Is AI Automating Your Work or Expanding It?

How can you tell whether an AI rollout is automating or expanding your work?

Compare one recurring workflow before and after the system is used: which steps it completes, which human contributions remain or grow, and what happens to review, exceptions, quality, workload, decision authority, and accountability. Automation means a smaller human share of a defined step. Augmentation means useful work expands or improves while the worker retains meaningful contribution and control. A rollout can do both. Faster output alone establishes neither worker expansion nor a staffing outcome.

Keep the comparison local. A model demonstration shows capability under demonstration conditions; an occupational exposure estimate indicates tasks that may be affected; neither tells you how your workflow changed. The OECD’s AI surveys of employers and workers illustrate why perspectives differ: its 2022 surveys covered 5,334 workers and 2,053 firms in finance and manufacturing across seven countries, with workers reporting experience and employers describing implementation. The OECD Employment Outlook 2023 reports varied effects on task content and autonomy, depending on implementation. These findings offer questions, not a diagnosis of your job. Compare work completed, including checking and exception handling, rather than classifying your occupation. Start with a work cycle and note what moved; do not infer a career decision from tool capability alone.

Sources: OECD AI surveys of employers and workers; Artificial intelligence, job quality and inclusiveness

What should you compare in the workflow before and after rollout?

Record the same recurring process before and after staff begin using the system, and compare periods with roughly comparable volume and case difficulty. Use the same definitions of “completed,” “corrected,” and “exception” in both periods; otherwise a changed reporting rule can look like a changed workflow. If the work varies widely, note the case type alongside each observation instead of treating every item as equivalent. Include review and correction, exceptions, coordination, elapsed time, and active human time. Then note who can approve, revise, or reject the result and who is expected to answer for an error. These observations show where work moved; they do not by themselves show what will happen to staffing. Keep the record simple enough to repeat. Add a brief note about quality: for example, whether required information was complete, a decision met the existing standard, or another person had to repair the output. If exact timing or production figures are inaccessible, describe a representative range or record a few non-confidential examples rather than inventing precision. Do not copy customer, patient, financial, or company-sensitive material into an unapproved tool or personal notes. The comparison needs a fair starting point. A launch demonstration shows what the system can produce under selected conditions; it does not show the ordinary mix of incoming work or the time required to check results. Likewise, the first days of use may include training, setup problems, or extra caution. Compare a normal baseline with a later period after people have had a reasonable chance to learn the approved process,. The OECD’s AI surveys of employers and workers illustrate why both views matter: its early-2022 study surveyed 5,334 workers and 2,053 firms in finance and manufacturing across seven countries, with employer questions about implementation and worker questions about first-person experience. Track tasks added as carefully as tasks removed. A system may draft or classify an item while a worker checks evidence, handles unusual cases, explains a decision, or coordinates a handoff. OECD’s qualitative study, “The impact of AI on the workplace: Evidence from OECD case studies of AI implementation,” draws on nearly 100 cases in manufacturing and finance across eight OECD countries. It reports both job reorganization and benefits such as reduced tedium, alongside challenges including higher skill requirements and reports of increased work intensity. Because this is a selected set of qualitative cases, it identifies possible mechanisms, not how often any one pattern occurs in every workplace. Finally, mark conditions that could make the periods hard to compare: a seasonal rush, a change in case complexity, different staffing, a new policy, or another software change. If one period contains easier requests, faster completion may have little to do with the AI rollout. If a clean baseline does not exist, say so; begin tracking a representative cycle now and revisit the comparison as more observations accumulate. The useful conclusion is bounded: these steps appear to have shifted, review or exception work has changed, or the evidence is not yet clear. That is enough to identify a question for a manager without claiming a precise productivity effect or predicting a job outcome.

Sources: OECD AI surveys of employers and workers; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

What does the evidence say workers experience after adoption?

The evidence supports several patterns rather than one default: a system may take over a recurring step, workers may acquire new duties, or both changes may occur in the same job. These findings map what to examine in a rollout; they are not a diagnosis of every workplace or a prediction about one employee. Keep separate what employers say they implemented and what workers say they experienced.

The OECD’s project record describes surveys conducted in early 2022 among 5,334 workers and 2,053 firms in finance and manufacturing across Austria, Canada, France, Germany, Ireland, the United Kingdom, and the United States. The paired design matters: employer responses provide information about how and why firms implemented AI, while worker responses capture first-person experience using it and responding to workplace changes. Neither perspective substitutes for the other. An employer can report a deployment without showing how it altered each employee’s day; a worker’s reported change does not describe organizational implementation choices or staffing plans.

