How AI Is Changing Legal Work: Research, Judgment, and Responsibility

How AI is changing legal work: the short answer

For an early- or mid-career U.S. legal worker, the practical conclusion is that AI may change bounded digital tasks, but capability or occupational exposure alone does not show that an employer has adopted a tool or that a job will disappear. Start with one recurring task in your own work. Before trying AI on it, confirm that the tool and data are permitted, a qualified person can review the result, and the team has a measurable standard for accepting or rejecting it. If any of those conditions is unclear, ask the workflow owner first; do not test with client material on assumption. Some assignments may still be reorganized, so use the answer to plan a controlled next step or investigate adjacent work, not to presume either safety or displacement.

Why does the job title hide the real change?

A job title is too broad to tell you which parts of legal work may change. The relevant starting point is the task: a bounded unit of work, such as locating authorities on a defined issue, organizing documents, or preparing a first draft from an established record. An exposure measure asks whether some tasks appear technically applicable to a capability. It does not say that the capability performs the task reliably in a real matter, that a firm has put it into a workflow, or that the workflow has been redesigned. Those are different questions, and they require different evidence.

The official U.S. occupational profile **23-1011.00 Lawyers, O*NET OnLine** makes the breadth visible. Its listed work includes interpreting laws and rulings, analyzing cases, gathering evidence, advising clients, preparing legal documents, negotiating, representing clients, and supervising legal staff. Some activities produce or transform digital text; others depend on facts developed through people and records, a client's objectives, interaction with opposing parties or institutions, or responsibility for directing work. The profile is a standardized occupation description, not a time-use study. It does not tell an associate, paralegal, or other legal worker how much of their week is spent on each task, nor does it describe a particular practice group or employer's workflow. Its value is to show why the label “lawyer” cannot stand in for an individual's actual task mix.

This distinction also prevents a common inference error. A system may be capable of producing a draft or summarizing a supplied set of documents. That capability says something about a bounded operation under some conditions; it does not establish the quality threshold needed for a particular use. Occupational exposure aggregates possible task-level applicability across a role. Employer adoption asks whether an organization has approved, integrated, and assigned a tool. Redesign asks whether responsibilities, staffing, supervision, or the sequence of work have changed. Labor demand concerns the number and kind of workers employers seek over time. Displacement is a worker-level outcome that can depend on those changes and other conditions. Moving from one rung to another requires evidence; a technical demonstration cannot supply the whole chain.

The International Labour Organization's 2025 **Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO Working Paper 140** illustrates both the usefulness and the boundary of an exposure measure. The index combines task-level information with worker survey data, expert input, and model predictions to estimate potential exposure across occupations. Its broad conclusion is that transformation is more likely than full replacement in the aggregate. This is a modeled global assessment of potential task exposure, not an observation of U.S. law firms adopting systems, a forecast of lawyer employment, or an estimate of any individual's chance of losing work. It can guide inquiry into which tasks merit closer inspection; it cannot determine what is happening inside a specific office.

O*NET and the ILO index therefore answer complementary but limited questions. O*NET describes the variety of tasks associated with an occupation; the ILO index estimates where generative AI may apply to tasks across a global occupational landscape. Neither establishes which tasks dominate a particular person's day, which tools their employer permits, or whether saved production time changes staffing or client demand. A broad taxonomy and a global model can both conceal local variation: court rules, practice area, matter complexity, client requirements, team design, and review capacity can change the practical relevance of a capability. Treat these sources as maps for asking better local questions, not as classifications of your job. For example, a document summary created from a closed, approved record is a different workflow from advice that depends on current law, missing facts, or a client's unstated priorities. The same apparent task label can therefore conceal different inputs, error costs, and review needs; exposure categories cannot resolve those distinctions without local workflow detail.

For a worker, the useful unit is consequently a recurring task in its real setting: what information it uses, what judgment it requires, who relies on its result, and what must be checked before anyone acts on it. Two people with the same title may have different mixes; the same task may also carry different consequences depending on whether it is internal preparation or client-facing work. Mapping that difference is not a prediction of job loss. It identifies where a permitted, reviewable workflow might be examined and where the title-level evidence stops.

Sources: Lawyers: Occupational Outlook Handbook; 23-1011.00 Lawyers, O*NET OnLine

Where can AI assist with legal research and review?

