Google's ATLAS study analyzed nearly 15 million AI interactions across more than 150 countries and found that, for the non-routine cognitive work where AI is used most heavily, fewer than 10% of conversations attempted to automate a task end-to-end. For now, the real-world picture looks much more like humans working with AI than AI simply replacing them.

Google has released an unusually broad picture of how people are using its AI products, and the pattern is less autonomous than the loudest predictions suggest.

The ATLAS v1.0 study examined almost 15 million de-identified interactions across more than 150 countries and territories. In the non-routine cognitive work that accounted for most observed workplace use, fewer than 10 per cent of conversations appeared to ask the model to automate a task from beginning to end. Most looked collaborative: people sought ideas, information, analysis, drafts or help with one part of a larger job.

This is one company’s research release, not settled consensus about the whole labour market. It describes activity on selected Google products during two weeks in April 2026. It does not show that jobs are safe, that every interaction improved the work, or that AI systems outside Google are being used in the same way.

What Google actually measured

The ATLAS working paper, published on 23 July 2026, analysed 14,653,926 interactions recorded from 6 to 19 April. The sample covered the Gemini app, Google AI Mode and the Gemini API. An automated pipeline classified the conversations and connected work-related activity with more than 800 occupations and roughly 4,000 tasks across 140 languages.

Only around 14 per cent of all observed interactions were classified as work-related. That boundary matters because this was not primarily a workplace survey. Personal learning, household questions and other non-work activity formed more than 86 per cent of the sample. Google’s own summary of the project describes a technology embedded in everyday life as much as in paid employment.

Within work, use was broad but relatively shallow. The researchers observed AI activity in 68 per cent of occupations, representing just over 88 per cent of US employment. Yet in the median occupation where any use appeared, AI touched about 21 per cent of tasks. Only 3 per cent of occupations showed use across more than three quarters of their tasks.

The 10 per cent result has a specific boundary

Non-routine cognitive tasks, including analysis, writing, planning and problem-solving, make up roughly 35 per cent of professional tasks in the economy but almost 65 per cent of the work-related AI interactions in ATLAS. This is where conversational models are currently finding their easiest fit.

Even there, fewer than one conversation in ten was classified as an attempt to automate the whole task. Users more often asked the model to contribute. The difference is visible in ordinary work: requesting a first draft is not delegating publication; asking for a summary is not making the decision that follows it; generating code is not testing, integrating and maintaining a system.

ATLAS measures the apparent intent of the interaction, not its result. It cannot tell us whether the output was correct, whether a worker used it, how much time it saved, whether the task was eventually completed, or whether an employer later removed a role. “Attempted to automate” is the paper’s careful phrasing. It should not be silently upgraded to “successfully automated”.

Collaboration can still alter a job

A collaborative pattern should not be confused with a harmless one. If a model handles 20 per cent of a job, an employer may still change staffing, output targets or the kind of experience it is willing to pay for. Partial automation spread across many people can affect labour demand without any single conversation handing an entire occupation to a model.

Our earlier coverage of a generative-AI assistant used by 5,179 customer-support agents is a useful example. The average productivity gain was concentrated among novices, while experienced workers gained little. The tool did not replace the agent. It changed where valuable knowledge sat and how quickly a newer employee could approach veteran performance.

A tool can be assistive in interface and consequential in practice.

Another Silicon Canals article on Stanford payroll research described disproportionate employment pressure on workers aged 22 to 25 in highly AI-exposed roles. That evidence remains observational, and it cannot assign every hiring change to AI. It still closes off an easy misreading of ATLAS: humans can remain in the loop while the entry-level rung weakens.

Workflow fit is the harder problem

There is a practical reason collaboration dominates. Most work is not a clean prompt followed by a finished product. It passes through permissions, source checks, customer context, revisions, exceptions, legacy software and accountability. A model may handle one visible fragment while the rest of the system remains stubbornly human and organisational.

This sits neatly beside Silicon Canals’ earlier look at MIT NANDA’s report on enterprise generative-AI initiatives. The reported obstacle was often not access to a capable model but poor workflow fit, weak feedback loops and ill-defined operational problems. A fluent answer is only one component of useful work. It still has to arrive in the right system, draw on permissible data, survive review and produce an outcome the organisation can measure.

This explains why usage can spread quickly without equivalent organisational returns. An employee can ask for help in seconds. Rebuilding a process around autonomous software requires data access, controls, escalation routes and accountability when the system is wrong.

ATLAS therefore captures a period in which people can readily recruit models into their work, while many organisations have not rebuilt entire processes around them.

What the study cannot see

The exclusions are substantial. Google Workspace and Gemini Enterprise conversations were not included. Paid Gemini API content was not part of the main text analysis, and neither were file attachments. Those gaps are especially relevant when the question is end-to-end automation, because corporate integrations and agentic workflows are more likely to appear in enterprise systems than in an ordinary consumer chat.

The sample covers two weeks and only Google surfaces. It does not capture ChatGPT, Claude, Copilot or private models, nor the review and integration that happen outside the conversation.

The classification pipeline used large language models and clustering to infer what people were trying to do. Independent domain experts reviewed samples of the resulting work and non-work clusters, but the exercise still turns messy human requests into categories. It is an ambitious measurement system, not a direct census of completed tasks.

Google calls ATLAS v1.0 an early view. That is the right frame.

A more precise argument about replacement

The dataset weakens one simple claim: that today’s conversational AI is already being used mainly as an autonomous substitute for human cognitive work. On these Google products, during this period, it was not.

It does not settle the larger argument about employment. Collaboration can raise productivity, transfer expertise, reduce demand for junior work and give employers reasons to reorganise teams. It can also fail because the tool does not fit the process. All of those outcomes can coexist with a human still typing the prompt.

The next ATLAS releases will be more informative if they show whether the automation share rises, which occupations move first and whether apparent intent becomes completed, reliable work. For now, Google’s large snapshot looks like people working with models inside human processes, not models quietly completing whole jobs on their own.