How to Use an AI Work Checker Before Rollout

Use the checker as a dated baseline, not as a team forecast

Use the free checker now as a dated baseline: it can identify tasks to examine, but cannot tell you whether your employer plans to adopt AI or whether your job will change. Record recurring tasks, routine steps, exceptions and review; date the record and annotate the signals with what you know. With no confirmed rollout, treat the result as a question list, not a forecast. Inspect one task and choose a learning step that fits your time and budget. Do not commit to costly retraining or a role change on a checker signal alone. Keep three observations separate. First is capability: what AI may be able to do with a task. Second is use elsewhere: whether comparable tasks appear in documented use beyond your team. Third is local adoption and redesign: whether your team has an approved tool, changed procedure, pilot, training plan or altered task assignment. Capability cannot prove performance in your setting; use elsewhere cannot show your employer has adopted it; an announcement does not establish that your workflow changed. Gaps point to what you should verify next, not an outcome to assume. The International Labour Organization’s *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* shows why task signals need this boundary. Its 2025 paper combined occupational task assessments, a survey of 1,640 employed people in Poland, expert discussions and AI-assisted scoring to produce global estimates across four exposure gradients. The paper says transformation is the most likely impact because most occupations include tasks requiring human input. These modeled occupational estimates are not observations of your employer’s workflow or plans, and are not individual job-loss probabilities. Use them to understand why a checker may flag parts of a role, not to infer what will happen to your team. *A Practical Way to Read Your Own Job’s AI Exposure* recommends listing actual tasks and considering routine work and time spent; it calls its checker a prompt for an individual audit. But “routine” cannot settle whether a task will be automated: exceptions, quality checks, information access and accountability matter too. Keep the record descriptive; do not enter work, client or personal information into an unapproved system. Check policy before workplace experiments. Revisit the baseline when an approved pilot, procedure, training plan or task assignment supplies evidence. Until then, a bounded learning move is preparation, not proof that a career pivot is needed.

Sources: A Practical Way to Read Your Own Job’s AI Exposure; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

What does a checker result measure before your team adopts anything?

A task checker can organize signals about which parts of a job may be affected by AI. It cannot measure your employer’s deployment decision, whether an output will be reliable in your setting, or your personal probability of losing work. Read a task signal as “investigate this task,” then inspect how often it occurs, how much time it takes, what exceptions it contains, and who verifies the result. With no confirmed rollout, it is a question map, not a forecast. Exposure concerns possible contact between technology and work activities; adoption is an organizational choice. A system might draft a summary while a worker checks facts, handles exceptions and approves the result. Capability on one step does not show that a full workflow can run without people, that an employer has approved a system, or what it will do with time saved. The International Labour Organization’s *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* makes task analysis central to its estimate. The 2025 paper used sampled tasks in Poland’s occupational classification, worker assessments, expert input and AI-assisted estimates. It reports 52,558 data points covering 2,861 tasks, across four exposure gradients. Its global estimates put one in four workers in occupations with some exposure and 3.3% in the highest gradient; the authors say transformation is more likely than full replacement because most occupations contain tasks requiring human input. These modeled occupational estimates are not observations of your workplace or individual job-loss odds. These estimates cannot tell you which process your team will adopt. NASK’s explanation of the ILO–NASK Index says its purpose is to identify where work is likely to change, not job disappearance. Use the index to prioritize inspection, not label a job safe or doomed. Roles combine repeatable steps with tasks involving context, judgment, physical presence, trust or accountability. These can raise implementation and checking costs, but offer no permanent protection. When a checker flags a task, break it into subtasks. Note the input, routine steps, exceptions, quality threshold, review and who owns the outcome. Consider importance as well as count: one activity may take substantial time. *A Practical Way to Read Your Own Job’s AI Exposure* also recommends mapping actual tasks and considering routine and time spent. Its “durable” task advice is not universal: judgment and accountability can change, and an outside article cannot establish what your team values. Use the checker’s change-pressure signal to choose what to examine first, then note frequency, time, exceptions, review burden and mistake consequences. This practical annotation is not a validated formula. A frequent, high-signal task merits attention; its treatment depends on evidence about the actual workflow. Imagine a report task that gathers figures and drafts a summary. A person may still confirm the figures are current, explain an unusual movement and approve the result. Reliability depends on data, error costs and checking. This illustration does not imply local use; it shows why “AI can help with this step” and “the job is replaceable” differ. End with a task note, not a career verdict: “This step merits investigation; here is what a person checks and owns; local adoption is unconfirmed.” Date it so a later approved pilot can be compared. Until then, exposure describes a possible pressure point. It does not establish adoption, role demand, a staffing decision or displacement.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; NASK; A Practical Way to Read Your Own Job’s AI Exposure

Why does adoption elsewhere fail to tell you what is happening in your team?

