Google’s ATLAS study (Google: AI updates, July 2026: the AI & Economy ATLAS study, opens in a new tab) analysed roughly 15 million de-identified interactions across the Gemini app, Google AI Mode and the Gemini API. It found AI use associated with occupations representing 88% of U.S. employment (arXiv (Google): AI & Economy ATLAS: measuring AI use across occupations, opens in a new tab).
Google reached an equally important second conclusion: penetration within those occupations remains relatively shallow. People overwhelmingly use AI collaboratively, while end-to-end task automation remains limited.
The distinction that matters
What “broad but shallow” AI adoption looks like
It helps to separate two dimensions. Breadth is how many jobs and workflows AI touches. Depth is how much of each workflow AI actually performs.
Breadth has expanded quickly. Federal Reserve Board researchers reported that 41% of the workforce (Federal Reserve Board: Monitoring AI Adoption in the U.S. Economy, opens in a new tab) was using generative AI for work by November 2025, up nearly 10 percentage points from a year earlier, while 54% of workers (Federal Reserve Board: Monitoring AI Adoption in the U.S. Economy, opens in a new tab) were employed by companies using large language models.
Depth is a different story:
- A marketer develops an outline, researches competitors or rewrites a paragraph.
- A developer asks AI to explain an error or generate a function.
- An analyst writes a formula, summarises a report or interprets a dataset.
- A manager prepares meeting notes or drafts an email.
AI participates in all of those jobs. But the worker still defines the task, supplies context, evaluates the output, makes the decision and moves the work to the next step. That is augmentation, not end-to-end automation.
Google isn’t the only dataset showing this pattern
Anthropic found a similar pattern when researchers analysed more than four million Claude conversations, classifying 57% augmentation (Anthropic (arXiv): Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations, opens in a new tab) and about 43% automation (Anthropic (arXiv): Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations, opens in a new tab) — augmentation being work where humans and AI proceed together through learning and iteration, automation being requests AI completed with relatively little human involvement. Software development and writing alone accounted for nearly half of observed usage.
Microsoft’s workplace data points the same way. An analysis of more than 100,000 Microsoft 365 Copilot conversations found 49% cognitive work (Microsoft WorkLab: Agents, human agency, and the opportunity for every organization, opens in a new tab) such as analysing information, evaluating options, solving problems and thinking creatively, with another 19% working with people, 15% finding information and 17% producing work.
These patterns look much more like a highly capable collaborator than an autonomous employee.
Giving workers AI does not automatically redesign their jobs
One of the clearest demonstrations comes from a randomised field experiment involving 7,137 workers (arXiv: The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments, opens in a new tab) across 66 companies. Workers given generative AI inside their existing email, writing and meeting tools eventually spent about two fewer hours per week on email. Outside those individual productivity gains, researchers detected no significant change (arXiv: The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments, opens in a new tab) in the overall quantity or composition of workers’ tasks.
The workers became more efficient at pieces of their existing jobs. Their jobs did not fundamentally change. That helps explain the “broad but shallow” observation: the easiest way to adopt AI is to insert it into something a person already does.
Assistance pattern
Existing workflow → AI assistance → human review → existing workflow
Automation pattern
Trigger → AI reasoning → access to company context → tool use → action → validation → next action
The second model requires far more than a good language model.
Why full automation is harder
A chatbot is useful with almost no infrastructure: give it a prompt, receive an answer. Automation has to operate inside a system. To meaningfully automate a business process, an agent may need email, documents, analytics platforms, CRMs, project-management systems and databases — plus permissions, context, rules for when to act, error recovery and clear boundaries around what requires human approval.
That suggests the current bottleneck is increasingly organisational and architectural rather than simply model intelligence. Microsoft’s 2026 Work Trend Index supports that reading: in a survey of 20,000 workers (Microsoft WorkLab: Agents, human agency, and the opportunity for every organization, opens in a new tab), organisational factors such as AI culture, manager support and talent practices were more than twice as strongly (Microsoft WorkLab: Agents, human agency, and the opportunity for every organization, opens in a new tab) associated with reported AI impact as individual mindset and behaviour. Microsoft concludes that many employees are moving faster than the systems around them.
The models may be capable of more. Organisations have not necessarily rebuilt their workflows to let them.
The progression from AI assistance to AI automation
Rather than sorting organisations into “using AI” and “not using AI”, it is more useful to treat AI maturity as a progression.
