How to Track AI Task Changes Without a Job-Loss Score

How can you keep a useful record without making a job-loss score?

Keep a dated record of tasks. Separate what AI appears capable of doing from whether it was tried or adopted, how the workflow changed, and what happened to time, quality, review, workload, and responsibility. Compare similar work across cycles, use occupational exposure as context, and choose a next step from evidence and constraints. The record can show how work is changing; it cannot calculate your odds of losing a job. A capability match is not workplace adoption; adoption does not show which tasks changed; task change does not settle labor demand or an individual outcome. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” combines theoretical measures with observed-use measures to compare occupations, while stating that its categories do not forecast employment effects, adoption probability, wages, or worker replacement. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that indicators vary by method and omit adoption conditions and labor outcomes. For example, record a drafting-tool trial, human review, and unclear revision time. Keep uncertainty visible and review comparable cycles before deciding whether to observe, adjust the workflow, or seek clarity at work.

Sources: Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t

What should each entry record?

Record each task as a small, dated observation, not as a point in a personal risk calculation. A useful entry has four layers: what work is being done, what the tool can or was allowed to do, what changed in the workflow, and what the worker observed afterward. Keep blank or uncertain fields marked “unknown.” Guessing a frequency, benefit, or risk to make the record look complete would add precision without adding evidence. Start with the task’s name and purpose in plain language. “Prepare a weekly variance summary for the operations meeting” is more useful than “analysis,” because the first description identifies the output and why it matters. Record the period covered and, if known, a rough count or frequency: once per week, several times per day, or only when a case is escalated. Similar outputs can serve different decisions and carry different stakes. Next separate capability from use. Capability means a tool appears able to assist with some part of the task in a demonstration or controlled test. Actual use means it was tried in the worker’s real context, with the relevant permissions and information. Routine use means it has become part of a repeated workflow. A tool’s ability to draft a summary does not establish that an employer has introduced it, that a worker is permitted to use it, or that the draft can be relied on without review. The International Labour Organization’s *Generative AI and Jobs: A 2025 Update* estimates occupational task exposure from task-level analysis; that is useful context for identifying possible overlap, but it is not a log of an individual’s adoption or workplace outcome. The third layer records the workflow around the task. Note which steps changed, which stayed with a person, and where work moved next. For example, a draft may be generated automatically while a worker still gathers inputs, checks figures, resolves exceptions, approves the final version, or explains it to a decision-maker. Record any new handoff or review step and who remains responsible for checking and approving the result. This avoids treating contact with a tool as proof that the whole task disappeared. The final layer is an observation of results and conditions, not a verdict. When practical, note elapsed time for the whole task, amount completed, quality indicators already used at work, corrections or rework, and review effort. Keep definitions consistent: drafting time and end-to-end time are not comparable. Include relevant context such as unusual volume, a deadline, a policy change, new software, or help from a colleague. These details do not prove what caused the difference; leave unknowns visible rather than inventing estimates. A compact entry might read: “Weekly variance summary; period ending 8 October; normally prepared once per week; drafting feature tested on two summaries; worker checked source figures and edited the draft; final review remained with the analyst; completion time not measured consistently; one correction noted; volume higher than usual.” This is an illustrative format, not evidence about any particular workplace. It preserves what was tested, what the person still did, and what remains unknown, without converting a few observations into a forecast. An occupational task list can help identify duties missing from a personal record, but it should remain a prompt. The O*NET Task Statements file maps task descriptions to U.S. O*NET-SOC occupations and documents whether each statement is core or supplemental, the number of incumbents responding, and an update date. It can broaden a checklist, but cannot say what a worker did this week or whether a workplace introduced a tool. The O*NET database is updated over time, so note the release or update date when using it as a reference. Keep the record minimal and safe to maintain. Do not copy customer details, confidential documents, internal prompts, or sensitive employer material into a personal file unless the workplace has approved that storage. Preserve a traceable account of task, use, workflow, and outcome so later decisions rest on observations rather than an exposure label.

Sources: Task Statements; O*NET Database at O*NET Resource Center; Generative AI and Jobs: A 2025 Update

Why can’t an exposure measure become a personal job-loss score?

