Which Work Tasks Will AI Automate First—and How to Audit Your Own Job

AI does not change work according to the organizational chart. It rarely takes over an entire position at once. Instead, it assumes or accelerates individual operations: drafting an email, extracting fields from documents, classifying requests or checking code. A more useful question than “Will my profession disappear?” is therefore: “What tasks make up my week, and which of them could be delegated to a machine with appropriate controls?”

The short answer: digital, repetitive and readily verifiable tasks are sensible places to investigate AI use. Tasks involving accountability, negotiation, substantial ambiguity, protected information or decisions with serious consequences generally call for human control. This is an editorial framework for prioritizing an audit, not a research-validated forecast of which tasks will be automated first.

Why you should analyze tasks, not occupations

A single position combines activities with very different risk profiles. A project manager fills out a report, looks for discrepancies in a plan, conducts difficult negotiations and decides whether to postpone a launch. The first two activities might be candidates for AI assistance or automation. AI could support the third by preparing materials, while the fourth still requires an accountable human decision-maker.

This task-level perspective is consistent with how researchers have assessed potential exposure to generative AI. An index from the International Labour Organization and NASK, published on May 20, 2025, combined nearly 30,000 occupational tasks with expert validation, AI-assisted scoring and harmonized labor data. It estimated that 25% of global employment falls within occupations potentially exposed to generative AI. The study stresses that exposure is not actual job loss: transformation is considered more likely than wholesale replacement.

Observed use also does not establish an economic outcome. An Anthropic analysis of approximately one million Claude.ai conversations found AI use in at least one-quarter of associated tasks across roughly 36% of occupations, while only about 4% of occupations appeared in use covering at least three-quarters of their tasks. In this sample, 57% of interactions were classified as augmentation and 43% as automation.

Those figures describe activity among users of one service, not a representative sample of the entire labor market. Anthropic also could not establish how users applied the responses, whether the outputs were accurate or whether an apparently automated result was subsequently edited. The findings support analyzing work at task level, but they do not validate the prioritization criteria or numerical recommendations in this article.

Which tasks are practical candidates to examine first

AI Feed editorial guidance: the characteristics below are practical screening signals selected by the editorial team. They are not validated study thresholds, and the cited research does not prove that tasks with these characteristics will necessarily be automated before other tasks.

• Digital input and output. The input is text, a spreadsheet, a document image or a structured record, and the output is another digital object. Transferring fields from an invoice into a register is one example.
• Repetition. A similar operation occurs frequently, such as assigning categories to requests, producing standard meeting summaries or formatting product listings.
• Clear output criteria. Required fields, calculations, formatting, source links or agreement with a reference answer can be checked.
• Available and permitted context. Approved examples, instructions and reference materials can be supplied without violating policy, contracts or legal obligations.
• Reversible consequences. An inaccurate draft can be caught before use and corrected without creating unacceptable harm.

Possible candidates for investigation include initial sorting of incoming email, extracting account details, drafting standard replies, converting notes into minutes, finding duplicate text, explaining formulas, generating test data and making basic code transformations. These examples illustrate the AI Feed editorial framework; they are not a universal ranking of automation readiness.

Clerical work does have comparatively high measured exposure in ILO research. An ILO overview published on September 29, 2025 identifies data entry clerks, typists, accounting and bookkeeping clerks, and administrative secretaries among the most exposed occupations. The same overview warns that automating clerical tasks in practice can be harder than a theoretical assessment suggests. Organizational adoption, infrastructure, skills, data access and process design can all constrain implementation.

When to augment rather than delegate completely

In this article, automation means that a system performs an operation within a defined process while a person handles exceptions or reviews results as required. Augmentation means that AI proposes options, identifies gaps, produces a draft or acts as a second reviewer, while a professional retains the substantive decision.

Augmentation may be the more defensible design when a task cannot be reduced to a single correct answer. Examples include developing strategy, editing an important document, analyzing possible reasons for declining sales, designing a product or preparing for negotiations. AI can generate hypotheses, summarize supplied materials and flag possible contradictions. It does not bear professional responsibility, know every informal commitment or reliably avoid unsupported claims.

