AI Work Is Broad, Shallow, and Weird

July 24, 202615 min read

Update: the career ladder is becoming the bottleneck

In July, Google's ATLAS report left one question frustratingly open:

Did AI make junior employees easier to train, or easier not to hire?

Anthropic now has an answer from inside one of the most AI-intensive workplaces on earth.

The answer is not comfortable.

On October 16, Anthropic published a detailed study of how employment is changing inside the company. It draws on employment records, agent telemetry, output metrics, surveys, calendars, hiring scorecards, internal communication, and 17 employee interviews.

Since early 2025, junior employees' share of Anthropic's technical workforce fell from 21% to 14%.

Their share of technical hires fell from 27% to 9%.

By August, interactive AI agents were logging six hours of active work for every assumed hour worked by an employee.

This does not prove that AI caused every hiring change.

It does show what a company can look like when AI stops being a tab and becomes part of the organization.

The company is still growing. Technical hiring increased in absolute terms. Hiring in sales, finance, and recruiting simply grew faster. Anthropic is also an unusual employer: young, expanding quickly, building the technology it uses, and willing to reorganize itself around it.

Do not turn one frontier lab into the entire labor market.

But do not ignore the lab either.

Anthropic is where the shallow phase is getting deep.

Six agent hours are not six employee hours

The six-to-one number will travel.

It needs a label.

Anthropic measures agent runtime, including inference and active tool use, and compares it with headcount multiplied by an assumed 40-hour work week. The agent figure excludes waits longer than 30 minutes, but it is still not equivalent to six hours of useful, reviewed human work for every employee hour.

In fact, Anthropic says interactive agents spend about 80% of their session time waiting for human input.

So this is not a sixfold productivity multiplier.

It is a sixfold expansion of activity that somebody still has to direct, inspect, interpret, and sometimes clean up.

The report shows both sides of that system.

Code commits per engineer rose to five times their velocity six months earlier. Lawyers processed more contracts. Salespeople closed more deals. Accounts payable handled more invoices. Following the same employees over time, output generally rose after they adopted coding agents.

But adoption was not random. People preparing a large project may have adopted agents because they were already about to produce more. Commits are not business value. Anthropic's code-quality measures cover observable signals such as reverts and later deletions; they cannot capture every subtle defect or maintenance cost.

Meanwhile, the review load grew.

For non-technical teams, reading all the output from interactive Claude sessions would take more than half the work week. Agent-management work rose from 10% to 22% of the average employee's interactive sessions. Median meeting time increased 20% over a year, including 16% for technical individual contributors.

The labor did not disappear.

It moved from making the first pass to steering a fleet of first passes.

Apprenticeship is infrastructure

This is where the hiring data matters more than the runtime headline.

Entry-level work is not only cheap implementation.

It is how people build judgment.

The junior engineer learns by tracing the bug, writing the ordinary test, reading the review, understanding why the obvious fix breaks something else, and slowly acquiring the context that later makes supervision possible.

If an agent absorbs that work, the company may save time today while removing the process that produces tomorrow's senior employee.

Anthropic is not the only signal. A Stanford study using payroll data found no broad economy-wide displacement, but employment for workers aged 22 to 25 in AI-exposed occupations was 19% below the path of less-exposed peers. The authors call the pattern descriptive, not causal.

A separate Stanford study of 41 countries found that companies adopting generative AI reduced the junior share of their foreign affiliates relative to comparable firms. That change came mainly from faster senior hiring, not mass junior layoffs.

And the outcome is not mechanically negative. In a field experiment with 673 German IT apprentices, AI access improved performance by 26 to 31 percentage points on tasks beyond the apprentices' formal training, without reducing immediate comprehension.

That is the actual choice.

AI can become an apprenticeship accelerator.

Or it can become a reason to stop offering apprenticeships.

The model does not choose between those futures.

Management does.

No layoffs is not enough

Anthropic deserves credit for publishing the data and for not reducing the story to a victory lap.

The company has built an early-warning system for roles being transformed by AI. It has moved employees across departments and redeployed whole teams after agents absorbed parts of their work. The report also argues for shared ownership, internal mobility, retraining, and income support.

Those are serious responses.

They mainly protect people who already got through the door.

A company can preserve incumbent jobs through redeployment while quietly closing the path for the people who would have become incumbents.

The report says this directly: protections for current employees do not help people who were never hired.

That should change what responsible AI adoption means.

An employer needs to measure more than output per worker and layoffs avoided. It should track:

the junior share of hiring and headcount
promotion rates and time to independent judgment
how much object-level work remains available for learning
whether AI expands junior capability or only senior leverage
review load, meeting load, and unresolved agent output
internal redeployments alongside opportunities for new entrants
whether productivity gains reach workers through pay, ownership, time, or training

And it needs an apprenticeship budget.