AI assistance is most plausible where a legal task has a defined question, a bounded set of digital records, and an output a professional can compare with an authoritative source. That makes research synthesis, comparison of clauses or authorities, extraction of specified facts, and a first draft from an established record candidates for a controlled workflow. “Candidate” matters: a model can help organize or propose material, while usefulness depends on whether the result is accurate and complete enough for its purpose after review. The task label alone cannot settle that. A synthesis intended to help a lawyer locate relevant authorities has a different acceptance standard from a conclusion intended to support advice or be filed with a court. For each use, define what the system receives, what it is asked to produce, who relies on the result, and which source will determine whether it is right.

The Stanford HAI research summary, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries,” describes a preprint evaluation built around more than 200 challenging legal research queries. The selected questions tested issues including legal doctrine, jurisdiction and time, false premises, and factual recall. Across tested systems, the researchers identified materially incorrect answers, incomplete answers, and responses whose cited authorities existed but did not support the propositions attached to them. These are different failure modes. An incorrect statement of law can send analysis in the wrong direction; an incomplete answer can omit an exception or relevant authority; a citation that looks real can still leave a claim ungrounded. A source list therefore cannot substitute for checking what each source actually says and whether it applies to the question asked.

The benchmark is a stress test of the evaluated systems on deliberately difficult, manually constructed questions, not a measured error rate for routine work in a live firm. It does not establish how often a particular current tool fails on a particular practice’s ordinary tasks, nor does it measure net time saved after review. The plan’s evidence record also notes that specialized retrieval systems reduced errors relative to general models within the evaluated set. That comparison is a reason to assess systems and workflow design separately: access to a focused legal research tool may change performance on tested queries, but it does not establish error-free output or automatic suitability for a matter. Nor should one transfer benchmark results from one version or query set to every tool, question, or deployment.

A practical test should measure the whole work product rather than the speed of the first response. Select a recurring, low-consequence task with permission to use the relevant material. Set a manual baseline and define in advance what counts as an acceptable result: for example, whether all material authorities in the approved record are represented, whether quoted language matches the source, and whether jurisdiction, date, and exceptions are handled. Then compare the assisted output against the authoritative record, recording unsupported propositions, missing relevant material, corrections, and any exceptions a reviewer finds. Count the time spent framing the request, tracing claims, correcting the draft, and obtaining review. If an apparently fast response needs extensive reconstruction, the workflow may shift effort rather than reduce it. This is a decision method for the local task, not a productivity result reported by the Stanford benchmark.

Review should also be designed around the task’s failure cost. For internal issue spotting, the acceptable role of a tool may be to suggest questions for a lawyer to investigate, with no suggestion treated as exhaustive. For extraction, the reviewer can check sampled entries or, where consequences warrant, every entry against the record; sampling is only appropriate if the responsible team accepts the residual risk. For a research memorandum, the person responsible can trace each legal proposition to current primary authority and check that the authority supports the stated rule in the relevant jurisdiction. For a first draft, review includes checking factual assumptions and omissions as well as wording. These examples are workflow choices, not claims that any category is reliably automated. They make explicit which errors the process is intended to catch and who must do the checking.

The useful question is therefore not simply whether a model can produce a summary, comparison, extraction, or draft. It is whether the complete process produces an acceptable result with a review burden the team can support, using a permitted tool and data. A workflow that fails its quality threshold or consumes more checking time than it saves has not shown useful assistance, even if its initial output is polished. A workflow that meets a narrow threshold in a controlled trial may justify further evaluation, but it does not by itself authorize broader deployment or prove gains across matters. The Stanford findings make the comparison concrete: evaluate the final, checked work against the underlying record and count the labor required to reach it.

Sources: AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries

Why does responsibility stay with the professional?

A system can produce text, but professional use begins only when a responsible person decides whether the output may enter the work, how it will be checked, and whether it is fit for its intended purpose. Verification here means more than proofreading: the reviewer must have the knowledge and access needed to assess the work, compare it with the proper record or authority, and correct or reject it. Accountability means a named professional or supervisor owns that decision and the consequences of relying on the result. The ABA Standing Committee’s “Formal Opinion 512: Generative Artificial Intelligence Tools” explains how existing ABA Model Rule duties apply when lawyers use generative AI. It treats the tool as part of a lawyer’s work process, not as a substitute for the lawyer’s duties.

