More than 95% Working Professionals in US, UK & EU Use AI Regularly, 76% Pay for Premium Tools

More than 95% of working professionals across the US, UK, and EU use AI for their professional or personal tasks, according to the State of AI Report.

Authored by Nathan Benaich, General Partner at Air Street Capital, the report surveyed 1,183 participants between July and September 2025. Over 90% of respondents were highly educated professionals from start-ups, public companies, and academia across the US, UK, and EU.

The report stated a striking 76% of respondents said they personally pay for at least one AI service (such as ChatGPT Plus, Claude Pro, or Perplexity Pro). On impact, 92% reported measurable productivity improvements, while 47% described those improvements as “large” or “transformative.”

APIs Over Self-Building

The report shows that companies are overwhelmingly choosing API-first approaches over building or fine-tuning large models in-house, even as organisations scale up AI budgets and confront practical barriers such as integration time, data privacy and scarce expertise.

The study, which also lays out ten headline predictions for the coming year and a technical tour of advances in reasoning, adaptation and agentic systems, signals a market maturing from experimentation to operational stress-testing.

It suggested that the clearest commercial trend is procurement-driven adoption. Most teams buy models via APIs rather than investing in full in-house model builds or fine-tuning pipelines.

When groups do fine-tune, toolchains centre on PyTorch, Hugging Face and LoRA. Popular consumer tools cited by practitioners include ChatGPT, ElevenLabs, Perplexity and Claude. Crucially, more than 70% of respondents reported year-on-year growth in their organisation’s AI budget, underscoring a shift from pilot projects to sustained investment.

Reasoning & Open vs Closed

On the capabilities front, the report highlights renewed progress in model reasoning driven by methods that use test-time compute, chain-of-thought-style supervision and reinforcement from verifiable rewards (RLVR).

Labs such as o1 and DeepSeek R1 have pushed “reasoning” methods that sometimes improve math and coding performance, although gains can be fragile and sensitive to evaluation setups. Branching or parallel-reasoning approaches, which spawn multiple inference threads and merge outcomes, are showing promise at reducing hallucinations and leveraging parallel hardware for long-context tasks.

Closed, heavily resourced model families (GPT-5-type families, o3 variants, Gemini, Claude, Grok 4) still lead on average, but strong open-weight contenders, DeepSeek, Qwen, Kimi and others, have narrowed performance gaps, helped by China’s fast-moving open ecosystem.

That competition is reshaping innovation incentives and accelerating practical adoption of open models.

Predictions

The report lists ten high-impact predictions insiders expect to see within a year.

Standouts include a major retailer deriving more than 5% of sales from an “agentic” checkout, a top lab choosing open-sourcing for strategic/political advantage, an open-ended agent making a meaningful end-to-end scientific discovery, a deepfake/agent cyber incident prompting NATO/UN debate, and a generative, real-time video game exploding on Twitch.

The set mixes realistic commercial inflections with lower-probability, high-consequence technological and geopolitical events.