Not a vague promise to hire people with better judgment after the old training path has vanished.

A real allocation of work, mentors, review time, and roles designed to produce judgment.

The map has a new layer

Google's ATLAS report showed AI spreading broadly but shallowly across the economy.

Anthropic shows what happens when that usage becomes deep, agentic, and organizational.

Output rises.

Human attention becomes scarce.

Work shifts upward toward judgment, coordination, and monitoring.

And the bottom rung of the career ladder starts to look optional.

That is not proof of mass automation.

It is something more immediate: evidence that employment can change through hiring composition before it changes through layoffs.

The first people displaced by an AI-heavy organization may not be the people it fires.

They may be the people it never teaches.

The Short Version

The AI jobs debate finally has a better map.

Not a perfect map.

Not a neutral map.

But a map.

On July 23, Google published the first AI & Economy ATLAS report, an ongoing study of how people use Google's AI tools. The full PDF is built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and the Gemini API. It maps usage across more than 800 occupations, 4,000 work tasks, 300 household activities, 150 countries, and 140 languages.

That is a serious dataset.

It is also a Google dataset.

Both facts matter.

The useful takeaway is not:

"AI is replacing everyone."

It is not:

"Relax, nothing is happening."

The useful takeaway is stranger:

AI use is broad, shallow, mostly collaborative, and leaking into places the normal jobs debate barely knows how to describe.

Google says workplace AI adoption spans occupations representing just above 88% of U.S. employment. That sounds enormous. But in the median occupation where Google observes AI use, AI appears in only 21% of tasks. Less than 10% of non-routine cognitive work conversations in the dataset look like end-to-end automation. More than 86% of conversational AI interactions happen outside formal work.

So yes, AI is everywhere.

No, that does not mean the economy has already automated itself.

Very inconvenient. The future has once again refused to fit inside a panel title.

This is why the report matters. It pushes the conversation away from job titles and toward actual task behavior: what people ask AI to do, where they use it, whether it helps or replaces, whether it happens at work or at home, whether it is text or multimodal, and which parts of the economy are invisible to normal productivity statistics.

That is a much better argument than "will AI take my job?"

The better question is:

Which parts of the job are being absorbed into software, and who gets the benefit?

Broad is not deep

The most important phrase in the report is "broad but shallow."

Google sees AI usage across every major sector: professional services, construction, leisure and hospitality, and more. It sees usage in occupations people always talk about, like software developers and market researchers. It also sees usage in less fashionable categories: farmers, foresters, industrial engineers, mechanics, repair workers.

That should kill one lazy story.

AI is not only a San Francisco office-worker phenomenon.

But the depth number matters just as much.

If AI appears in 21% of tasks for the median occupation with AI use, that is not a job replacement number. It is an adoption-surface number. It means workers are trying AI on slices of work. It means the tool is entering workflows unevenly. It means a job can be "touched by AI" without being remotely close to automated.

That distinction is essential.

A marketing manager asking Gemini to brainstorm campaign angles is not the same thing as replacing the marketing department.

A mechanic using multimodal AI to interpret a test result is not the same thing as a robot repairing the car.

A lawyer using AI to summarize precedent is not the same thing as the legal system becoming a chatbot in a robe, which, please, let us avoid for everyone's sake.

The question is not whether AI shows up.

It clearly shows up.

The question is where it becomes load-bearing.

Assistance is not harmless

Google's report says most work-related AI use is collaborative and assistive. People use it for ideation, strategy, information retrieval, learning, partial drafting, review, and refinement. In non-routine cognitive work, full task automation appears in less than 10% of conversations.

That is an important correction to the doom loop.

It is also not a reason to stop worrying.

Assistance can still change labor markets.

If AI drafts the first version, reviews the second version, finds the bug, explains the policy, generates the test, summarizes the meeting, or helps a support agent answer twice as many tickets, that may not look like "automation" inside a classifier.

It can still change the job.

Sometimes that change is good. The worker gets leverage. The boring task gets shorter. The error gets caught. The person with less formal training gets a useful tutor. The repair technician can make sense of a weird diagnostic output at 9:47 p.m. without waiting for someone else.

Sometimes the change is less friendly. The same worker gets more volume, more monitoring, more pressure to do the work of two people, or less bargaining power because the first pass has become cheaper.

This is why "assistive, not automated" is not the end of the debate.

It is the beginning of the measurement problem.

What did the tool change?

Did it save time?

Did quality improve?

Did the worker keep the time, or did the company absorb it?

Did the junior role become easier to train, or easier to avoid hiring?

Did the expert become more valuable, or did their judgment get wrapped into a workflow someone else controls?

Those are the questions that matter.

One of the better parts of ATLAS is that it does not treat physical work as AI-irrelevant.