Adoption statistics show that AI use is spreading across specified populations. They do not tell you that your team has an approved tool, connected it to work systems, or changed who performs a task. Treat outside data as context for questions. A local conclusion needs local evidence: an approved tool, a named owner, changed procedures, training, a pilot, or observable shifts in handoffs and review. An employee may use a public tool personally, a company may license software, a department may test it, and later a workflow may change. Each is a different observation, and a checker result cannot fill the gaps between them. The OECD announcement “AI use by individuals surges across the OECD as adoption by firms continues to expand” reports that 20.2% of firms in countries with available data used AI in 2025, up from 14.2% in 2024 and 8.7% in 2023. It also reports that 52.0% of large firms used AI, compared with 17.4% of small firms. These are firm-level indicators, not counts of teams or redesigned tasks. They tell you that use is present in the measured population and varies by firm size; they do not reveal which system is used, which work it touches, or whether staffing changes. The OECD separately measures individual use, including employed people. That describes people using tools, not their employer's deployment. Observed activity on one product answers a narrower question. Anthropic's “Economic Index: New building blocks for understanding AI use” analyzes conversations on Claude.ai, which captures consumer use, and its first-party API, used mostly by businesses according to the report. Its November 2025 sample was predominantly Claude Sonnet 4.5; conversations were classified by measures including task complexity, purpose, autonomy and success. This can illuminate activity on those products. It is not a census of all workers, systems, employers or rollouts. Product data can show patterns within the observed service, but cannot establish how common they are elsewhere or whether your organization has adopted AI. Use them as clues about possible applications, not forecasts for your team. The OECD report “Generative AI and the SME Workforce” shows why use somewhere in an organization does not necessarily mean core work has changed. In its 2024 survey of SMEs across seven countries, 28.7% of SME users of generative AI said they used it in core activities, defined as activities producing revenue or the main goods and services. The surveyed firms also reported more use for simple, one-off and less important tasks than for complex, recurring and important tasks. These are self-reports from SMEs, not evidence about every employer, large enterprises or a different industry; they do not show that peripheral use has no consequences. They do demonstrate that “uses AI” is too broad to describe a particular workflow. A tool may help with a peripheral task without being embedded in the repeated work that defines a team's output. Use the gap between broad statistics and your own knowledge as a reason to ask a bounded question, not to guess. Record whether each item is external context or local evidence, with its date and source. For local evidence, note what specifically changed: an approved tool, pilot, revised procedure, training, handoff or review burden. If all you know is that no rollout has been announced to your team, record that precisely. It does not prove that no one elsewhere is experimenting, and it does not show adoption is imminent. The missing team-level information is a real limit on what you can conclude. Outside figures can frame what may be worth watching; observable evidence about your task is needed before treating them as a career signal.

Sources: AI use by individuals surges across the OECD as adoption by firms continues to expand; Economic Index: New building blocks for understanding AI use; Generative AI and the SME Workforce

How can one task change without the whole job disappearing?

A task can change while its job remains a bundle of responsibilities. A system may draft, classify, summarize, or retrieve information while a worker checks fit with a client’s situation, handles exceptions, corrects errors, and owns the final communication. Keep four stages distinct. Capability means a system can perform a task under conditions. Observed use means people use it for comparable work. A local pilot puts it into a defined workflow. Job redesign changes assignments, review, staffing, or service expectations. Displacement is a further outcome; none of the earlier stages establishes it. A checker can point to tasks worth examining, but cannot tell which stage your team has reached. “The tool did the task” may mean it drafted text a person edited, handled a routine case with light review, or completed the first step before a person resolved exceptions. If typing takes less time but verification takes more, work has shifted. Use also depends on information access, integration, responsibility, and escalation paths. “Anthropic Economic Index: New building blocks for understanding AI use” illustrates why observed use needs boundaries. Anthropic analyzes Claude.ai and first-party API conversations, the latter mostly business use. In its November 2025 sample, augmentation accounted for 52% of Claude.ai conversations and automation for 45%. These are categories in one provider’s data, not shares of workplace tasks or jobs. They do not show a particular employer deployed AI; the report says product conversations may not map clearly onto real-world job changes. A productivity gain is narrower than a job outcome. If an approved pilot shows a routine summary takes less time with acceptable quality, that supports a claim about the task under those conditions. Management might use time to shorten a queue, serve more requests, handle harder work, raise output expectations, or reorganize staffing. The result depends on demand, budgets, quality requirements, and choices a checker cannot observe. Exposure and productivity should not become a personal job-loss forecast. The International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” combines task data, worker input, expert discussions, and AI-assisted estimates. Its May 2025 record estimates one in four workers globally are in occupations with some exposure; 3.3% of global employment is in the highest category. It says transformation is most likely because most occupations include tasks requiring human input. These modeled estimates are not observations of an employer’s rollout or probabilities for an individual. They support inspecting task bundles, not declaring a role safe or doomed. If a team pilot arrives, compare before and after: task and cases tested; worker actions; system output; remaining review and exceptions; and who owned the result. Note quality and workload alongside elapsed time. Ask whether work moved to a colleague or changed expected volume. A faster first step is not an end-to-end gain if checks take longer or another task becomes the bottleneck. Look for repeated evidence. A demonstration may show capability; a pilot may show performance in one setting; recurring use and changed assignments may show adoption. Employment consequences remain open until separate evidence about demand and staffing. Record what task changed, what work moved, and what responsibility stayed. Treat speed as a workflow measure, redesign as an organizational choice, and displacement as an outcome requiring its own evidence.