The OECD Employment Outlook 2023 chapter, “Artificial intelligence, job quality and inclusiveness,” reports that workers experienced both gains and losses in control. In the 2022 AI surveys, 58% of finance workers and 59% of manufacturing workers said AI increased their control over task sequence; 20% and 21%, respectively, said it decreased that control. This is a specific measure, not a general score for autonomy, workload, or decision authority. These results resist a simple label such as “AI gives workers more control” or “AI takes control away.” The reported direction varied, and the measure cannot tell an employee whether they can reject a recommendation, obtain time to check it, or decide how much new work to accept.

The method and date set limits. The surveys were commissioned between mid-January and mid-February 2022. Employers were surveyed by telephone through management representatives at firms with at least 20 employees; workers responded through online access panels, which the OECD notes may exclude people without internet access and may undermine representativeness. Results concern two sectors, selected countries, and respondents’ perceptions at that time. They do not establish how common a pattern is across all occupations, whether AI caused every reported change, or what a new rollout will do elsewhere. Nor do these surveys measure a particular team’s before-and-after task mix directly.

The OECD project page says the surveys were repeated in 2026 with revised questionnaires and broader country and sector coverage. That update does not provide findings from the repeat on the page itself. Until results are available and checked, the 2022 figures remain historical context, not a current trend line. For a live rollout, use the survey evidence to frame local questions: which steps moved, what duties appeared, and whether control over sequence, pace, and review changed. Then compare the employee’s recurring work over a representative period. Survey patterns can guide that observation, but cannot replace it or predict an individual employment outcome.

Sources: OECD AI surveys of employers and workers; Artificial intelligence, job quality and inclusiveness

When can AI expand what a worker can do?

AI expands a worker’s contribution when assistance helps them complete useful work more effectively, with quality and the worker’s role still visible. More cases per hour can be one signal, but it is not enough by itself. Ask whether the person can resolve unfamiliar problems, produce a sounder answer, or learn a reusable method, with time and authority to judge recommendations. If the only change is a higher volume target, output has expanded while the worker’s capability may not have. A useful bounded example comes from “Generative AI at Work,” a field study of the staggered introduction of a conversational assistant among 5,172 customer-support agents at one Fortune 500 business-process software firm. The researchers measured issues resolved per hour and reported an average increase of 15% after access to the tool. The assistant offered response suggestions and links to internal technical documentation; agents remained responsible for the conversation and could edit, ignore, or accept suggestions. This describes one tool used with human discretion in one support workflow, not a general effect of AI on knowledge work. The average also concealed meaningful differences. Less experienced and lower-skilled agents improved in both speed and quality, while the most experienced and highest-skilled agents saw small gains in speed and small declines in quality, according to the study’s published abstract. The authors also found evidence consistent with learning: during periods when the system was unavailable, exposed agents retained some productivity gains relative to their pre-tool performance. They reported larger benefits on moderately rare problems, where agents had less personal experience but the system had enough training data to offer suggestions. That pattern points to a mechanism: useful assistance can fill a knowledge gap, while adding little—or creating a quality risk—when the worker already performs strongly or the system lacks relevant examples. For an employee, the practical comparison is not “AI made me faster, therefore it augmented my job.” Check a task where you can compare similar cases: did the tool help you address more difficult or unfamiliar work, and did the result remain accurate after review? Did you gain a reusable method, or did you spend the saved time correcting confident but unsuitable suggestions? Note what judgment you still supplied and whether you could reject an answer that did not fit. An approved, low-risk trial can answer these questions better than a demonstration, provided workplace rules are followed and work is checked. Even a well-measured productivity gain cannot establish higher pay, lower workload, greater decision authority, lasting demand, or job security. The study’s authors explicitly limit their findings to one tool, one firm, and one occupation, and say their data do not show aggregate employment or wage effects. Treat its result as evidence that expansion is possible under particular conditions, not as a forecast for your role. If your own review shows better quality or broader contribution, ask whether expectations and review time reflect that improvement. If the evidence shows only more output, separate throughput from capability before deciding what to learn or request next.

Sources: Generative AI at Work

What does a mixed rollout look like in real work?