Software development shows why the distinction is not absolute. In an Anthropic study of 500,000 coding-related interactions across Claude.ai and Claude Code, the specialized coding agent was more often used to perform operations directly: 79% of Claude Code conversations were classified as automation. Yet feedback-loop interactions still involved human input and validation. The authors did not measure the resulting code’s quality, its eventual use or productivity gains, and they caution that their early-adopter sample may not represent developers generally.

AI Feed editorial checklist: inventory your actual work

The entire checklist in this section, including every duration and interval, is an AI Feed editorial recommendation rather than a validated study protocol. Organizations should adapt or replace it according to their schedules, risks, policies and governance requirements.

1. Keep a task log. As an AI Feed editorial starting point, record work for one or two representative weeks. This period has not been established as a minimum or optimal threshold by the cited studies.
2. Choose a workable level of detail. AI Feed suggests logging activities in roughly 15- to 30-minute increments when practical. These numbers are organizational aids, not evidence-derived measurement standards.
3. Use a verb and an object. Instead of “worked on the report,” write “collected metrics from three spreadsheets,” “explained the variance” and “confirmed the conclusion with finance.”
4. Record operating context. Note frequency, approximate duration, input data, intended recipient, current tools, required approvals and the method used to verify quality.
5. Separate superficially similar activities. Drafting a template email and replying to an angry key customer both involve correspondence, but their context, risk and accountability differ.
6. Mark prohibited or sensitive processing. Identify personal data, client material, trade secrets, regulated records and intellectual property before testing any external service.

For an initial map of possible uses by role, treat examples as hypotheses rather than a diagnosis. AI Feed’s catalogue includes a map of 30 roles and their task connections. Organizations considering on-device deployment can also consult the catalogue’s guide to local AI models for Mac. These internal articles were published in September 2026 and were added as supplementary reading during this article’s 2026 update. They were not part of the external research evidence available at the September 29, 2025 evidence cutoff described below.

Local execution can change where data is processed, but it does not remove the need for access controls, lawful use, security review, logging and output verification. It also does not by itself establish that a model is suitable for a particular task.

AI Feed editorial assessment matrix

This matrix and every number associated with it are AI Feed editorial recommendations, not a validated scientific instrument. The suggested 1-to-5 scale is intended only to make internal comparisons easier. It is not a research-backed cutoff, employment forecast or measure of automation probability. A team may use words such as low, medium and high instead.

If a team chooses the AI Feed editorial 1-to-5 format, higher scores for frequency, time, standardization, available data and verifiability can represent greater practical interest. Higher scores for error cost and confidentiality should represent greater risk. Do not total the columns into a supposedly decisive automation score. A single high-risk condition can outweigh several convenience factors.

Potential stop conditions include a prohibition on transferring data, an applicable requirement for professional approval, unacceptable potential harm, inability to verify outputs or an action that cannot be reversed. The exact stop conditions must come from the organization’s policies, contracts, professional duties and applicable law—not from this editorial matrix.

Four possible decisions for each task

These categories are AI Feed editorial decision aids rather than outcomes established by the cited studies. A task can move between categories as tools, evidence, regulation or the organization’s ability to verify results changes. “Automate with review” should not be interpreted as permission to remove oversight where law, policy or professional standards require it.

AI Feed editorial checklist: run a controlled pilot

The complete pilot checklist below, including all quantities, ranges and suggested measurements, consists of AI Feed editorial recommendations rather than validated study thresholds.