An exposure measure cannot become a personal job-loss score because it answers a different question. It estimates whether AI capabilities overlap with tasks associated with an occupation, or maps selected recorded AI use onto those tasks. It does not observe whether your employer has adopted a system, how your workflow has been redesigned, whether demand for the work is changing, what staffing decisions will follow, or what will happen to you. Treat an occupational category as a prompt to inspect your actual task mix, not as individual odds. Those stages are related but not interchangeable. Capability means a system might assist with or perform some task under specified conditions. Observed use means interactions with a particular system have been recorded and mapped to work activities; it does not establish typical use across employers. Adoption means an organization has chosen to deploy a system in its own setting. Workflow redesign describes how steps, handoffs, checks, and responsibilities change after that choice. Labor demand concerns how much work employers or customers require, while displacement is an employment outcome for workers. The U.S. Bureau of Labor Statistics (BLS) makes this boundary explicit in “Artificial Intelligence (AI) exposure categories.” For its 2025–35 employment-projections context, BLS combines five external sources: three theoretical exposure measures and two measures based on observed AI use mapped to occupational tasks or work activities. It calculates percentile ranks within each source, takes median ranks for theoretical and observed dimensions, then groups occupations into four relative categories: low, moderate, high, and very high. The construction is useful for orientation, but its details matter when interpreting a label. BLS reports 831 occupations across five source measures, or 4,155 possible occupation-source combinations; 211 values were imputed, affecting 75 occupations. The categories depend on source coverage, occupational crosswalks, imputation, normalization, and clustering choices. BLS also says the theoretical inputs describe capabilities available no later than mid-2023, while observed measures may lean toward early adopters of particular models. “Workers’ exposure to AI: What indicators tell us – and what they don’t,” an April 2026 International Labour Organization (ILO) research brief, reinforces the distinction across measurement approaches. It says exposure indices vary with their definitions and methods, often rely on static lists of existing tasks, embed judgments, and omit adoption constraints such as economic conditions and institutional barriers. The brief describes technological susceptibility, not labor-market outcomes: an indicator does not establish whether firms find a use feasible, how they will change workflows, or how employment, wages, and demand will adjust. This is a measurement review, not evidence about a particular worker’s employer. A separate ILO brief, “Generative AI and jobs: A 2025 update,” illustrates both the value and the boundary of aggregate estimates. Its method combines task-level data, expert input, and AI predictions across nearly 30,000 tasks at six-digit ISCO-08 occupational detail, grouping occupations into four exposure gradients. It estimates that one in four workers globally is in an occupation with some GenAI exposure and concludes that most jobs are more likely to be transformed than made redundant. So a high exposure band is not a dismissal notice, and a low band is not a safety certificate. Use the label to ask which tasks in your role resemble the measured task bundle, which are absent from it, and whether you have observed an actual local change. Keep the answer to those questions in your dated record. If the record and an occupational category disagree, preserve the mismatch: it may reflect local task mix, timing, tool access, or limits in the index. Neither signal, alone or combined as a homemade score, supplies a validated probability of personal job loss. That distinction keeps technical possibility separate from an organizational choice and an employment result.

Sources: Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t; Generative AI and Jobs: A 2025 Update

What kinds of change belong in the same record?

