Somebody finally counted.
For the past few years, most conversations about AI adoption have run on vibes. Someone says AI is everywhere. Someone else says nobody they know actually uses it. Both people cite anecdotes. Nobody has numbers. Google’s AI & Economy ATLAS is an attempt to change that, and the September update added new data visualizations to make its findings easier to read.
I want to walk through what this thing is, why I think it matters more than the average product announcement, and what it does not tell us yet.
What ATLAS actually is
ATLAS launched in July 2026 as an ongoing, large-scale, de-identified study of how people use AI at work and in everyday life. Two details in that sentence do a lot of work.
“Ongoing” means this is not a one-off report that gets stale in four months. It is designed to keep running, which means we should eventually get trend lines rather than snapshots. That difference matters. A single measurement tells you where things stand. A series tells you where things are going.
“De-identified” means the data has been stripped of details that tie it back to individuals. If you are going to study behavior at this scale, that is the part you want to see stated up front.
The work sits inside Google’s Chief Economist’s Office, with Zanna Iscenko as AI & Economy Lead and Scott Strand as Head of StratOps. That placement is a signal in itself. This is not a marketing project dressed up as research. It is economists asking economic questions about a technology that everyone keeps predicting will reshape work without much evidence about how.
Why a usage study beats another benchmark
Here is my honest bias as someone who explains AI agents for a living. I have grown tired of benchmark scores. A model gets a higher number on a test, the number gets posted, and none of us learn anything about whether the thing helps a bookkeeper close her month faster.
Usage data is different. It answers questions people actually have:
- Which tasks do people hand to AI, and which do they keep?
- Does adoption look the same across countries, or wildly different?
- Is this a work tool, a personal tool, or both?
- Where does usage stall out after the novelty fades?
Those questions are harder to answer than a benchmark, and far more useful. If you are deciding whether to bring AI tools into a small team, “what do people like me actually use this for” is the only question that matters.
The visualization update is not cosmetic
The September change made ATLAS data easier to explore visually. That sounds like a minor thing. I do not think it is.
Research that lives in a PDF reaches researchers. Research you can click through reaches everyone else, including the manager trying to make a decision on a Tuesday afternoon without a data science background. Making findings legible to non-specialists is how research actually influences behavior. I have watched plenty of good studies sink because nobody outside the field could read them.
So when a study aimed at understanding the AI economy invests in making its own data browsable, I read that as the team wanting to be understood rather than cited.
The model updates in the background
ATLAS is not landing in a vacuum. The same stretch brought updates across Google’s Gemini models and NotebookLM gaining integration with design tools. That second one is quietly interesting to me. NotebookLM started as a way to reason over documents you supply. Connecting it to design tools pushes it toward being part of a workflow rather than a destination you visit.
That shift, from tools you go to toward tools that live where you already work, is the direction I keep seeing. And it makes measurement harder. When AI is a chatbot you open, usage is easy to count. When it is stitched into the software you already use, “did you use AI today” becomes a question people cannot reliably answer about themselves. Studies that measure activity rather than self-reported opinion will hold up better as that happens.
What I am still waiting for
I will not pretend ATLAS settles anything yet. A first iteration is a first iteration. The interesting findings come from the second, third, and tenth updates, when you can compare and see what moved.
There is also the obvious limitation. This is Google studying usage of a space Google competes in. That does not make the data wrong. It does mean the ideal version of this research is several organizations running similar studies so findings can be checked against each other.
Still, I would rather argue about real data than trade anecdotes. If you have been trying to figure out whether AI adoption is genuine or hype, a study built to keep measuring is the most useful thing to come out of this cycle. Go poke at the visualizations. Bring your own questions.
🕒 Published: