When people ask which jobs artificial intelligence will automate, they may be asking the wrong question.
Jobs are rarely single activities.
An accountant does not only calculate numbers.
A programmer does not only write code.
A lawyer does not only research cases.
A customer-service representative does not only answer questions.
Each occupation is a collection of tasks—some repetitive, some analytical, some interpersonal, some physical, and some dependent on judgment.
Artificial intelligence is therefore entering the workplace task by task.
It can draft an email before it can manage a department.
It can summarize a contract before it can negotiate one.
It can generate code before it can take responsibility for a production system.
It can classify customer requests before it can resolve every unusual customer problem.
That makes the most useful question:
Which tasks are being automated or accelerated first—and what do those tasks have in common?
The Short Answer
The tasks most exposed to today's generative AI tend to involve information that is already digital and work that can be expressed through language, code, documents, or structured data.
The International Labour Organization's 2025 global analysis found that clerical occupations remain the most exposed to generative AI, while exposure has also increased in highly digitized professional and technical occupations as AI capabilities improve.
Experimental evidence reviewed by the OECD shows measurable gains in tasks including writing, summarizing, editing, translation, coding, customer support, consulting and related information work.
Real-world usage data collected by Anthropic show another important pattern: AI use remains concentrated. In its November 2025 sample, the ten most common work tasks represented about 24% of Claude.ai work-task usage, while computer and mathematical tasks accounted for roughly one-third of Claude.ai conversations and nearly half of first-party API traffic. Because these are observations from one AI provider rather than the entire economy, they should be treated as evidence about usage patterns, not a universal measure of AI adoption.
| Tasks seeing early AI adoption | Why they fit current AI |
|---|---|
| Drafting routine text | Output can be generated and reviewed quickly |
| Summarizing documents | Input and output are primarily language |
| Translation | Clear transformation from one representation to another |
| Customer-support assistance | Repeated questions and searchable knowledge |
| Routine coding | Code can be generated, executed and tested |
| Document classification | Categories can often be defined clearly |
| Data extraction | Information can be pulled from standardized documents |
| Research assistance | AI can search, summarize and organize information |
| First-pass analysis | Humans can review before decisions are made |
| Administrative communication | Highly repetitive patterns and formats |
These are not necessarily the first jobs to disappear.
They are the first kinds of work inside jobs that organizations can plausibly hand to AI.
What Makes a Task Easy to Automate?
Current AI works best when several conditions appear together.
The information is already digital
AI can easily interact with:
- documents;
- databases;
- email;
- websites;
- code;
- spreadsheets;
- digital images;
- transcripts.
A task that exists entirely inside a computer is easier for software to reach than a task requiring unpredictable interaction with the physical world.
The task can be described clearly
“Summarize this contract into five key obligations” is easier to automate than:
Figure out what is really going wrong with this client relationship.
The first has a relatively clear input and output.
The second requires context, interpersonal interpretation and judgment.
The answer can be checked
Automation becomes more attractive when correctness can be evaluated.
Code can be tested.
A calculation can be recomputed.
A document can be compared with its source.
A database query can be validated.
This helps explain why coding has become such an important arena for agentic AI: software can often test the software generated by another piece of software.
The work occurs repeatedly
Automating something that happens once may not be worth the effort.
Automating something employees perform thousands of times can create substantial savings even if the improvement per task is modest.
Mistakes are recoverable
Organizations are more comfortable automating low-stakes first drafts than irreversible high-stakes decisions.
The cost of a mediocre marketing headline is different from the cost of an incorrect medical treatment.
Clerical Work Is Highly Exposed
The ILO's task-level analysis continues to identify clerical occupations as having the highest exposure to generative AI.
That makes intuitive sense.
Clerical work often includes:
- scheduling;
- document preparation;
- data entry;
- correspondence;
- information retrieval;
- form processing;
- record updating;
- basic bookkeeping;
- administrative coordination.
Many of these tasks involve taking information in one structured form and moving, summarizing or rewriting it into another.
That is precisely the kind of work modern language models can often perform.
But a job such as administrative assistant can also include:
- handling unusual requests;
- coordinating people with conflicting schedules;
- protecting confidential information;
- understanding organizational politics;
- recognizing what is urgent;
- dealing with visitors;
- resolving unexpected problems.
Those responsibilities are harder to capture in a simple automation statistic.
Drafting Is Being Automated Before Final Responsibility
One of AI's clearest workplace roles is producing the first version of something.
Examples include:
- emails;
- reports;
- proposals;
- job descriptions;
- meeting summaries;
- marketing copy;
- documentation;
- presentations;
- policy drafts.
This is important because many jobs contain a large amount of writing without being writing professions.
A salesperson writes follow-up emails.
A manager writes performance summaries.
An engineer writes documentation.
A doctor creates clinical notes.
A lawyer drafts correspondence.
AI can therefore affect many occupations through a relatively narrow capability: turning instructions and information into plausible text.
The human role often shifts from:
write
to:
direct → review → correct → approve.
That can be a substantial productivity change even when no employee is removed.
Summarization Is Another Early Target
Organizations produce enormous quantities of information.
AI can reduce:
- a long meeting to action items;
- a report to key findings;
- a contract to major clauses;
- customer history to a briefing;
- research papers to an overview;
- correspondence to unresolved issues.
The OECD specifically identifies summarizing among tasks where experimental evidence shows generative AI can improve efficiency.
Summarization works well because the source material is available for comparison.
But high-stakes summaries still require verification.
A model can omit an exception that turns out to be the most important sentence in the document.
Customer Support Is a Natural Testing Ground
Customer service combines repetitive work with measurable outcomes.
Many requests resemble earlier requests:
- Where is my order?
- How do I reset my password?
- What is your return policy?
- Why was I charged?
- How do I change my account?
AI can retrieve relevant information, propose responses, summarize account history or sometimes handle straightforward conversations directly.
Experimental research reviewed by the OECD has found productivity gains in customer-support settings, with particularly notable benefits in some studies for less-experienced workers.
But unusual cases still reveal the boundary.
A customer with contradictory account records, an unusual hardship, a legal complaint or an emotionally sensitive problem may require escalation.
The likely architecture is therefore often:
AI handles the routine path.
Humans handle the exception path.
Coding Is Moving Rapidly Toward Agentic Automation
Programming is especially important because code has an unusual property:
it can often be tested automatically.
An AI can:
- write code;
- run it;
- observe an error;
- revise the code;
- run tests;
- continue until the tests pass.
That feedback loop turns code generation into something closer to autonomous task completion.
Current real-world AI usage is already heavily concentrated in computer and mathematical work according to Anthropic's Economic Index data.
This does not mean software developers become unnecessary.
Software engineering also includes:
- deciding what to build;
- understanding users;
- architecture;
- security;
- debugging ambiguous production failures;
- integration;
- tradeoffs;
- maintenance;
- responsibility.
But the amount of code a human needs to type manually is clearly becoming less important as a measure of programming productivity.
Research Tasks Are Being Decomposed
A research assignment might once have required a person to:
- search for sources;
- open documents;
- extract relevant passages;
- compare claims;
- organize notes;
- prepare a first draft.
Modern AI systems can increasingly assist with several of these steps.
The danger is that research is not simply information retrieval.
A system must distinguish:
- authoritative from unreliable sources;
- recent from outdated evidence;
- primary from secondary reporting;
- correlation from causation;
- fact from interpretation.
AI can make research faster.
It can also make poorly verified research faster.
The human role increasingly moves toward source judgment and verification.
Data Entry Is Becoming Data Extraction
Traditional data entry requires a person to look at one representation and type information into another.
AI increasingly changes the workflow.
A system can read:
- invoices;
- receipts;
- contracts;
- applications;
- PDFs;
- screenshots;
- scanned forms;
and transform them into structured fields.