A mixed rollout changes both sides of a workflow: the system takes over a defined step, while the employee’s remaining work shifts in content, difficulty, pace, or responsibility. The test is whether removed work is replaced by more varied or demanding work, and whether workers have support to do it well. A documented example appears in the OECD Employment Outlook 2023 chapter “Artificial intelligence, job quality and inclusiveness.” A financial services provider in the United Kingdom introduced a customer-service chatbot that directs customers to frequently asked questions. Representatives then handled fewer basic queries and a broader range of topics. The chatbot absorbed a routine contact while the human role shifted toward more varied conversations. A representative described the work as more interesting and varied. That single reported experience does not prove complex calls are better for every worker. The OECD chapter explains why this shift can cut either way. Fewer routine calls may leave time for involved conversations, but routine calls can also provide a mental pause. When they disappear, a shift of demanding cases may become more taxing. “The employee now handles harder cases” shows changed work, not necessarily a better job. Check whether new tasks bring learning and judgment, and whether training, pace, and recovery time remain workable. Other OECD cases caution against treating every transfer as an upgrade. The Employment Outlook chapter describes AI automating easy tasks while employees kept the same workload with fewer low-demand moments. In a Canadian manufacturing case, an AI developer said the firm retained some easy tasks as a mental break. This does not show that automation always worsens work; it shows workload can stay level as average effort rises. A useful comparison therefore has three parts: what the system now does, what the employee does instead, and what changed in the conditions for doing it. Name the removed routine and added task: “the chatbot resolves common questions; representatives handle a wider range of customer issues.” Then ask whether those issues require new knowledge, whether training and review time are available, whether expectations changed, and whether workers can manage the pace. Responsibility is not meaningful expansion if accountability rises without authority, information, or time to verify. The qualitative OECD study draws on nearly 100 workplace cases in finance and manufacturing across eight OECD countries. It identifies mechanisms, but does not estimate their frequency across jobs, establish an average causal effect, or reveal an employer’s plans. Treat the chatbot example as a case, not a forecast for every service team. Record both the step that shrank and the work that grew; compare difficulty, pace, discretion, training, and review capacity. Supported broader tasks may expand your contribution while automating part of the process. A heavier queue of demanding tasks may change the job without improving it. Task change alone does not establish a staffing decision: ask whether the mix is sustainable and conditions match responsibility.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; Artificial intelligence, job quality and inclusiveness

Why does more output not always mean more capability?

Higher throughput can reflect useful assistance, automation of an upstream step, a higher target, More completed work alone does not show gains in skill, discretion, conditions, or control. Interpret a productivity figure alongside checking, correction, difficult cases, handoffs, pace, and errors. If a system drafts routine responses faster, saved time might go toward unusual cases, broadening a worker’s contribution. It might also leave the same workload with fewer pauses or require close review before suggestions can be used. A response count alone cannot distinguish these outcomes.

The OECD Employment Outlook 2023 chapter on artificial intelligence and job quality describes both benefits and costs. In its finance and manufacturing worker surveys, 58% and 59%, respectively, said AI had increased their control over task sequence; 20% and 21% said control had decreased. These self-reports concern one aspect of autonomy. They do not establish how much review time a particular employee has or whether a rollout caused change. It also discusses AI increasing work pace or intensity. One mechanism: automation removes easier tasks but leaves the same overall workload, concentrated in harder cases and without the pauses those tasks provided. Output may rise or volume stay steady as the day becomes more demanding. The evidence supports checking pace, difficulty, and discretion alongside volume, not treating speed as proof of either empowerment or deterioration.

The OECD’s workplace case studies show varied outcomes. They describe task substitution that changed the range of duties, as well as implementations that left tedious new input work or targets that absorbed productivity gains. These cases show possible mechanisms, not how common they are. Task counts are a weak proxy for capability: fewer steps may leave a worker with greater responsibility for reviewing outputs, handling exceptions, or explaining decisions. Assistance may also support varied cases while retaining judgment. Count both work moved to the system and work retained or added around it.

In “Generative AI at Work,” a staggered introduction of a conversational assistant at one company covered 5,172 customer-support agents. The study reports a 15% average increase in issues resolved per hour, with effects differing by prior experience and skill: less-experienced workers benefited more, while the most experienced gained little speed and had small quality declines. It also reports evidence consistent with learning. This is evidence about one tool, company, and workflow. It does not establish durable skill gains for every worker, increased authority, lower workload, higher pay, or employment security; the study measured outcomes in that setting, not a general promise for other roles.

