Google’s ATLAS found broad workplace AI use. Task coverage was still shallow.

Google’s first AI & Economy ATLAS study found workplace AI use in occupations representing 88.4% of U.S. employment. Among occupations with at least one task above Google’s 25-user threshold, the median occupation showed AI activity in 21% of its tasks. Only 3% of occupations reached task saturation in at least three-quarters of their tasks.

Those figures support a more careful business decision than the broad adoption headline suggests. Occupational reach shows that AI has found entry points across much of the economy; it does not show that most workers use AI, that whole workflows have changed, or that roles are ready for automation. Separating occupational reach, task saturation, interaction intent, and business outcome makes it easier to decide what to measure before expanding access, budget, or autonomy.

ATLAS measures patterns in three Google products

Google published its overview of ATLAS on July 23, 2026. ATLAS stands for Activity, Task, Landscape, and Adoption Study. The accompanying 100-page ATLAS v1.0 report analyzes about 15 million aggregated, de-identified interactions across the Gemini App, Google AI Mode, and Gemini API, then maps them to more than 800 occupations and 4,000 tasks. The broader dataset spans more than 150 countries and 140 languages.

The detailed methodology places the sample between April 6 and April 19, 2026. Google sampled approximately five million interactions from each included surface that passed basic data-cleaning and materiality filters, then reweighted the three samples by each surface’s overall share of interactions. It removed identifying information, summarized conversations, grouped similar summaries into anonymized clusters, and used automated classifiers to associate those clusters with occupations and tasks from the U.S. O*NET taxonomy.

That design gives Google a wide view of how its products are being used, with privacy protections that prevent researchers from following a person through a complete work process. Occupation and task assignments are probabilistic inferences from conversation data. Google says its granular findings carry more uncertainty than broad occupation-group patterns.

The sample boundary matters as much as its scale. ATLAS v1.0 covers users of three Google surfaces, so it is not a representative census of all workers or all AI products. Paid Gemini API content and Google Cloud enterprise usage are excluded from the taxonomized data, which may understate professional enterprise use. Several other Google products, including Workspace and AI Overviews, are also outside the dataset. Current non-users and work that never appears in a Google interaction are absent by design.

Occupational reach and task saturation answer different questions

Occupational reach asks where any meaningful use appears. Google counted a detailed occupation as observed when at least 50 users globally were associated with its tasks. Using that threshold, AI use appeared in 68% of detailed occupations. When Google mapped those occupations to U.S. employment data, they accounted for 88.4% of employed civilian workers.

That 88.4% is an employment-weighted description of the occupations in which Google saw use. It does not mean that 88.4% of U.S. workers used Gemini during the study period. A small number of users can place a large occupation inside the observed group, so the measure establishes reach across the economy without establishing adoption within each workplace.

Task saturation moves one level deeper. Google counted an O*NET task when at least 25 users globally appeared to use Gemini for it, then calculated how much of each occupation’s task list crossed that threshold. Among occupations with at least one observed task, median saturation was 21%. About 30% of occupations reached at least a quarter of their tasks, 11% reached half, and 3% reached at least three-quarters.

The contrast between 88.4% employment coverage and 21% median task saturation is the central operational finding. AI use is broadly distributed across occupations and still concentrated in a minority of tasks within a typical observed occupation. A company considering role-level automation needs evidence about the work inside the role, because an occupation-level presence cannot supply it.

Interaction intent shows collaboration more often than end-to-end automation

ATLAS goes deeper by estimating what users wanted AI to do. Google’s preliminary intent classifier sorted anonymized conversation-cluster summaries into task automation, partial drafting and generation, review and refinement, ideation and strategy, or information retrieval and learning. It defined task automation as asking AI to execute the core task or a major sub-task end to end.

For non-routine cognitive work such as hypothesis testing and creative design, less than 10% of conversations were classified as seeking task automation. Use centered instead on drafting part of an output, reviewing existing work, developing ideas, and retrieving information. These activities can improve a task, but they usually leave a person to supply context, integrate the output, make a decision, or complete the downstream action.