Sources: Economic Index: New building blocks for understanding AI use; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

What counts as useful local evidence if the rollout has not reached you?

Useful local evidence names a task, a work setting and an observable change. An approved pilot is more informative than a general announcement because it identifies where evaluation is permitted. Written guidance, training, a workflow owner, or a connected tool can show that preparation has become operational. Stronger still is a change you can observe: a task moves to a different person or system, review time changes, exceptions accumulate, quality checks change, or workload shifts. None of these alone proves that a job will disappear. They help answer the narrower question: has this team's work actually begun to change? Keep a short dated record so later comparisons do not depend on memory. For one recurring task, note how often it occurs and roughly how much time it takes; what information it uses and its sensitivity; the routine steps; common exceptions; the quality threshold; who reviews the result and remains accountable; and the current policy or tool status. Add the source of each observation, its date, and your confidence in it. Distinguish a lead saying evaluation is being considered from written pilot guidance or an actual task reassignment. These are different kinds of evidence. This record is an editorial aid for comparing a task before and after a change, not a validated assessment instrument. The NIST AI Risk Management Framework Core helps explain why the details matter. Its Map function asks organizations to define the intended task, context, system limits, intended use and human oversight. Its Measure function calls for evaluation in conditions resembling deployment and monitoring once a system is in use. A product demonstration does not show that a tool performs acceptably with the team's information, exceptions and review requirements. The framework describes organizational risk management; it does not tell an employee to run a test or authorize use of an unapproved system. Evidence about implementation also argues against treating interest as rollout. In the OECD's Generative AI and the SME Workforce survey, more than 5,000 small and medium-sized enterprises across seven countries were surveyed using data collected at the end of 2024. Among surveyed SMEs using generative AI, 28.7% reported using it in core company activities; the report says use was more often in peripheral, simple or one-off tasks than in core, complex or recurring work. This self-reported SME result cannot describe every employer or establish what is happening in your team. It does show why “the company is exploring AI” and “this workflow has changed” should remain separate statements. The same OECD report lists barriers among surveyed SME non-users, including whether the technology fit the work, concerns about legal or copyright issues and input information, and worker skills. Those figures are not rates for your employer; they are reasons to check actual fit rather than assume that organizational interest becomes deployment. Update the record when policy, a pilot, training, access, task assignment or workflow observation changes. Keep notes at the level needed to compare work, and do not copy confidential, client, personal or regulated information into an unapproved tool. If there is no local evidence yet, the honest entry is “not confirmed,” alongside the source and date you checked.

Sources: AI RMF Core; Generative AI and the SME Workforce; Generative AI and the SME Workforce

How can you investigate safely without creating an unauthorized pilot?

