Google Says 86%. Anthropic Says 35%. Both Are Right.

Both numbers are from this summer’s research. Google’s 86% is the share of AI interactions happening outside of work. Anthropic’s 35% is the share of conversations that are personal on a typical Tuesday. Same question, 51 points apart.

Neither lab is wrong.

Three major AI labs have now published serious research into how people actually use their AI. Read one at a time, each looks like a census. Read side by side, they contradict each other on the single statistic everyone quotes.

Here’s what each actually says, and what it should change about your AI marketing strategy.

What Google’s ATLAS found

ATLAS is the broadest of the three by a wide margin: 15 million aggregated, de-identified interactions across the Gemini app, AI Mode in Search, and the Gemini API. Those surfaces serve more than a billion monthly users between them. It spans 150+ countries covering 99% of the world’s population, 140 languages, 800 occupations, and 4,000 individual tasks.

The workplace findings stack up in a specific order, and the order matters:

• AI shows up in 68% of all occupations
• Those occupations represent 90% of total US employment
• But within a typical job, AI touches only about 21% of tasks

Broad reach, shallow penetration. AI has arrived nearly everywhere and taken over almost nothing.

This is what I see as well: very few companies have successfully automated anything meaningful. Fewer than 10% of work interactions fully automate a task. The vast majority are collaborative: ideation, strategy, information retrieval, learning. Non-routine cognitive work (creative design, testing) shows up at roughly 65% of AI interactions versus 35% in the baseline economy. People reach for AI disproportionately for the parts of their jobs that are hardest to script.

Trades and technical workers are using it too, mostly for diagnostics and troubleshooting, and they’re about 2x more likely to use image and video input. If you’ve written off trades and field work as an AI-free zone, I strongly disagree. It’s one of the areas I see the most adoption.

Then the headline: more than 86% of interactions in ATLAS happen outside of work. Household administration. Government services. Purchases, appliances, taxes, licensing, fines. The unglamorous friction of being an adult.

What Anthropic’s Economic Index found

Anthropic’s Economic Index is narrower in surface area and deeper in method. The June 2026 edition, Cadences, samples conversations continuously between April 10 and June 10, 2026, and links roughly 9,700 survey respondents to their actual usage. It covers Claude.ai, Claude Desktop, Claude Code, and first-party API traffic.

The picture is close to the inverse of Google’s.

Personal conversations run around 35% on weekdays, rising to just under 50% on weekends. This is a work-dominant tool and the weekly rhythm shows it. Anthropic can see the calendar in the data: tax questions are eight times more common on April 14 than the May average, news requests peak at 7 a.m., recipe questions run 2.3x higher at 6 p.m.

93% of conversations produce an identifiable artifact. Here are the top three:

Again: assistance, not replacement.

The Index’s longest-running metric is its automation-versus-augmentation split. In the November 2025 sample, augmentation led at 52% of Claude.ai conversations against 45% automation. That’s a reversal from August 2025, when automation was ahead 49% to 47%. The February 2026 report, Learning Curves, found automation on the first-party API had “decreased sharply.”

Consumer use is diversifying while API use is concentrating. The top 10 tasks accounted for 19% of Claude.ai traffic in February 2026, down from 24% in November 2025. On the API the top 10 moved the other way: 33%, up from 28% the previous August. People are finding more things to do with the chat interface but developers are industrializing a narrowing set of jobs.

Experience compounds. Users six or more months past signup show roughly 4 percentage points higher task success rates, after controlling for task type, model, language, and country. They’re also 7 percentage points more likely to be using it for work.

Anthropic is the only one of the three that asked people how they feel about it. About 10% rated losing their own job as likely or very likely. But more than a third put the odds of a junior colleague losing their job above 60%. Meanwhile 57% say their skills are becoming more valuable and 68% say they’re learning more.

Credit to Anthropic who states the limitation of their data plainly. Its survey respondents aren’t representative. Computer and mathematical occupations make up 30% of respondents versus 4% of US employment.

What OpenAI’s research found

OpenAI’s contribution is an NBER working paper published September 2025 with Harvard economist David Deming, drawing on 1.5 million conversations, plus a Q1 2026 refresh published in May.

Its most-cited framework splits usage three ways:

People value the tool most as an advisor, not a task executor. Three-quarters of conversations fall into practical guidance, seeking information, and writing. Writing is the most common work task. Coding and self-expression stay comparatively niche on consumer plans.

On the work split: roughly 30% work-related, 70% non-work, with non-work having grown from 53% of messages in June 2024 to 73% by June 2025.

The demographic story is most interesting. In January 2024, 37% of users with classifiable names had typically feminine names. By July 2025 that had passed half at 52%, and it kept climbing through Q1 2026. Adoption growth in the lowest-income countries was running more than 4x the rate in the highest-income countries by May 2025. The fastest-rising countries by messages per capita in Q1 2026 were the Dominican Republic, Haiti, Japan, Mexico, and Tanzania.