1. Choose known cases. AI Feed suggests starting with 20–50 historical, synthetic or appropriately anonymized examples for which an acceptable result is already known. This range is a manageable editorial starting point, not a statistically validated sample size. High-impact deployments may require a larger evaluation designed by qualified specialists.
2. Document the baseline. Record total time, observed error types, returns, review effort and current operating cost before introducing AI.
3. Define quality in advance. Specify required fields, factual checks, permitted sources, formatting rules and errors that automatically fail a case.
4. Limit consequences. During the initial test, do not allow the system to send messages, modify production databases or make external decisions unless separately authorized and controlled.
5. Measure the whole workflow. Include prompt preparation, generation, review, corrections, escalation and complete rework—not only the model’s response time.
6. Track severity as well as frequency. A rare but serious error can matter more than numerous harmless formatting mistakes.
7. Include operating costs. Count model, integration, monitoring and human-review costs. For a separate budgeting method, see AI Feed’s catalogue article on AI API costs for chat, RAG and agents. Published in September 2026, that internal guide is supplementary material and is not part of the 2025 external evidence base used for the research claims in this article.
8. Assign ownership. Name a process owner, define escalation and rollback procedures, and record model, prompt and integration changes.

A pilot should improve the complete process, not merely accelerate generation. If a draft appears quickly but requires lengthy verification or introduces unacceptable uncertainty, the workflow may not have improved. A favorable result on a limited sample also does not establish reliability in every future case, especially when inputs, models or surrounding systems change.

Methodology, chronology and limitations

Publication chronology: this article was updated for publication on October 1, 2026. Its external research claims rely on the ILO and Anthropic materials listed in the sources, with an evidence cutoff of September 29, 2025. The three linked AI Feed catalogue articles were published in September 2026 and added later as supplementary practical reading. They do not form part of the evidence base supporting the article’s claims about occupational exposure, observed AI use, automation or augmentation.

The evidence synthesis combines the ILO’s task-oriented analysis with the automation-versus-augmentation distinction used in Anthropic’s research on observed Claude use. Findings from those sources have not been projected onto a particular employee, employer or industry.

The ILO index measures potential occupational exposure, not actual displacement. Anthropic’s analyses describe activity within its own products and do not establish output quality, productivity gains or representative labor-market adoption. Neither source validates this article’s task characteristics, matrix, scoring scale, logging schedule, pilot sample range or reassessment intervals.

All practical checklists and numerical suggestions in this article—including one or two weeks, 15–30-minute increments, a 1-to-5 scale, 20–50 pilot examples, and quarterly or six-month review intervals—are explicitly AI Feed editorial recommendations. They are not validated study thresholds. They should be adapted or replaced where risk specialists, professional standards, organizational policy, contracts or applicable law require a different method.

Frequently asked questions

Does high AI exposure mean that an occupation will disappear?

No. Exposure indicates that the technology may be applicable to some tasks associated with an occupation. Actual effects depend on capability, quality, cost, regulation, infrastructure, process design, demand for the work and adoption decisions. The ILO explicitly distinguishes potential exposure from actual job losses.

Should an audit begin with the most time-consuming task?

Not necessarily. Under AI Feed’s editorial framework, a frequent, verifiable and reversible operation with manageable risk may be a more informative first pilot. This is an AI Feed editorial recommendation, not a conclusion validated by the cited studies.

Can employees upload work documents to a public chatbot?

Only when organizational policy, contractual terms and applicable data requirements permit it. Personal information, trade secrets, client materials, regulated records and intellectual property may require specific approval or may be prohibited from external processing.

How often should the audit be repeated?

Repeat it after material changes to models, tools, policies, data, regulations or workflows. As an AI Feed editorial planning suggestion—not a validated threshold—rapidly changing teams might schedule a review quarterly or every six months. Other organizations may need event-driven, more frequent or less frequent reviews according to risk and governance requirements.

Does a high matrix score justify full automation?

No. The optional 1-to-5 matrix is an AI Feed editorial comparison aid, not a validated automation score. Legal restrictions, confidentiality, inability to verify outputs or the possibility of serious harm can rule out delegation regardless of other ratings.

Are the 2026 AI Feed links sources for the research claims?

No. They are verified internal catalogue links added as supplementary reading during the October 1, 2026 update. The external evidence supporting the research claims is listed separately below and has a cutoff of September 29, 2025.

AI-assisted article, checked against the primary sources cited above.

Originally published by AI Feed: Which Work Tasks Will AI Automate First—and How to Audit Your Own Job.