Routine cognitive work showed a different pattern. More than a quarter of those conversations were classified as targeting task automation. That result supports closer inspection of structured, repeatable tasks, while its scope should stay exact: it applies to conversations classified as routine cognitive work, not to all work interactions or all routine jobs.

The classifier estimates the role users asked Gemini to play from summaries that can contain more than one intent. Google cannot observe the activity outside Gemini, and it describes this classification as preliminary. The result is useful evidence about the collaboration and automation mix, but it is not direct observation that an end-to-end workflow ran successfully.

Manual and technical occupations make the same distinction visible. Google observed automotive technicians and industrial mechanics using AI to interpret test results, debug electrical wiring, and inspect machinery for wear. Their share of multimodal conversations was more than twice the average for work interactions. These are meaningful uses in physical work, yet the interaction data does not show whether the diagnosis was correct, the repair was completed, or the asset returned to service.

Business outcomes remain outside the study

The fourth layer is the one a manager needs for a rollout decision, and ATLAS does not measure it. The report states that a completed conversation does not establish that the user completed the intended task, saved time, or produced measurable economic value. It also does not measure output quality, rework, downstream errors, or whether the interaction added another review step.

Even a higher rate of end-to-end task automation would still fall short of job automation. A role contains complementary tasks, coordination with other people and systems, exception handling, and organizational friction. Google makes this boundary explicit in the report. The same boundary applies inside a company: a technical completion can leave the surrounding workflow unchanged.

Usage data still has a purpose. It can reveal demand, show where employees are experimenting, and identify tasks worth studying. It becomes risky when leaders treat conversation volume, active seats, or occupational coverage as proof that work improved. Our related analysis of AI assistant stickiness explains why engagement metrics need to be paired with evidence from completed work.

Measure one task before widening the rollout

A local evaluation can be much smaller than ATLAS because it has a different job. Start with one recurring task, record the AI’s current role, name the work a person still completes, and choose outcome evidence that can be compared with the prior process. Retrieve, ideate, draft, review, and automate are useful role labels because they describe different amounts of work and risk.

The examples below are illustrative. They show the level of specificity needed before a team expands a tool or gives it more autonomy; they do not claim measured results.

Scroll sideways to see all 4 columns.

Task	Current AI role	Remaining human work	Outcome evidence
Prepare a response to a billing dispute	Retrieve policy and draft the response	Verify account facts, decide any exception, approve and send	Cycle time, policy errors, rewrite rate, reopened cases
Classify an incoming equipment issue	Retrieve guidance and suggest a diagnosis	Inspect the asset, run tests, choose the repair, document completion	Time to diagnosis, misclassification rate, repeat visits
Review a recurring supplier invoice exception	Compare records and draft an explanation	Resolve contract ambiguity and approve payment	Resolution time, rework rate, payment errors

Collect a baseline before changing the rollout, then compare the same task after AI enters the process. Access can expand when the evidence shows better task performance without unacceptable errors or hidden review load. Autonomy deserves a separate decision because drafting a useful output and completing a workflow carry different consequences.

Some tasks will expose judgment, missing context, or exceptions that the initial usage data did not reveal. In those cases, study the work under human control before automating it. Our guide to choosing a safer first AI automation explains how to examine that boundary without assuming the entire task is ready.

Expand from measured task improvement

ATLAS is strong evidence that people are bringing AI into many kinds of work, including roles far outside conventional office software. Its depth measures show that these uses remain selective, and its intent measures show that collaboration still dominates many complex tasks. The business case has to be completed locally with evidence from the task and workflow that a team wants to change.

A request for more licenses or automation budget should therefore name the task, the AI’s role, the remaining human work, and the result that improved. Occupational coverage and conversation volume can help locate an opportunity. They cannot justify the budget decision.

BaristaLabs helps teams connect one AI use case to the workflow and before-and-after evidence needed for a rollout decision. See our process automation work, or review one task before expanding AI access.