OpenAI is explicit about its blind spot. The analysis covers consumer plans only. It excludes Codex and all enterprise and education products, and therefore “understates total workplace and educational usage.”

The same question, three answers

One caveat: the units aren’t identical. Google counts interactions, Anthropic counts conversations, OpenAI counts messages.

Where all three agree

1. Assistance is winning, decisively. Under 10% full automation at Google. Augmentation ahead of automation at Anthropic. Asking ahead of Doing at OpenAI. Three companies, three methods, one agreement: today we are not seeing mass replacement.

2. Reach is broad, depth is shallow. Google’s 21%-of-tasks figure is the cleanest version, but Anthropic’s finding that consumer task concentration is falling says the same thing from another angle. People are using these tools for more things, a little at a time. Penetration to work and society takes time.

3. The early-adopter era is over. OpenAI’s gender gap closed. Google covers 68% of occupations. Anthropic sees management traffic growing. Whoever you sell to is now plausibly a regular AI user.

Where they contradict, and why

Now the interesting part. Why does Google see 86% non-work when Anthropic sees 35%?

Because none of these reports measures how people use AI. Each one measures how people use that vendor’s product, and those aren’t remotely comparable.

Google’s ATLAS includes AI Mode, which lives inside Search. Search is the highest-volume consumer surface on the internet, and the overwhelming majority of search behavior has always been personal. Google’s 86% is, in large part, a statement about Search. It reflects Google’s distribution, not humanity’s priorities.

Anthropic’s customers are businesses, and Claude Code is a professional developer tool. There’s no consumer search product feeding the denominator. A work-heavy result is exactly what we should expect.

OpenAI sits between the two and says so out loud: consumer plans only, excluding Codex and enterprise, which “understates total workplace usage.”

So the right reading of these three reports isn’t “AI is 86% personal” or “AI is 65% work.” It’s this:

There’s no such thing as “how people use AI.” There’s only how people use a particular category of AI tool. The reported numbers tell you more about the vendor’s distribution than about user behavior.

Any statistic you see this year claiming to describe universal AI usage is describing a product’s install base. Ask which surface it came from before you build a strategy on it.

The geography puzzle

The geographic findings look contradictory. They aren’t. They answer three different questions, and separating them is worth two minutes of your time.

Levels. Google: usage mirrors national wealth. Rich countries use more AI per person right now. True.

Growth rates. OpenAI: the lowest-income countries were growing 4x faster than the highest-income ones. Also true, and entirely compatible. Growth rates get computed off small bases.

Concentration. Anthropic: the top 20 countries account for 48% of per-capita usage, up from 45%, so country-level concentration is increasing. Yet inside the US, the top 5 states’ share fell from 30% to 24%, with projected convergence in five to nine years.

What to do with this

Run the 21% audit. Google’s finding that AI touches roughly a fifth of tasks in a typical job is a benchmark you can actually test. List your team’s recurring tasks. Not tools. Tasks. Mark which ones involve AI today. If you’re well under 21%, you’ve got room. If you’re well over it and not seeing output gains, you’ve got an education problem.

Stop treating vendor usage stats as market sizing. When a deck cites “86% of AI use is personal,” the right question is which surface produced that number. For most B2B marketers, Anthropic’s work-dominant profile is the relevant comparison set and Google’s is close to irrelevant.

Build for the collaborative middle. Under 10% of work interactions fully automate anything. If your product, your content, or your internal pitch is built on full autonomy, you’re targeting the smallest slice of real behavior. The volume sits in the workflows of ideation, drafting, explanation, and review.

Take the life-admin category seriously. Taxes, licensing, appliances, fines, government services. Google found enormous volume in exactly the friction categories consumer brands usually ignore. If you sell anything with a complicated onboarding, purchase, or support process, your customers are already asking an AI tool for help with it. That conversation’s happening whether or not your content is in it.

Check your non-English and multimodal assumptions. Two-thirds of AI conversations aren’t in English. Trades and technical workers are 2x more likely to use image or video input. Text-only, English-only content strategy is now a deliberate narrowing, not a default.

Near-term

The near-term story isn’t mass job replacement. It’s AI quietly removing slices of work and life friction, one task at a time. 21% of the tasks in a job. Under 10% of those fully automated. Mostly inside a collaborative workflow, where a person is still driving.

That’s a less dramatic story than the one being sold by your local tech bro. It’s also the one with three independent datasets behind it.

Pull up your team’s recurring task list and mark the AI-assisted ones. Whatever percentage you land on is the most useful number you’ll get about your own AI adoption this year.

Sources

• Google, Understanding the AI economy (ATLAS) (July 23, 2026)
• Anthropic, Economic Index report: Cadences (June 2026)
• Anthropic, Economic Index report: Learning curves (March 2026)
• Anthropic, Economic Index: New building blocks
• OpenAI, How people are using ChatGPT (September 15, 2025)
• OpenAI, How ChatGPT adoption broadened in early 2026 (May 11, 2026)
• Chatterji et al., How People Use ChatGPT, NBER Working Paper 34255