Research

How Fast Is AI Making Switzerland?

Mar 10, 2026

Key findings

  • Tasks that would take 2.5 hours become 14-minute conversations. Across 5,097 Swiss Claude sessions, the mean estimated speedup is 10.5x.
  • Efficiency gains are consistent across cantons — from Zürich (10.3x, n=2,229) to Bern (8.9x, n=353), every canton above the sample threshold shows at least an 8.9x speedup.
  • The median tells a different story. At the median, tasks drop from 75 minutes to 10 — a 7.5x gain. The gap between mean and median suggests a long tail of complex tasks where AI saves the most time.

The data

In January 2026, Anthropic released the fourth version of its Economic Index — a dataset of roughly one million Claude.ai conversations from November 13–20, 2025. Each conversation was classified by Claude itself using privacy-preserving methods, estimating variables like how long a task would take a human alone, how long it took with AI assistance, and the degree of autonomy delegated to the model.1

The dataset includes geographic breakdowns. Switzerland contributes 5,097 conversations, split across 19 cantons. Two of the economic primitives — human_only_time (estimated in hours) and human_with_ai_time (estimated in minutes) — let us measure the perceived efficiency gain directly.

10.5x
Mean speedup
150 min → 14.3 min
7.5x
Median speedup
75 min → 10 min
136
Minutes saved (mean)
Per conversation
Figure 1. Swiss national efficiency metrics. Human-only time is converted from hours to minutes for comparability. N = 5,097 conversations, Nov 13–20, 2025.

The mean speedup of 10.5x means that, on average, what Claude estimates would take a competent professional 2 hours and 30 minutes took roughly 14 minutes with AI assistance. The median is lower at 7.5x — a typical task drops from 1 hour 15 minutes to 10 minutes. The gap between mean and median reflects a right-skewed distribution: some tasks involve substantial human-only time estimates (research, complex coding, multi-step analysis), and these are precisely where AI collaboration saves the most time.

Where AI is actually being used

The efficiency gains are consistent across cantons. But the distribution of who is using AI is not. Of the 5,097 Swiss conversations in the dataset, nearly half come from a single canton.

43.7%
of all Swiss Claude conversations come from Zürich alone

Zürich’s dominance is partly explained by economics: the canton is home to Switzerland’s largest concentration of tech companies, financial institutions, and ETH Zürich. But its 44% share of AI usage far exceeds its roughly 18% share of the Swiss population. The top five cantons — Zürich, Geneva, Vaud, Bern, and Basel-Stadt — account for 76.7% of all conversations.

Figure 2. Share of Swiss Claude AI conversations by canton. Darker shading = higher share of national usage. Grey = no data in dataset. Seven small, rural cantons have no recorded conversations.

Seven cantons — Appenzell Innerrhoden, Appenzell Ausserrhoden, Glarus, Jura, Nidwalden, Obwalden, and Uri — have no recorded conversations at all. These are small, rural cantons, but their absence from the dataset is still striking. If AI literacy is built through exposure, and exposure is concentrated in a handful of urban centres, the risk is clear: Zürich and the other major hubs are building AI fluency at a pace that the rest of the country cannot match.

This matters because the efficiency gains are real everywhere they are measured. The speedup is not a Zürich phenomenon — Bern, the lowest among high-sample cantons, still shows an 8.9x gain. The bottleneck is not whether AI helps, but who gets to use it. A canton where almost no one interacts with AI tools is a canton that will struggle to adapt when AI reshapes the skills its workforce needs.

A note on methodology. These are Claude’s own estimates of how long tasks would take with and without AI. They reflect the model’s assessment, not ground truth. Anthropic notes that human_only_time is estimated in hours by design, while human_with_ai_time is in minutes — the unit gap itself encodes an assumption about the scale of the difference.2 The data comes from a single week in November 2025 and reflects Claude.ai usage only, not all AI tool usage in Switzerland.

What this doesn’t tell us

Efficiency is not productivity. A 10x speedup on the wrong task is still wasted time. And a single platform’s data cannot stand in for the full picture of AI adoption. Swiss professionals also use ChatGPT, Copilot, Gemini, and a growing list of domain-specific tools that are invisible here.

But the concentration pattern is likely to generalise. If nearly half of one platform’s Swiss usage comes from Zürich, it would be surprising if the other platforms told a fundamentally different story. The question is not whether AI is useful — the 10.5x speedup answers that. The question is whether Switzerland’s AI dividend will be shared across cantons, or whether it will widen the gap between those who are already ahead and those who are not yet in the game.

• • •
Notes
  1. Anthropic, The Anthropic Economic Index: Insights from Claude.ai, v4, January 2026. The dataset uses privacy-preserving classification where Claude assesses each conversation against the economic primitive definitions from Table 2.1 of the report. ↑
  2. The classifier prompt asks for human_only_time in hours and human_with_ai_time in minutes. See data_documentation_v4.md and the economic primitives report, Table 2.1, for the full classifier definitions. ↑