Track both substitution and assistance, and record the whole task rather than only the step a tool touches. For a recurring report, note whether a draft was generated, whether it was adopted, and what happened through delivery: time spent preparing and checking, number of reports completed, corrections or rework, exceptions, quality and who approved the judgment. A faster first draft does not by itself show that the task became shorter, workload fell, or responsibility moved. Imagine a reporting task where a tool can assemble a first draft from approved notes. In one pattern, the worker checks routine fields, corrects errors, and spends less total time per report. That is assistance which may reduce effort in one part of the task. In another, the quick draft requires checking sources, reconciling figures, and documenting the conclusion. The writing step changed, while verification and accountability remained or grew. A third pattern is that easier drafting leads a team to request more reports or more tailored versions; time per item may fall while total volume and coordination rise. A fourth is that routine reports are generated with limited intervention, but unusual cases still require a person to investigate, explain, and sign off. These are distinct changes, not points on a single “safe” to “at risk” scale. Preserve both the changed part and the surrounding work. Record whether minutes were saved, moved into review, or replaced by another task; whether output volume changed; whether error rates, corrections, or escalations changed; and whether the worker’s discretion, approval duty, or exposure to consequences changed. Write mixed results plainly. “Draft time shorter; total time unclear; review still required” is more informative than “AI saved time.” If output rose, distinguish a higher count from better quality or lower workload. If a human check catches problems, include that check as work rather than treating it as friction that can be assumed away. Describe observable changes; leave their significance open for comparison. The OECD’s 2023 working paper, “The impact of AI on the workplace: Evidence from OECD case studies of AI implementation,” draws on nearly 100 qualitative case studies in manufacturing and finance across eight OECD countries. In those cases, job reorganization appeared more prevalent than displacement; the paper also reports possible reductions in tedious work and improvements in engagement or physical safety, alongside changed skill requirements and reports of increased work intensity. This shows implementation can change task bundles and job quality in more than one direction. It is not a representative estimate of how often any outcome occurs, and the cases were gathered before the latest wave of generative AI use. It cannot tell a particular worker what their employer will do. The ILO’s “Generative AI and jobs: A 2025 update” likewise frames its occupational estimates as potential exposure and says, in aggregate, that transformation is more likely than redundancy for most jobs. That conclusion is not an observation of a particular workplace or a promise that a role will remain. Taken together, these sources support recording task reorganization and its consequences, not predicting an individual outcome. Workload may also reflect staffing, seasonality, policy, demand, or other software changes. A personal record can show that a change followed a tool trial; without a comparison that separates these factors, it cannot establish the tool caused every difference. Keep those competing explanations beside the entry, and do not equate faster first output with less total work or assistance with task elimination.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

How should you compare changes across work cycles?

Compare a task with itself under similar conditions, not with a different task that happens to use the same tool. Ask whether work changed across cycles: what was produced, how much checking it required, and who handled exceptions. A demonstration can show that a capability exists in a particular setting. It does not show that a workflow adopted it or that routine results will match the demonstration. Keep capability tests, approved trials, and ordinary production work in separate entries. If there is no reliable baseline, mark the first entry as the starting point rather than reconstructing a comparison from memory. Choose a recurring task with a clear purpose, such as preparing an internal report, and define the unit to compare. It might be one similar report, a recurring batch, or a regular work period. Record the date and context that could affect results: volume, complexity, deadline, input quality, staffing, policy, and tools or steps in use. Then use the same outcome labels each time where feasible. Note elapsed time, completed output, corrections or rework, verification required, and handoffs or escalations. Use observations you can make; if a measure is unavailable, record “not tracked” instead of filling the gap with an estimate. The goal is a legible comparison. After several comparable cycles, or sooner if the workflow changes materially, review the entries together. Look for consistency across dimensions rather than asking whether one number went down. A shorter first draft may coincide with more checking, corrections, a higher output target, or a different division of responsibility. A summary can label the pattern “repeated,” “mixed,” or “not yet clear,” followed by a sentence describing the observations behind that label. These are recordkeeping aids, not validated categories or thresholds. Do not assign points or a combined risk total to make the pattern look more certain than the notes. Separate early signals from outcomes that matter to the whole task. A successful demonstration, faster first pass, or fewer manual steps may be worth noting, but does not establish lower total effort or lasting change. More consequential observations include whether completed work meets the relevant quality standard, how much review and rework it takes, whether exceptions remain manageable, and where approval and accountability sit. The BLS “Artificial Intelligence (AI) exposure categories” methodology combines occupational measures into relative categories; those categories are not observations of a worker’s cycles. They provide context for what to watch, but cannot replace the local record or turn it into an individual forecast. When a pattern appears, write down competing explanations before attributing it to AI. A changed deadline, workload, policy, staff turnover, different input mix, or parallel process change may also explain a result. Note what changed and whether the comparison still seems fair. This does not require proving causation; it prevents a timeline from becoming a causal claim. “The Impact of AI on the Workplace: Evidence from OECD Case Studies of AI Implementation” draws on qualitative workplace cases. It illustrates that implementation may reorganize work and affect job quality, but cannot establish a universal sequence for a worker’s task. Document local conditions rather than assume a tool’s arrival explains every movement. A disciplined review can follow a short sequence: select one recurring task; identify the earliest reliable comparison; record later cycles with stable outcome labels; note workflow and context changes; then describe the pattern and its uncertainty in plain language. The NIST AI Risk Management Framework Core discusses documenting measurement outcomes in organizational AI risk management. Here, that is only an analogy for repeatable, traceable observations; NIST does not prescribe a worker’s career log or labor-market test. If evidence remains mixed, the honest result is unresolved. Keep observing or clarify the workflow. If repeated entries show a material change in review burden, responsibility, output expectations, or task volume, preserve that finding for a proportionate conversation or next-step decision. No fixed number of cycles settles every task, and a known workplace decision may make waiting unnecessary.

