment rose from $77.7 billion to $128.3 billion.

Across 27 years of Spanish manufacturing firms, robot adopters raised output by 20 to 25 per cent within four years, RAISED their own employment by about 10 per cent, and cut their labour cost share by 5 to 7 percentage points — while average wages did not move at all. The gain reached workers as headcount, and never as pay.

The same study finds that firms which did not adopt lost employment. The adopters’ gain was in part market share taken from them.

A Danish study covering eleven occupations found precisely estimated zeros — no significant effect on earnings, hours or wages, with confidence intervals ruling out average effects larger than 1 per cent. Chatbots saved about 2.8 per cent of total work hours, and between three and seven per cent of that saving reached workers as pay.

Employment for workers aged 22 to 25 in the most AI-exposed occupations is about 19 per cent below where it would be had it kept pace with similarly aged workers in less-exposed occupations. Pay for those who keep these jobs has not visibly fallen — the entry door narrowed instead.

The hires rate has been falling since mid-2021. It stood at 4.6 per cent in July and again in November 2021, and had already fallen to 4.0 per cent by September 2022 — before ChatGPT was released.

Paralegals, one of the most task-exposed occupations, grew about 11 per cent. The two steepest declines among exposed occupations have the strongest non-AI explanations: computer programmers were affected by an occupational reclassification, and telemarketers by call-blocking legislation.

The International Labour Organization revised its own global exposure index. Working Paper 140, published on 20 May 2025, supersedes the widely quoted 2023 figures; the top exposure band moved from 2.3 to 3.3 per cent of global employment while the broad middle band fell from 29.2 to 20.5 per cent. The authors describe their earlier scores as overly optimistic about automation.

A calibration by one of the field’s leading economists puts the total gain in US total factor productivity from AI at no more than about 0.66 per cent over ten years, with the capital share of national income rising about 0.38 percentage points.

In a writing experiment, generative AI cut time on task by about 40 per cent. In the Danish payroll data, the same class of tool saved about 2.8 per cent of total work hours. Forty per cent of a task is not forty per cent of a job.