The human then checks exceptions.
This is an important pattern:
automation does not always remove the workflow.
It changes where humans enter it.
Instead of processing every document, a person investigates the minority that the system cannot handle confidently.
AI Is Also Automating Parts of Professional Work
Generative AI is unusual because exposure extends beyond routine clerical work.
The ILO notes increasing exposure in strongly digitized professional and technical occupations.
Stanford's 2026 AI Index reports strong benchmark performance in professional-domain evaluations involving tax, mortgage processing, corporate finance and legal reasoning, although reliability remains a significant challenge.
That means AI can assist with portions of work traditionally requiring substantial education.
Examples might include:
- finding relevant legal material;
- drafting financial commentary;
- reviewing tax documents;
- preparing first-pass analyses;
- extracting terms from contracts.
The consequence is not necessarily the elimination of high-skilled occupations.
It may instead change their internal division of labor.
The Entry-Level Problem
Routine tasks often have another purpose.
They train beginners.
A junior analyst learns by preparing basic analyses.
A junior lawyer learns by reviewing documents.
A young programmer learns by fixing straightforward bugs.
An assistant learns the organization by coordinating routine work.
If AI performs those activities, employers may gain efficiency while weakening traditional apprenticeship pathways.
Stanford researchers analyzing U.S. payroll data found that workers ages 22–25 in the most AI-exposed occupations experienced a 13% relative employment decline after widespread generative-AI adoption, while more experienced workers in those same occupations were comparatively more stable. The researchers found the declines concentrated in occupations where AI was more likely to automate rather than augment work. The authors describe this as early evidence rather than a final verdict on AI's labor-market effects.
A separate Stanford policy review in July 2026 continues to describe aggregate employment effects as small so far and emphasizes the difficulty of separating AI from broader labor-market conditions.
Both findings can be true:
little aggregate disruption
and:
meaningful disruption for particular workers.
Automation and Augmentation Are Happening Together
The word automation suggests that AI takes over a task.
But another pattern is augmentation:
A worker performs the task with AI assistance.
Anthropic's November 2025 usage sample classified 52% of Claude.ai conversations as augmentation and 45% as automation. The same research also found a gradual increase in automation's share over longer comparisons. Again, these numbers describe one provider's usage, not the entire economy.
The distinction matters.
Consider two accountants.
Accountant A
Uploads transactions and asks AI to produce the complete analysis.
Accountant B
Analyzes the accounts and uses AI to identify unusual patterns worth checking.
Both are using AI.
Only the first is primarily delegating the task.
What Is Harder to Automate?
Current systems struggle more when tasks require combinations of:
Ambiguous goals
“Make this business successful” is not a well-defined task.
Accountability
Someone must ultimately be responsible for high-stakes decisions.
Physical dexterity
Software cannot repair a leaking pipe merely by generating instructions.
Unpredictable environments
Real-world circumstances contain exceptions that training data may not anticipate.
Relationships
Trust, negotiation, persuasion, leadership and care depend heavily on human interaction.
Tacit knowledge
Organizations contain unwritten rules that rarely appear in formal documentation.
Difficult verification
If no one can tell whether the AI answer is correct, automation becomes dangerous.
The Exception Economy
One useful way to imagine the future workplace is:
AI handles the normal cases. Humans handle the abnormal ones.
That makes exception handling more valuable.
A travel system can book routine flights.
A human resolves the visa complication.
An accounting system categorizes normal transactions.
A human investigates the strange one.
A customer bot answers routine questions.
A human talks with the furious customer whose situation does not fit policy.
Paradoxically, this can make remaining human work more difficult.
The easy cases disappear.
Humans increasingly see the edge cases.
Automation Can Happen Without Layoffs
Suppose an eight-person team produces 800 reports each month.
AI raises output per employee by 25%.
The company now has several options.