A checker can help you decide what to examine; it cannot authorize an experiment at work. Check your employer’s policy and identify approved tools, data, and use cases before trying any AI system. Do not paste company documents, customer details, personal information, or other confidential material into a public or unapproved service, Keep any personal learning exercise outside live work systems and use public or synthetic material that contains no real client or employer information. Use approved tools only within their stated boundaries. This boundary reflects how responsible organizational assessment is supposed to work. NIST’s *AI RMF Core* describes governance as including policies, documented roles and responsibilities, training, and attention to third-party software and data. Map defines context, purpose, tasks, limits, and human oversight; Measure concerns evaluation and monitoring in relevant conditions. These are organizational practices, not a worker’s permission to conduct an informal pilot. It is voluntary guidance, not a legal ruling. Its practical lesson for an employee is narrower: a personal checker result cannot supply missing details about an approved system, its data, its limits, or who is accountable for its use. The OECD report *Are SMEs prepared for generative AI?* helps explain why those questions matter without implying that its results describe your employer. It surveyed more than 5,000 small and medium-sized enterprises in seven countries at the end of 2024. Among surveyed SMEs that did not use generative AI, respondents cited poor fit with their work, legal or copyright concerns, input-information concerns, and worker skills as barriers. Among SME users, the report also found that training and staff guidelines were not universal. These self-reported SME findings do not establish rates for every company or facts about your team. They show why task fit alone cannot settle appropriateness: data handling, rules, skills, and workflow fit matter too. If the policy leaves the question open, ask through a normal channel such as your manager, team lead, or designated technology contact. Ask neutrally: “For [task], are there approved tools or guidance? Is an evaluation or training planned? If a pilot happens, which handoffs may change, and who reviews and owns the final work?” Seek clarification, not a promised date or jobs prediction. A reply records stated policy or plans, not proof that the workflow has changed. If no evaluation is planned or nobody knows, record that as unknown rather than filling the gap with a market statistic or checker score. You can still prepare while waiting. Choose one recurring flagged task and practice a transferable skill: break it into steps, check source quality, or evaluate an output against a stated standard. Practice with public information or invented examples, Note what you would verify before relying on an answer. Keep the exercise modest enough to fit your available time and budget; there is no need to buy a credential just to satisfy uncertainty. If an approved pilot later appears, use the official process and ask what can be measured, how exceptions and errors are handled, and who remains responsible. This builds capability without turning a personal assessment into an unauthorized test or employer diagnosis.

Sources: AI RMF Core; Generative AI and the SME Workforce

When does the baseline justify training or a career-path comparison?

Before rollout, a checker signal is a reason to inspect a task and perhaps learn one transferable skill, not to buy an expensive credential or resign. The International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” estimates exposure from tasks using worker input, expert review and model estimates. It concludes job transformation is more likely than whole-job replacement because most occupations include tasks requiring human input. It does not measure your employer’s plans, local demand or your chance of losing work.

Use an evidence ladder. A checker result prompts task decomposition: which steps are routine, how often do they occur, and where are judgment, exceptions, review or accountability needed? Similar use elsewhere prompts a question about fit, not a conclusion about your workplace. OECD firm adoption figures describe firms across countries ; they cannot establish whether your team has approved a tool or changed a procedure. A local pilot is a reason to monitor the workflow. Repeated changes in task assignments, quality, workload and required skills justify comparing paths.

When evidence appears, ask whether an upgrade in your current role addresses the gap. Then consider an adjacent role that uses your experience, and a larger change if those options do not fit. Check actual openings, duties and credential requirements where you can work. Exposure is not demand, and neither establishes displacement. Match learning to the gap: problem framing, data literacy, evaluation and verification can transfer across tools; training on one vendor interface may need refreshing. A certificate or self-study alone does not prove readiness .

Choose one small learning step you can afford in time and money, while protecting salary needs and accounting for location, health and family responsibilities. Consider costly training only when changed work and a plausible opportunity make its purpose clear. If later comparing an upgrade, an adjacent move and a larger change against those constraints, the paid roadmap can organize scenarios and a 30/60/90-day plan; it guarantees neither income nor employment.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; AI use by individuals surges across the OECD as adoption by firms continues to expand; Generative AI and the SME Workforce

What should you do this week?

This week, use the free checker to identify a task worth examining and save the result with the date. Choose one recurring task; note its routine steps, exceptions, review, and who remains accountable. Add what you know about tool or policy status, labeling unknowns clearly. The checker gives task-level change-pressure signals. It does not estimate your chance of job loss or establish that your employer has adopted AI. Treat adoption elsewhere as context, not local evidence. Check your workplace’s approved-tool policy before experimenting. Ask through your normal channel: “Is there an approved pilot or training plan for this task, and where can I find the guidance?” The NIST AI Risk Management Framework Core emphasizes understanding a system’s intended task, context, limits, and human oversight. This helps identify useful deployment evidence; it does not authorize testing an unapproved tool. Choose one learning step tied to the task, such as practicing how to check a generated summary against public material, within your time and budget. Revisit your baseline when you learn of an approved pilot, new procedure, or change in assignments or review. If later evidence raises a larger career decision, the free checker at /ai-job-risk-checker provides task-level signals and first actions. The optional roadmap at /career-roadmap compares scenarios and constraints and provides a 30/60/90-day plan; it guarantees neither income nor employment.

Sources: AI RMF Core

Questions readers ask