Your AI Adoption Has a Shape. Anthropic Just Published It.
Anthropic published its latest Economic Index report on June 26th. It’s the first one built on a data pipeline that can see what a Claude session actually produces, not just what was said in it — a new classifier that labels the output of every conversation, higher-rate sampling, and the first results from a survey linked to real usage data across roughly 9,700 respondents.
I read it the way I suspect most people in this business read it: looking for the places where the data disagrees with what I see in the field, because that’s where the learning is. There were fewer of those than I expected. The report is, more than anything, a very large, very well-instrumented confirmation of a pattern that anyone spending their weeks inside mid-market businesses has been watching form for about eighteen months.
Here is that pattern, stated plainly. AI adoption is not happening at the level of the organization. It is happening at the level of the person. And the gap between the organizations that will compound that and the ones that won’t has almost nothing to do with which model they bought.
Adoption is a person, not a department
Start with the survey’s composition, because it tells you something before you read a single finding.
Computer and Mathematical occupations make up roughly 30% of survey respondents and 4% of US employment. Management is 23% of respondents against a 7% employment share — but accounts for only 4% of sessions. Transportation, food service, construction: badly under-represented, in the survey and in the usage data both.
The obvious read is that this is a sampling problem, and Anthropic says so directly. The more useful read is that the skew is the finding. This is not a random slice of the working world. It’s a self-selected population of people who went and got the tool themselves, without waiting for anyone to hand it to them. And it is wildly concentrated in a handful of occupations — plus a large cohort of managers who, tellingly, are using Claude for something other than managing.
That is exactly the shape of adoption in the accounts I work in. It is almost never a department. It’s two people. Sometimes one. Someone in ops who quietly automated the reporting pack, someone in sales who rebuilt their entire research workflow, a controller who now writes SQL she couldn’t write a year ago. They’re not on a rollout plan. Nobody asked them to. They’re four or five hundred sessions deep while the org chart around them hasn’t moved an inch.
Every mid-market business I walk into has these people. Most leadership teams cannot name them.
Where the value actually shows up
The second thing the report gets right is what the value looks like when it arrives, and it isn’t what most owners are budgeting for.
The classifier found that 93% of Claude conversations produce an artifact — a document, a piece of code, an explanation, an analysis. The most common are explanations (17%), documents and reports (15%), and guidance (11%). Narrow it to work conversations and documents and reports lead at 20%, followed by explanations at 9%, email drafts at 7%.
Read that as a cost-savings story and you’ll be disappointed, and so will your clients. Anthropic’s own survey backs this up: 86% of respondents report gains in speed, 82% in scope, 69% in quality — and only 27% report gains through cost savings on services they’d otherwise have bought.
Scope. That’s the one. When I ask a client six months in what they’d point to and say “that paid for itself,” almost nobody says “we cut headcount” and almost nobody says “we saved on the agency.” They say some version of: we’re doing work we simply could not do before, and we’re answering customers in a day instead of a week.
The report’s token analysis explains the mechanism better than I’ve managed to. Compute tracks the value of the work: conversations mapped to top-wage-tercile occupations consume 2.07 times the tokens of bottom-tercile ones. Marketing managers earn about twice what editors do and their conversations burn roughly 2.5x the tokens. Building an app costs more than three times the median conversation; a typical explanation costs about a fifth.
So when an owner tells me their AI spend is reassuringly low, I no longer hear discipline. I hear that nobody has pointed it at anything expensive yet. Cheap AI usage is a symptom, not an achievement.
And here’s the part that should settle the augmentation debate for anyone still having it. In higher-wage conversations, Claude produces 1.34x more output per turn while the user engages 1.53x more turns. More machine does not mean less human. They move together. The tool doesn’t get valuable when people step back — it gets valuable when people lean in on harder problems.
The two things that are actually blocking you
Now the part where the report and my field notes converge on something uncomfortable.
The blockers I run into most are not fear, and they are not leadership refusing to fund anything. They are habit and technical ceiling — people doing the new thing in the old shape, and people who don’t yet have the skill to ask for more.
Anthropic quantified the first one so cleanly that I’ve started putting the number in slides.
They compared how much autonomy people give Claude across surfaces, on a 1-to-5 scale. Claude Code sessions run +0.37 points more autonomous than chat and Cowork sessions on average, higher on 26 of the 31 output types measured. And the single most clarifying statistic in the entire report:
Same task. Same output. Same underlying models. Thirteen turns, or one.