The opinion’s competence discussion turns tool selection into a practical question: does the lawyer understand enough about the tool’s capabilities and limits to make a reasoned choice for this task? A responsible process identifies the tool and version approved for the work, the kinds of tasks it may support, and what training or consultation is needed before use. It also names a reviewer with relevant legal and factual competence. If the person assigned to review cannot evaluate the output, a nominal approval step does not provide meaningful oversight. The right control may be to narrow the task, add qualified review, learn enough to evaluate it, or refrain from using the tool for that purpose. Those decisions should be set before the output is relied upon, not inferred from fluency afterward.

Confidentiality makes data permission an operational gate. Before entering information, ask whether the tool and account are approved for that class of client or matter data, what the applicable policy permits, and whether the information would be retained or used in ways the organization has not authorized. The ABA opinion discusses confidentiality duties in relation to generative AI; it does not turn any particular product setting into permission to disclose. If the rules for a client record, sensitive facts, or third-party material are unclear, the safe workflow question is to obtain guidance from the appropriate supervisor or information-security contact before testing. A task can be suitable in principle and still be unsuitable with real client inputs. A synthetic or otherwise approved test set may allow evaluation while the data question is resolved, if workplace policy permits it.

Communication duties require a separate decision about whether a client should be told that generative AI was used. Under the ABA opinion’s discussion of the Model Rules, that depends on the circumstances, including the engagement, client instructions, the nature of the use, and whether disclosure is material to informed decisions or the representation. The operational check is not “always disclose” or “never disclose”: determine whether a rule, agreement, client direction, or material feature of the service makes communication necessary, and who will make that communication. A workflow owner should specify how client expectations are recorded and escalated. This avoids treating internal tool use as automatically invisible or making broad disclosure claims beyond the governing circumstances.

Candor and supervision connect the output to filings, representations, and delegated work. The opinion addresses duties that arise when a lawyer submits material to a court or makes representations to others, as well as supervisory responsibilities for work performed by subordinate lawyers and nonlawyers. The practical questions are who checks every authority and factual assertion before external use, who confirms that required representations are accurate, and how the supervisor knows the reviewer actually performed the check. A tool-generated citation or draft cannot be treated as self-authenticating. Internal standards can define sign-off points, required records of verification, escalation for uncertain results, and tasks that may not proceed without a named reviewer. These controls locate responsibility in the existing professional workflow rather than in the software.

Fees and client value also affect whether a proposed use is appropriate. The ABA opinion discusses the Model Rules governing fees and billing, including the need for charges to be reasonable and for billing practices to reflect work actually performed. A firm considering an AI-assisted process should ask how time, tool costs, review effort, and client billing will be handled under the engagement and applicable rules. Faster generation alone does not answer whether a fee is reasonable, and a lawyer should not represent saved labor that has not been measured. The person responsible for pricing or billing needs a clear account of which work was done, which was delegated to a tool, and what human review remained. This is a governance question about transparent, permissible practice, not an assumed financial benefit of adoption.

Taken together, the ABA guidance translates into a short approval record: permitted tool and data; defined task and intended use; competent reviewer and supervision; required client communication; checks before relying on or submitting the result; and a billing approach consistent with the engagement and governing rules. A team can add task-specific acceptance criteria and a record of exceptions so that approval is auditable and a failed workflow can be stopped. Formal Opinion 512 is interpretive guidance on the ABA Model Rules, not a nationwide statute or a binding rule in every jurisdiction. State rules, court orders, employer policies, and matter-specific obligations may add requirements or differ. The responsible professional must therefore confirm the controlling local rules and workplace policy; a generic checklist cannot settle that inquiry.

Sources: Formal Opinion 512: Generative Artificial Intelligence Tools

What happens to the judgment people learn through junior work?

The available studies do not measure changes in law-firm assignments, the development of junior associates, durable independent legal judgment, or employment. They examine law students learning in bounded educational settings. They can therefore inform a question about how instruction and practice might interact with AI, but they cannot establish whether firms are removing useful repetitions from junior work or replacing them with better supervised learning.

In “Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis,” a randomized study of 164 law students compared three conditions around an issue-spotting exam: no optional LLM access, optional access without training, and optional access after approximately ten minutes of instruction. The abstract reports that tool use rose from 26% in the untrained access condition to 41% when access was paired with instruction. The trained-access group scored 0.27 grade points higher than the untrained-access group (p=.027); access alone did not improve scores against the no-access condition. The result is consistent with a bounded point: access and productive use are not interchangeable, and a small amount of task-relevant instruction may matter in this exam setting.