Sources: AI Risk Management Framework Core; Artificial Intelligence (AI) exposure categories; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

How should a personal record and occupation profile be compared?

Use an occupation profile to ask, “What task might I be overlooking?” Use a personal record to answer, “What is changing in my work here, over time?” Compare coverage, geography, dates, and methods; do not average them into a personal risk score. A profile suggests where to look; repeated local observations help decide what to clarify or test. O*NET offers a structured U.S. reference, not a diary of any one worker’s current week. Its database organizes work characteristics by O*NET-SOC occupation and includes task statements, ratings, and work activities. The O*NET Database page reports that the collection is updated quarterly, with a primary update in the third quarter, and directs readers to update summaries and archived releases. Record the version or update date consulted: a profile is a dated snapshot, not a live description of every workplace. Use the task statements as a prompt. If your notes mention preparing routine summaries but omit checking sources, handling exceptions, or explaining results, compare the profile and ask whether those duties belong to your role. A standardized occupation may include tasks you rarely perform, while your team may rely on specialized duties the profile omits. A mismatch does not make the reference useless: an occupation label is a starting map, while your record identifies recurring tasks in your setting. Exposure indicators answer another question again: how modeled AI capabilities or mapped use relate to occupational tasks under a particular method. In “Artificial Intelligence (AI) exposure categories,” the U.S. Bureau of Labor Statistics describes a 2025–35 projection supplement that combines five external sources, three theoretical and two based on observed usage mapped to occupational tasks or activities. BLS normalizes source results as relative percentile ranks and groups occupations into four categories. It warns that categories are not forecasts of employment growth, wages, productivity, adoption, or worker replacement; theoretical inputs concern capabilities available no later than mid-2023, while observed inputs may lean toward early adopters. A relative category cannot tell you whether your employer introduced a tool or changed your assignment. The International Labour Organization’s April 2026 brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” makes the methodological point broader: exposure indices differ in how they define and measure exposure, often rely on static task lists, and leave out adoption constraints and labor-market outcomes. It describes susceptibility, not what firms will implement or how employment and demand will adjust. This helps interpret a profile but does not measure your workplace. BLS is a U.S. reference tied to its occupational universe and projection cycle; O*NET is also a U.S. occupational framework. Elsewhere, use them as prompts, not local forecasts. When sources disagree with your notes, preserve the disagreement and investigate its shape. A high occupational exposure category alongside little observed change in your task record may mean the relevant capability is not used in your workflow, adoption is constrained, or the category covers tasks outside your current mix. A low category beside a recurring change in one duty may point to a narrow task-level shift that an occupation-wide label smooths over. These are hypotheses, not conclusions. Check task fit, source date, tool availability and approval, and whether the observation was a trial or routine use. Note workload, staffing, policy, or process changes before attributing an effect to AI. Keep the evidence types in separate columns: your dated record for observed local workflow; O*NET for standardized task prompts and its update context; and exposure indicators for method-specific occupational comparison. Compare like with like and note omissions. Do not combine them mathematically: one is an account of observed work, another a structured occupational reference, and the third a modeled or mapped indicator. For an immediate decision, use your record to identify the task, handoff, review burden, or responsibility to discuss. Use the profile to find blind spots and the indicator to frame questions. This triangulation gives context without turning unlike evidence into a forecast about one person’s job.