Before you write that off as a coding-tools artifact, note that the gap survives when you hold the model constant — among Sonnet conversations, Claude Code still shows +0.26 more autonomy. Anthropic’s own conclusion: “the product used is likely more important than the underlying model.” And roughly two thirds of the gap comes not from a different mix of tasks, but from the same tasks executed with more delegation.
That is habit, measured. It is the single largest lever in the report and it has nothing to do with procurement.
Most of the people in your business are turn-taking with a system that is capable of being handed the whole job. They ask a question, evaluate, refine, refine again, thirteen times — running the tool like a faster search box, in the conversational shape they learned in 2023 and never revisited. The capability moved. The habit didn’t.
The second blocker — technical ceiling — shows up in the report’s most counterintuitive finding, though you have to read it sideways to see it.
People who use Claude in the most automated way — handing over whole tasks — are the most optimistic about AI’s effect on their own pay, job security, ability to find work, autonomy, meaning, and human interaction. Every one of the six dimensions. Not the augmenters. Not the careful ones. The delegators.
The intuitive story is that this is just selection: enthusiasts delegate more. Anthropic tests that and the effect holds after controlling for tenure on the platform. The more likely story is the one I’d bet my practice on: delegation is how you find out what the thing can do. People who hand over a whole task get direct evidence of the ceiling. People who never let go of the wheel are guessing about it forever — and, mostly, guessing low.
Which is why the experience gradient in the survey should worry every leadership team reading this. Workers with 15+ years of experience rate AI’s current capability roughly 10 percentage points lower than people in their first year. Those are the people who decide what gets automated, what gets funded, and what gets dismissed as a toy. Your most senior, most trusted, most influential people are systematically the ones with the least accurate picture — because they delegate the least, so they’ve seen the least.
How to accelerate it
Five things, in the order I’d actually do them.
Find your power users and put them on payroll for it. They already exist. Go find them — not by asking who’s “using AI,” which gets you a room full of hands, but by asking who has automated something nobody asked them to automate. Then give them a title, a budget line, and four hours a week that are explicitly theirs. The single cheapest acceleration available to a mid-market business is to stop letting its most advanced adopters do this work as a hobby.
Audit your artifacts, not your seats. Every business I know can tell me how many licenses it has. Almost none can tell me what came out. Pull last month’s work and sort it by what it actually produced. If the honest answer is that your organization’s AI output is overwhelmingly explanations — the 17% category, the one that costs a fifth of a median conversation — then you have bought a very expensive tutor. Nothing shipped. That’s a fine place to have started and a terrible place to still be.
Attack the thirteen turns directly. This is the highest-leverage training intervention available and nobody is running it. Take the workflows people already do conversationally and teach them to specify once and delegate. Not “prompt engineering” — scope engineering: writing a brief good enough that the work can run without you standing over it. This is a skill, it’s teachable in an afternoon, and it is worth more than any tool you’ll buy this year.
Move the surface, not just the model. The report’s own conclusion is that the product shapes delegation more than the model does. If your entire organization’s experience of AI is a chat box, you have capped its autonomy by architecture. Get the agentic surfaces — Claude Code, Cowork, scheduled agents — in front of the people whose work can bear them, and you will get the delegation shift for free, because the tool asks for it.
Aim it at expensive work on purpose. Tokens track value. So point it at the thing that is genuinely hard, genuinely valuable, and genuinely bottlenecked — the analysis nobody has time for, the code review nobody can afford, the research that gets skipped. The businesses getting real returns are not the ones who found the cheapest way to do what they already did. They’re the ones who found work they couldn’t previously afford to do at all, and then did it.
The clock
One last number, because it reframes the whole thing.
Nearly 6 in 10 respondents expect AI to handle a larger share of their work tasks next year than it does today. More than a third expect it to handle most or nearly all of them. And critically — a software engineer and a construction manager anticipate roughly the same increment of progress in their own field over the next twelve months.
That uniformity is the tell. Everybody expects a leap. Almost nobody has a plan that assumes one.
The gap between the organizations that compound this and the ones that don’t will not be a gap in tooling. Everyone has the same tools; they cost forty dollars a month. It will be a gap in habit — between the businesses that learned to delegate whole work and the ones still typing thirteen turns into a chat box, waiting for someone to tell them it’s safe to let go.
The people in your business who already let go are the most optimistic people you employ. That’s not a coincidence. It’s the whole finding.
Steven Nichols is the founder of Banyan Business Outcomes, which helps mid-market and privately held businesses adopt AI in ways that produce measurable outcomes. Data in this piece is drawn from Anthropic’s Economic Index report: Cadences, published June 26, 2026.