That finding does not show that ten minutes of instruction produces lasting expertise. The source record available for this assignment is the study’s abstract, not its full paper, so claims about detailed mechanisms, subgroup effects, or the exact quality of the exam intervention would exceed what was reviewed. The measured outcome was performance on an issue-spotting exam, not later independent work, retention, transfer to unfamiliar matters, or professional judgment under client responsibility. A score on a single assessment may reflect how a student used the tool during that assessment; it is not the same construct as being able to identify a novel issue months later, explain why an authority controls, or recognize when a seemingly plausible answer rests on a missing fact. The difference between the randomized arms supports the reported comparison in that setting; it does not tell a firm what training package to adopt or how much supervision a junior lawyer needs. Because the available record is abstract-only, even the study’s reported association between use and score should not be expanded into a detailed causal account of what students learned.

A separate study, “AI-enhanced legal pedagogy: a case study of generative AI integration in property law courses,” followed 180 third-year law students at one Chinese research university through a semester-long property-law course. It used a mixed-methods quasi-experiment with non-equivalent intact groups: the course integrated instructor-guided use of iFlytek Spark for legal reasoning, while the comparison group followed the other course condition. The article reports higher post-test legal-reasoning performance for the AI-integrated group, with a between-group effect size of d=1.17. This is a substantial immediate difference in the study’s measured outcome, but assignment was not randomized. Differences between intact groups, instructor enthusiasm, student motivation, or other features of the course could contribute to the result.

The semester study provides a different kind of evidence from the randomized exam study: it concerns guided integration over a course rather than brief training around one assessment. Yet its scope remains narrow. The study took place at one institution, in one country and subject, using a particular platform and course design. It did not include a delayed post-test to show that the measured advantage persisted, and it did not test transfer to workplace matters or unsupervised professional decisions. An immediate post-test can capture near-term course learning, but cannot establish retention after instruction ends or independent use in a new setting. The article’s mixed methods add context to the course implementation, but they do not remove the selection and group-comparability limits of a quasi-experiment. Its effect size should not be read as a prediction for U.S. law students, law firms, or junior associate development.

Read together, the studies make guided learning a plausible condition to investigate, not a proven employment strategy. In the randomized exam, optional access without instruction did not produce the reported score gain; in the semester course, AI use was embedded in instructor-guided learning, but non-random groups leave alternative explanations open. The studies do not isolate which repetitions build judgment, whether AI displaces those repetitions, or whether a supervisor can replace them with richer assignments. They also do not establish that reviewing AI output alone teaches the reviewer how to reason independently. Tool access or review exposure, by itself, is not evidence of learning.

For a workplace, a cautious practical inference is to treat junior development as a workflow question alongside speed and quality. If a firm introduces AI into research, drafting, or review, it can preserve explicit opportunities for juniors to frame issues before seeing a generated answer, check reasoning against primary material, explain corrections, and receive feedback from a qualified reviewer. Supervisors can distinguish a task delegated for production from one assigned as deliberate practice, and track whether the junior can later perform or explain the reasoning without the tool. For example, a supervisor might ask a junior to write an initial issue map before tool use, compare it with the resulting research, identify a material omission, and then discuss why the correction matters. That sequence makes the learning objective visible and leaves a person accountable for the legal reasoning. The team could also rotate who drafts and who critiques, so that tool-assisted production does not become the only repeated form of junior participation. Any such arrangement should be judged by observed work quality and learning over time, not assumed effective because it includes a tool or a review step. It is an illustrative way to operationalize guided practice, not an intervention evaluated by either source. These are proposed mentorship practices inferred from the educational evidence and the nature of the work; neither study tested them in firms or demonstrated that they offset lost assignments.

There is a credible counterpossibility: if routine preparation takes less time, teams could redirect that time to complex analysis, client context, and closer coaching. Student studies make it plausible that structured guidance can coexist with AI use in learning environments. They do not show that employers have made that redesign, that freed time will become teaching time, or that any richer assignments compensate for a loss of repeated foundational work. For a junior legal worker, the relevant local evidence is therefore the actual assignment mix and feedback available—not an assumption that AI either destroys or automatically improves judgment development.

Sources: Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis; AI-enhanced legal pedagogy: a case study of generative AI integration in property law courses

What can labor evidence tell you—and what can it not?