Sources: O*NET Database at O*NET Resource Center; Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t

When should the record change what you do next?

A single experiment usually calls for another observation or a small workflow test, not a career verdict. Repeated changes can justify learning a verification or domain skill, clarifying who owns review and workload, or comparing adjacent paths. A known employer decision or material change in role expectations can warrant action before a long record exists. Use the record to choose a next step, not calculate personal odds. When evidence is thin, keep observing the same recurring task. Note whether the tool was merely available, tried once, or used routinely; then compare similar cycles. If the task remains important and your goal is to do it better, test one improvement, such as a clearer input or review checklist. Judge the whole workflow: faster first output may still require more checking or correction. A trial result concerns that setup and task, not proof that the occupation has changed. Repeated changes call for a specific response. If drafts are routinely produced with a tool but you remain responsible for accuracy, identify which subject knowledge helps you catch errors and what verification the work requires. Choose a bounded course, guided practice, or small project according to the gap, time, cost, and feedback needed. If handoffs, approvals, review burden, output targets, or accountability have shifted, seek clarity. Ask which tasks are changing, what quality standard applies, who makes the final decision, and what training or time is available. This can distinguish a local workflow change from a capability demonstration or occupational estimate. Keep notes factual and do not put customer or employer information in an unapproved record. Consider a broader move when expected work repeatedly diverges from the role you want or can sustain. Compare an adjacent path using existing experience with redesigning the current role and with a larger change. Check actual openings and requirements where you can work; a task record says nothing about local hiring demand or wages. Set a salary floor and account for location, training time and cost, health, and family commitments. A changed task does not erase the value of experience. This ladder is a practical decision rule, not a tested threshold. The International Labour Organization’s 2026 brief says exposure indicators differ by method, rely on static task lists, and omit adoption constraints; they indicate technological susceptibility, not employment outcomes. The U.S. Bureau of Labor Statistics says its categories combine theoretical and mapped observed-use measures and do not measure job loss, productivity gains, automation probability, or wage effects. Neither source measures an individual workplace or supplies an action threshold. Workplace evidence also shows why a perfect score is unnecessary. The OECD’s 2023 qualitative study draws on nearly 100 manufacturing and finance cases across eight member countries. It reports job reorganization as more prevalent than displacement in those cases, alongside reduced tedium and challenges including increased work intensity. Its sector-bounded cases cannot predict your organization. They support recording task reassignment alongside workload and skill changes, then checking what is happening locally. Do not wait for several cycles if a credible staffing decision, announced role redesign, changed performance expectation, or immediate loss of work already affects you. Confirm what is known, preserve relevant nonconfidential records, and seek timely information or support available in your workplace and location. You can investigate options without deciding at once to leave. Let evidence and personal constraints set the scale of action, not a dramatic score.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; Artificial Intelligence (AI) exposure categories

What is the smallest useful next step?

Choose one recurring task and make a dated entry before concluding. Name the task and purpose, note whether AI was merely capable of assisting, actually tested, or used routinely, and describe which steps changed. Record the human checks that remained and one observable outcome, such as time, rework, quality, or workload. Mark unknowns rather than estimating them. Review the entry after comparable cycles, or sooner if the workflow changes materially. The aim is a usable record of work, not a personal loss score.

Keep the interpretation bounded. Capability does not establish workplace adoption; observed use does not by itself establish lasting task redesign; and neither reveals labor demand or predicts an individual displacement outcome. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” describes relative occupational comparisons and says they do not forecast adoption, employment, wages, productivity, or replacement. Treat an exposure label as a prompt to inspect your tasks, then let repeated local observations and context shape your next step. If evidence is thin, keep observing or test one workflow. If responsibilities or workload have already changed, clarify expectations rather than waiting for a longer record.

The free task checker can organize task-level change-pressure signals, but it does not provide a validated probability of displacement. Start with one entry; decide what it supports after observing the work.

Sources: Artificial Intelligence (AI) exposure categories