Treat throughput as a signal to investigate, not a verdict. Check whether quality held up, rework or review time changed, case difficulty shifted, and expectations rose with speed. If output increased while review time or discretion shrank, discuss workload, quality standards, and protected verification capacity with the person responsible for the rollout. More output is meaningful expansion when the workflow supports greater contribution, judgment, or scope. Without that evidence, a higher count remains a higher count.

Sources: Artificial intelligence, job quality and inclusiveness; The impact of AI on the workplace: Evidence from OECD case studies of AI implementation; Generative AI at Work

How can you tell whether expanded output comes with control?

Expanded output comes with meaningful worker control only when a person can understand the basis for a result, exercise a real choice about using it, and carry responsibility with enough time and authority to act. Ask who frames the task, checks the evidence, can correct or reject a result, makes the consequential decision, and answers for an error. If the system hides its inputs or leaves no practical time to review, human sign-off may be only a checkpoint. Control is not one thing. The OECD Employment Outlook 2023 measures autonomy as control over the sequence in which workers perform tasks. In the OECD’s 2022 surveys of finance and manufacturing, 58% of finance workers and 59% of manufacturing workers using AI said it increased that control; 20% and 21%, respectively, said it decreased control. These findings concern reported control over task sequence. They do not establish whether a worker could overrule an AI recommendation, choose assigned cases, or set the final quality standard. Output totals alone cannot describe autonomy. The same report finds that faster pace can accompany greater control, so pace alone does not settle whether a rollout is helpful. It reports that 75% of surveyed finance workers and 77% of manufacturing workers using AI said it increased their work pace. The report also describes more pervasive monitoring and cases where automating easier tasks left workload unchanged but removed lower-intensity intervals. These are bounded findings, not proof that every faster pace is harmful. They do show why volume, discretion, and intensity should be observed separately. For a particular workflow, make control visible with practical questions. Can you see the records or source material behind a recommendation? Do you know which parts are generated and which are verified? Are you permitted to correct an error or decline a result, and is there a named route for escalation when the tool fails? Does the review window match the cases’ complexity? Who decides acceptable quality, and who explains a consequential decision? A “yes” should describe ordinary work, not only what rollout documentation permits. The OECD project record says its 2022 employer survey asked how and why AI was being implemented, while its worker survey captured first-person experiences. Together, they covered 5,334 workers and 2,053 firms in finance and manufacturing across seven countries. The page notes a 2026 repeat with revised questionnaires and wider coverage but reports no results from it. Neither perspective establishes an individual’s override authority or reveals staffing plans. If you are accountable for an outcome but lack access to evidence, time to verify, a workable override, or a clear escalation route, describe that mismatch concretely. Ask the manager or rollout lead to define sign-off rules, review capacity, training, and who handles exceptions. The aim is to make responsibility workable and the process observable. This conversation cannot by itself establish the tool’s wider effects or predict employment consequences.

Sources: Artificial intelligence, job quality and inclusiveness; OECD AI surveys of employers and workers

What can occupational exposure tell you about a live rollout?

An exposure index can flag that an occupation includes tasks a technology may affect. It cannot establish that your employer adopted it, that a step in your workflow changed, or that your own job is at risk. Use the index to decide what to inspect; use evidence from the rollout to decide what is happening. The International Labour Organization’s 2025 paper, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” estimates how generative AI could affect occupational tasks by combining task information, worker input, expert assessment and model-assisted predictions. It draws on a representative sample of 29,753 tasks in Poland’s occupational classification, a survey of 1,640 employed people, and 52,558 data points about perceived automation potential for 2,861 tasks. Researchers compared this input with international expert surveys and discussion before developing scores for tasks in the classification system. That is a structured estimate of potential, not a record of live workplace use. The paper’s exposure gradients describe how occupational tasks may be affected; they do not show whether a team deployed a system, whether staff use it routinely, or whether managers redesigned the process. The paper says transformation is the most likely impact because most occupations still contain tasks requiring human input. That is a global assessment, not a verdict on an employee’s position or a guarantee that a role will remain unchanged. Keep the inference steps separate. Capability means a system can perform a defined operation under some conditions. Applicability asks whether that operation resembles an occupational task. Exposure estimates that potential across tasks. Actual use requires evidence workers or an employer use the system. Workflow redesign means duties, handoffs, review or authority changed in practice. Labor demand concerns the organization’s or market’s need for work; displacement is a further employment outcome. Evidence at one step does not prove the next. The OECD’s “AI surveys of employers and workers” shows why perspective and date matter. Its 2022 survey covered 5,334 workers and 2,053 firms in finance and manufacturing across seven countries, using separate instruments for worker and employer experience. The OECD page notes a 2026 repeat with revised questionnaires but does not report those results. Neither survey can diagnose an individual’s current workflow. For a live rollout, start with tasks an index highlights, then check the work itself: which steps the system completes, which remain with a person, what review or exception work appears, and whether quality, workload or decision authority changes. An index or task checker is a prompt for that comparison, not a personal job-loss score.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; OECD AI surveys of employers and workers