The current U.S. national baseline is growth, not a prediction about AI: the Bureau of Labor Statistics’ “Lawyers: Occupational Outlook Handbook” projects lawyer employment to grow 5% from 2025 to 2035 and reports about 28,700 openings per year on average. BLS explains that annual openings include positions arising when workers transfer to other occupations or leave the labor force, as well as new positions. These are occupation-wide projections for the United States; they do not isolate junior hiring, a practice area, a city, or the effect of AI. They should be read as a baseline for the occupation’s projected size, not as evidence that every legal employer will expand or that demand will be evenly distributed across roles. “Openings” also should not be paraphrased as newly created jobs: replacement needs contribute to that annual average. The forecast provides a demand baseline across the occupation, not a count of positions available to a particular graduating class, nor a statement about the quality, pay, or experience level of those positions. It cannot tell a reader whether a specific office will hire this year.

Set beside that outlook, the International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO Working Paper 140” estimates potential exposure of tasks in a global model and says transformation is more likely than full replacement in the aggregate. The two sources can both be true because they measure different things: BLS projects national employment and openings across a decade, while ILO models task-level applicability to generative AI. The exposure model does not explain the BLS projection, and the projection does not show how exposed tasks will change. Neither supplies a forecast of displacement for a particular junior lawyer. An occupation could have positive projected employment while employers alter the mix of tasks, change the number of entry-level positions in some settings, or expect different capabilities from new hires. Conversely, task exposure in a model can coexist with continued demand if work expands, other tasks remain valuable, or organizations adopt tools slowly. These are ways the measures could coexist, not findings that any particular mechanism is occurring in legal services. They also show why a national forecast and a task index should not be averaged, ranked, or converted into one synthetic “risk” figure: one counts projected employment change, the other classifies modeled task applicability. A percentage in one measure has no common denominator with a percentage in the other. Even a valid occupation-wide forecast can conceal differences by seniority, geography, specialty, and employer strategy, while an exposure estimate says nothing by itself about the timing or scale of adoption. To claim that exposure caused hiring contraction would require evidence linking adoption and redesigned work to employment outcomes, which these sources do not provide.

For an individual U.S. legal worker, those national figures are context for a local inquiry. The decision-relevant signals are which assignments are actually changing, whether the employer has approved tools for the relevant data, who reviews the work, and whether junior roles still provide supervised practice. Current postings in the worker’s geography can show what employers are asking for in that market, but a sample of postings is a local snapshot, not a validated forecast of hiring or job security. Read postings for repeated requirements, experience levels, and evidence of actual role design, then compare them with accessible conversations or public employer information where available; one listing may be aspirational or atypical. Employer-approved tools and supervision matter because a technical task may remain human-led when policy, data permissions, or review capacity constrain use. Positive national growth does not guarantee a stable entry-level pipeline or an opening in a preferred practice; modeled exposure does not establish contraction. Together, local assignment patterns and geographically relevant demand signals can guide what to investigate next, while remaining observations rather than a personal employment forecast.

Sources: Lawyers: Occupational Outlook Handbook; Generative AI and Jobs: A Refined Global Index of Occupational Exposure, ILO Working Paper 140

Which move fits your legal experience and constraints?

The right next move depends on what your workplace permits, what work it recognizes, and what you can afford to change. For many early- or mid-career legal workers, the most reversible option is to deepen capability inside the current role: learn to frame a task clearly, use an approved tool only within its permitted data rules, check its work against authoritative material, and explain where it fails. That is an upgrade in legal practice, not automatically a move into AI engineering. It makes sense when a real workflow is available to examine, a qualified person can review the result, and the employer values the added work of verification, escalation, or process design. If those conditions are missing, self-directed experimentation with client material is not a career strategy; first clarify policy and whether a safe, authorized trial can be arranged.

An in-role upgrade is relatively low-cost because it builds on legal knowledge and existing relationships, and it can be tested before a larger commitment. But it is not cost-free. Learning, documenting checks, correcting errors, and helping colleagues adopt a process take time. Ask who owns that responsibility, whether it will be recognized in assignments or performance discussions, and whether the workflow actually benefits from assistance after review effort is counted. An employer may not approve a trial, may not provide a suitable reviewer, or may expect new responsibilities without time or recognition. In those cases, the useful evidence is the organization’s actual policy and role design, not a general claim that every lawyer should become an AI specialist. The upgrade is attractive when it strengthens work you want to keep doing and when the institution supplies enough authority and review to do it responsibly.