Google says workers in manual and technical trades are using conversational AI as a collaborator for diagnostics, troubleshooting, and real-time learning. The report mentions automotive technicians and industrial mechanics using AI to interpret test results, debug electrical wiring, and inspect machinery for wear. It also says multimodal AI use is more than twice as common in these contexts compared with the overall work baseline.

That makes sense.

If your job touches the physical world, text-only AI is often a bad abstraction. A technician does not only need a paragraph. They may need to show the machine, the wiring, the part, the gauge, the error code, the broken thing that refuses to explain itself politely.

This is where the "AI replaces white-collar work first" story gets too clean.

Manual jobs contain cognitive tasks.

Technical jobs contain documentation tasks.

Physical work contains diagnosis.

Trades contain learning loops.

AI can enter through those adjacent tasks without taking over the whole job.

That matters for workers, training programs, unions, tool vendors, insurers, schools, and anyone trying to understand productivity outside the laptop class.

The economy is not just documents and code.

Though, tragically, there are still plenty of documents.

Home use may be the hidden economy

The report's weirdest finding may be outside work.

Google says more than 86% of conversational AI interactions in ATLAS happen outside formal work. The report maps non-work activity to American Time Use Survey categories and argues that AI conversations span activities representing about 98% of Americans' non-sleep time.

That does not mean 98% of life is automated.

Please do not pitch that deck.

It means AI is showing up across the texture of ordinary life: household management, purchasing research, education, personal care, travel, legal questions, finance, government services, appliances, tools, and other high-friction admin.

This part is easy to undercount because standard productivity statistics are built for formal work and market transactions. If AI helps someone understand a benefits form, compare a repair estimate, write a school appeal, translate a medical instruction, or figure out which part to buy for a dishwasher, the value may be real even if no employer records it.

Google estimates that if household AI use saved 30 minutes per week on average, unpaid productivity gains in the U.S. alone could be worth about $100 billion using standard valuation methods.

Treat that as an estimate.

A very estimate-shaped estimate.

But do not ignore the underlying point.

Some of AI's early economic value may appear less as "a company reduced headcount" and more as "millions of people got a little help navigating annoying systems after business hours."

That is still economic value.

It is just harder to count.

Google's map is not the territory

Now the skepticism.

Google has incentives.

Google would rather the public debate believe AI is broadly useful, globally distributed, assistive rather than job-destroying, and good for everyday people trying to solve practical problems. That does not make the report false. It does mean we should read it as evidence from one of the largest AI platform owners in the world.

ATLAS is based on Google surfaces: Gemini App, AI Mode, and Gemini API. It does not capture all AI usage. It does not fully capture Workspace, Google Cloud Gemini Enterprise, AI Overviews, agentic coding, world models, other labs' products, private enterprise deployments, local open models, or AI embedded silently inside software.

It also does not prove outcomes.

The report says this clearly. ATLAS can observe what people are doing with AI. It cannot see the final productive output, whether the interaction worked, whether time was saved, whether quality improved, or whether a manager later used that workflow to restructure a team.

That is the gap.

Usage is not impact.

A prompt is not productivity.

A conversation is not a labor-market outcome.

This is especially important for entry-level hiring. The report says ATLAS cannot shed much light yet on whether AI is dampening entry-level hiring. That is one of the biggest open questions in the whole AI economy. If senior workers use AI to do more of the first-pass work that used to train juniors, employment effects could appear before job titles disappear.

The map is useful.

It is not the territory.

What this changes

Two weeks ago, I wrote that the AI economy needs instruments, not panic.

ATLAS is an instrument.

An imperfect one.

But still an instrument.

It shows the direction measurement needs to go:

from job titles to tasks
from tasks to workflows
from usage to outcomes
from outputs to quality
from automation claims to intent
from office work to physical work
from paid work to household value
from U.S. and China narratives to global adoption patterns
from "AI is happening" to "where, how deeply, and for whom?"

That is the serious story.

AI is not a single wave hitting every worker the same way.

It is a messy diffusion pattern across jobs, chores, languages, countries, trades, companies, and software surfaces. Some of it is shallow. Some of it will deepen. Some of it will stay as autocomplete with better manners. Some of it will become load-bearing infrastructure. Some of it will be measured only after the workplace has already changed.

So the smart response is not panic.

It is also not reassurance.

It is instrumentation.

If you run a company, measure which tasks AI is actually touching, whether output improves, and where review time moves.

If you build AI products, show users what changed, what the model read, how confident it is, and where human review happened.

If you make policy, stop arguing about one national job number and start funding faster labor-market telemetry, training experiments, sector-level adoption data, and ways to share gains when workflows compress.

If you are a worker, do not ask only whether AI can do your job.

Ask which parts of your work are becoming software, which parts still require judgment, and whether you are becoming the person who directs the system or the person whose work gets quietly absorbed by it.

That is the line to watch.

Not "AI or no AI."

Broad, shallow, weird, and getting deeper.