It can:
- produce 1,000 reports;
- improve quality;
- shorten deadlines;
- reduce overtime;
- assign employees to other work;
- hire fewer people as employees leave.
Only the final option reduces employment immediately.
This is why employment effects may appear gradually through hiring decisions rather than dramatic layoffs.
Smaller Companies May Experience Automation Differently
AI can allow small businesses to perform work they previously outsourced.
An OECD survey of small and medium-sized enterprises found that about one-third of surveyed AI-using SMEs reported reduced employee or owner workload, while 14% said generative AI had reduced reliance on external contractors.
That creates another form of labor-market effect.
The person affected may not work inside the business.
It may be the freelancer, agency or contractor the business no longer needs for a routine task.
Five Misconceptions About Task Automation
“If AI can perform one task in a job, the whole job disappears.”
Occupations combine many tasks.
“Routine means low-skilled.”
Highly educated professionals perform many routine information-processing tasks.
“Automation always means layoffs.”
It can instead increase output, reduce outsourcing or slow future hiring.
“AI automatically handles easy work better than beginners.”
Humans still need to evaluate the output, and some research finds inexperienced users can be less effective at recognizing AI errors.
“The tasks being automated today tell us exactly what will happen in ten years.”
Capabilities, prices, regulations and organizational behavior will change.
What Workers Should Watch in Their Own Jobs
Instead of asking:
Is my occupation safe?
Break the job apart.
Ask:
- Which tasks are repetitive?
- Which tasks consist mostly of reading and writing?
- Which tasks have standardized inputs?
- Which tasks can be checked automatically?
- Which tasks require unusual judgment?
- Which tasks require relationships?
- Which tasks make me responsible for the outcome?
- Which tasks teach junior workers how to become experts?
The answers reveal much more than the job title.
What Employers Should Measure
Companies should not evaluate AI by counting prompts.
Measure outcomes.
Useful metrics include:
- time saved;
- error rate;
- rework;
- customer satisfaction;
- output quality;
- escalation rate;
- employee learning;
- cost per completed task.
A task that is 50% faster but produces twice as many serious errors has not been successfully automated.
What to Watch Next
Coding agents
Software development offers strong feedback loops and is likely to remain an important test bed for autonomous work.
Back-office processing
Document-heavy finance, insurance, administration and compliance workflows are natural candidates.
AI agents using business software
The next step is moving from generating information to taking actions across applications.
Entry-level hiring
Watch whether junior positions continue weakening in occupations where apprenticeship tasks are increasingly automatable.
Voice and multimodal systems
Tasks currently requiring telephone or visual interaction may become more accessible to automation.
The Bottom Line
AI is not entering the labor market one occupation at a time.
It is entering one task at a time.
The earliest targets share recognizable features:
They are digital.
They repeat.
They have clear inputs.
They produce outputs that can often be checked.
They involve language, documents, code or structured data.
That explains why drafting, summarization, customer support, coding, research assistance and administrative processing are changing rapidly.
The important dividing line is increasingly not:
Which jobs can AI do?
It is:
Which parts of this job can AI do reliably enough that an organization is willing to reorganize work around it?
That is where automation becomes economically real.
Questions People Ask
What kinds of tasks is AI automating first?
Writing, summarization, translation, coding, document processing, routine customer support and other digital information tasks are among the clearest early areas.
Which occupations are most exposed to generative AI?
The ILO continues to identify clerical occupations as having the highest exposure, with growing exposure in highly digitized professional and technical work.
Is AI actually eliminating jobs yet?
Aggregate effects remain limited in current U.S. evidence, but researchers have found more significant effects for some young workers in highly exposed occupations.
Is AI mostly automating or helping workers?
Both. One large AI-provider usage study found augmentation slightly ahead of automation in late 2025, while also observing a longer-term increase in automation's share.
What is hardest for AI to automate?
Tasks requiring ambiguous judgment, physical interaction, trust, responsibility, unusual circumstances and difficult-to-verify decisions remain harder to delegate completely.