Stage 1
Ask
The employee asks AI questions.
Research, explanations, brainstorming and troubleshooting.
Stage 2
Create
AI produces an artifact the employee can use.
An email, report, presentation, image, block of code or analysis.
Stage 3
Collaborate
The human and AI iterate on a task together.
Weighing alternatives, editing content, debugging software or working through a decision.
Stage 4
Act
AI can interact with external tools.
Retrieving analytics, updating a CRM record, creating a ticket or querying a database.
Stage 5
Orchestrate
AI coordinates multiple tools and steps within a workflow.
Find pages losing organic traffic, retrieve Search Console data, identify likely causes, pull keyword data, recommend changes, create optimisation tasks.
Stage 6
Automate
The workflow runs largely without human intervention.
Escalating only when judgement or approval is genuinely required.
Most workplace AI adoption today appears concentrated in the first three stages. The major shift ahead may come from stages four through six.
Connectors are part of the missing infrastructure
This is where AI connectors, APIs and agent tooling become important. A general-purpose assistant knows a great deal, but it does not automatically know what happened in your CRM yesterday, which pages lost traffic last week, which support tickets are unresolved or which campaign creative was approved. Connecting AI to those systems changes what it can do, which is why we document documented connector capabilities rather than simply listing integrations.
Instead of asking “how should I analyse an SEO traffic decline?”, a connected agent could pull declining pages from Search Console, compare traffic and conversion trends in analytics, retrieve keyword and competitor data, inspect the affected pages, diagnose likely causes, produce prioritised recommendations, open tasks in a project-management system and ask a human to approve high-impact changes. Our Daily AI SEO Analyst guide walks through exactly that workflow on official APIs, with a human reviewing every recommendation.
The intelligence matters. But the connections, permissions, context and workflow design are what turn intelligence into automation, and our verified connection records are the practical starting point for judging what is available today.
Agent adoption is beginning to accelerate
Shallow adoption should not be mistaken for permanently shallow adoption. Microsoft reports that active agents across its Microsoft 365 ecosystem increased 15x year over year (Microsoft WorkLab: Agents, human agency, and the opportunity for every organization, opens in a new tab) through March 2026, and 18x in large enterprises (Microsoft WorkLab: Agents, human agency, and the opportunity for every organization, opens in a new tab), alongside substantial variation in how deeply companies integrate those agents into workflows.
The first phase of enterprise AI was giving workers access to AI. The second embedded copilots into existing tools. The next connects AI to the systems where work happens and grants permission to perform increasingly complex sequences of actions. The useful question is shifting from “are employees using AI?” to “how much of the workflow can AI actually complete?”
What this means for businesses
The opportunity is no longer simply convincing employees to use ChatGPT, Gemini, Claude or Copilot; adoption is already spreading quickly (Federal Reserve Board: Monitoring AI Adoption in the U.S. Economy, opens in a new tab). The harder and more valuable work is identifying where an organisation can move from isolated assistance to connected workflows.
Start by looking for work that is:
- repetitive, but still requires some reasoning
- distributed across several software tools
- dependent on retrieving and synthesising information
- governed by relatively clear rules
- high-frequency enough that small time savings compound
- easy for a human to review when necessary
Then ask a more useful question than “can AI do this task?”: what prevents AI from completing the next step? Sometimes the answer is model capability. Increasingly it is missing context, missing integrations, missing permissions, poorly structured data, or a workflow designed entirely around humans. Solving those problems is what turns AI from a tool people use into infrastructure that work can run on, which is where connector ROI evidence becomes relevant.
The bottom line
The available evidence suggests workplace AI is currently far more transformative at the task level than at the job level. Google (Google: AI updates, July 2026: the AI & Economy ATLAS study, opens in a new tab) sees usage spanning occupations representing more than 88% of American employment while describing penetration as shallow. Anthropic (Anthropic (arXiv): Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations, opens in a new tab) similarly finds augmentation more common than automation. And a randomised workplace study (arXiv: The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments, opens in a new tab) shows meaningful individual time savings without a change in the composition of employees’ jobs.
That does not mean automation is not coming. It means the transition is happening differently than the simplest predictions suggested: first AI helps people perform tasks, then it gains access to tools, then organisations redesign workflows around it. Only after that does meaningful end-to-end automation become possible.
We are well into the first stage. The more consequential transition is what happens next, and it is decided by which assistant platform reaches the systems your work depends on.