What can this workflow audit—and what remains unknown?

A workflow audit can establish what changed in the work you observed: which tasks moved, whether quality held, whether review or exception work grew, and whether authority or workload shifted. It cannot show a private staffing plan, settle what is happening across an industry, or predict your employment outcome. A bounded conclusion might be: “During these weeks, this process removed these steps and added these checks.” It does not establish a staffing decision. The sources answer different questions. OECD AI surveys of employers and workers describe complementary perspectives: employers report how and why AI is implemented, while workers report their experience using it. The OECD says its early 2022 survey covered 5,334 workers and 2,053 firms in finance and manufacturing across seven countries; surveys were repeated in 2026 with revised questionnaires and broader coverage. The project record does not report results from that repeat. They describe sampled populations, not what happened in your team. An occupational exposure estimate sits further from the local question. The International Labour Organization’s Generative AI and Jobs: A Refined Global Index of Occupational Exposure combines task data, worker input, expert discussion, and model-assisted predictions. Its 2025 method drew on a representative sample of 29,753 tasks in Poland’s occupational classification, responses from 1,640 employed people across ISCO-08 groups, and 52,558 data points about perceived automation potential for 2,861 tasks. This estimates potential exposure across occupational tasks. It does not establish whether your employer deployed a tool, whether staff routinely use it, how a team redesigned review, or whether staffing changed. An index can help select duties to inspect; it cannot fill those gaps. Before-and-after differences have competing explanations: case difficulty, volume, seasonality, policy, staffing, or learning may have shifted. A rollout announcement may precede routine use, and a target may change independently. These factors limit causal claims. If a task disappears during one unusual week, treat the observation as provisional. If the same allocation persists across comparable cycles, quality checks remain stable, and workers consistently describe the changed process, confidence in the local description improves. It still does not establish why management made a staffing decision. Look for sustained task changes across normal cycles, quality trends that include corrections and errors, changes in approval or override authority, and staffing evidence. Treat these as separate signals. Describe what was observed, the period and limits, and what remains unknown. Revisit if workload, authority, process design, or staffing information changes. This supports a proportionate next step without turning exposure into a forecast or local task change into a labor-market claim.

Sources: OECD AI surveys of employers and workers; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Generative AI at Work

What is a proportionate next move?

Start with one representative work cycle, not a career verdict. Record what the system completed, what you produced, what needed checking or correction, and whether quality, pace, or decision authority changed. Keep examples non-confidential and compare similar cases. The OECD Employment Outlook 2023 describes both job-quality benefits and pressures associated with workplace AI, including changes in autonomy and work intensity. That is why workload and control belong beside output. Bring your observations to a manager and ask: Which outcome should improve? How will quality and workload be assessed? Who can accept, correct, or escalate an output, and who is accountable when it is wrong? Agree on a review date to see whether the change persists in normal use. The ILO’s Generative AI and Jobs: A Refined Global Index of Occupational Exposure estimates task-level exposure; it does not show what your employer adopted or what a rollout means for your role. Choose the smallest step supported by what you observe. If an approved tool helps with a recurring task, learn its limits and verify results. If responsibility grows without review time, training, or usable authority, request those conditions. If routine duties remain removed, explore adjacent roles while accounting for salary, location, credentials, learning time, health, and family constraints. The free task checker can organize duties into change-pressure signals; it does not predict job loss. Revisit your decision if sustained workflow evidence changes.

Sources: Artificial intelligence, job quality and inclusiveness; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Questions readers ask