An adjacent move can preserve legal context while changing the balance of tasks. Legal operations may involve process design and matter workflows; compliance work may apply rules and controls; knowledge management may organize reliable internal guidance; contract management may coordinate review and obligations; implementation work may help configure or introduce systems for legal teams. These are examples of directions to investigate, not claims that they are expanding, easier to enter, or protected from AI. Before investing in a course or changing your title, inspect current postings in your location and at the level you could realistically seek. Compare repeated requirements with your evidence of capability: for example, process mapping, policy interpretation, contract review, data handling, stakeholder communication, or experience supporting a rollout. Separate essential qualifications from preferred ones, note whether legal credentials are required, and look for a credible route to fill gaps through a bounded project, supervised assignment, or targeted learning. A list of attractive job titles cannot show that openings exist locally or that your experience transfers without additional proof.

A larger change—such as leaving legal work for software development, data work, or a different profession—can be rational when the current path has no viable route or the destination better fits the person’s aims. It carries more uncertainty and usually requires a clearer test before substantial spending or lost income. Specify the target role and geography first, then check its actual credential and experience requirements. Estimate the time needed to reach them, tuition or course costs, periods of reduced earnings, and whether part-time learning or a project can test interest and aptitude before committing. Include practical constraints that career advice often leaves out: caregiving, health, immigration or licensing conditions where relevant, commute and remote-work availability, savings, and the effect of a transition on household income. These factors do not make a large change wrong; they determine which route and pace are feasible. A degree may be necessary for a particular destination, while a portfolio, internal transfer, or supervised project may be a more proportionate first test for another. Verify the destination’s requirements rather than assuming one credential opens every door.

The OECD working paper “Artificial intelligence and the changing demand for skills in the labour market” reports that most workers exposed to AI will not need specialized AI skills, based on cross-occupation labor-market analysis. That finding is a reason not to make an AI degree the default response to change in an existing legal role. It does not establish what a particular firm wants, whether any course pays off, or what skills a specific legal-technology position requires. Deeper technical training is sensible when the work you actually want—such as building software or implementing technical systems—calls for it, and when the requirements recur in credible local postings or are confirmed by the employer. For an in-role upgrade, foundations such as problem framing, data literacy, evaluation, domain knowledge, verification, and basic automation may be more relevant than a specialized credential. Choose learning against a named task or destination, and ask what observable work it will let you demonstrate. The practical comparison is not “AI training or no AI training”; it is whether a particular investment closes a verified gap at an acceptable cost and time.

Use a staged decision rule. First, try an in-role upgrade only if the task, tool, data permission, reviewer, and recognition of added responsibility are real. If the workplace offers no such route, explore adjacent roles by checking local openings and attainable requirements, then test the most important gap with a small piece of relevant work. Consider a larger change when the destination is specific and its credential, training time, cost, income effect, location, health, and family implications are explicit. Keep options open where evidence is thin: a conversation with a workflow owner or a small learning project can reveal more at lower cost than an immediate degree commitment. None of these paths is inherently secure. An adjacent role can also be redesigned; an in-role trial may be unavailable; and a larger transition may be the sound choice when current work has no credible future for you. The point is to match the size of the commitment to the strength of evidence and to constraints you cannot wish away.

Sources: Artificial intelligence and the changing demand for skills in the labour market

What should you do in the next two weeks?

Start by asking the workflow owner three concrete questions: which tool and data are approved for this task, who has authority to review the work, and what errors or exceptions mean it must be rejected? If any answer is missing, do not test with client material; clarify policy or use the time to compare local adjacent-role postings with your experience. If all three are clear, choose one recurring, bounded task and compare the approved assisted workflow with the manual baseline. Record elapsed time, corrections, missed exceptions, review effort, and what you learned; one short comparison informs a next decision but cannot prove lasting productivity or job security. Keep, narrow, or stop the trial based on those observations, note any unresolved risks, before making another commitment, then decide whether to build the skill in your current role or investigate another path. The free [AI job risk checker](/ai-job-risk-checker) offers transparent task-level change-pressure signals, not a validated probability of displacement. If you need to compare scenarios against your constraints, the [career roadmap](/career-roadmap) provides a personalized plan; it does not guarantee income or employment.

Questions readers ask