2026年新工作新世界：AI如何以超预期速度重塑工作报告（英文版）.pdf

Our freshly updated research on AI and jobs reveals disruption that is more extensiveand swiftthan we anticipated three years ago.What we predicted would happen in a decades time is already here.How AI is reshaping work faster than expected83Three big changes in AIExecutive summary25Getting ready for the$4.5 trillion labor shift 31About the authors10The most and least impacted jobs19What this means for the workplace Table of contents3 New work,new worldA few years ago,we made news with a study predicting an astounding 90%of jobs would be disrupted by AI in less than a decades time.As it turns out,though,we underestimated the impact of the technology.What we projected might take until 2032 to unfold is happening now before our eyes.Todaysix years ahead of schedule93%of jobs could be impacted in some way by AI.In the US alone,this could add up to about$4.5 trillion worth of labor shifting from humans to AI.The technology,in short,is affecting more jobs,faster,and to a greater extent than we anticipated.We came to these findings by updating our 2023 research on AI and jobs,in which we assessed 18,000 tasks performed by 1,000 professions in terms of the degree to which they could be automated or assisted by AI.A lot has happened in the ensuing three years.Since then,AI models have become increasingly adept at interpreting many types of input,including images,diagrams and video.Further,more sophisticated AI models have emerged with advanced reasoning capabilities.And finally,AI agent-driven systems are now capable of completing complex workflows with minimal human oversight.Understanding the$4.5 trillion economic impact of AI We wanted to quantify the economic value of the total amount of work that AI could assist with or automate today.To do that,we used data from the US Bureau of Labor Statistics on the number of employees in each of the occupation groups in our study and then multiplied that by these employees median annual salaries.Then,using our exposure scores,we assessed the amount of this total economic value that could theoretically be exposed to AI.The result4.5 trillionis based on the assumption that the traditional way a task is completed would shift seamlessly to AI.Despite being theoretical,however,the calculation provides a glimpse into the sweeping economic change AI could bring.4 New work,new worldWith these three advancementsmultimodality,advanced reasoning and agentic AIit was time for a new look at how AI could reshape the workforce.So,we conducted a thorough reevaluation of the 18,000 tasks,this time through the lens of AIs enhanced potential to assist or automate them.What we found:Across all occupations,average exposure scores(i.e.,the degree to which an occupation could be affected by AI)are an astounding 30%higher than what wed forecast theyd be by 2032.(See explainer box for more on the exposure score.)In fact,while our original analysis found an average 2%annual increase in exposure scores among the jobs studied,we are now seeing a 9%annual score increase.As a result,some jobs that seemed safe from change when large language models(LLMs)first became mainstream are now capable of being affected much more quickly(see Figure 1).Understanding exposure scoresTo calculate the exposure score,we used the same approach as in our original research.We examined 18,000 tasks and close to 1,000 jobs in the O*NET database,assessing the tasks for automatability on a five-point scale(not automatable,minimally AI-assistable,partially AI-assistable,mostly AI-assistable and fully automatable).In addition to how many of the jobs tasks could be automated or assisted by AI,we also considered the relative importance of the task.We used an AI model to get an initial assessment of task classification but then reviewed and reclassified the findings where necessary.Using that analysis,we calculated an exposure score for each profession.The score reflects the degree to which an occupation could be affected by AI.A higher score means a higher percentage of the jobs tasks are automatable,so people in that profession could be greatly affected.The analysis in this report is based on a fresh assessment of AIs capabilities,reflecting the rapid evolution of the technology over the past three years,specifically its multimodal,reasoning and agentic capabilities.The resulting exposure scores represent a theoretical maximum:what current AI technology could potentially accomplish with optimal implementation.The scores do not account for enterprise adoption,employee acceptance,regulatory frameworks,quality control requirements,ethical considerations or the substantial organizational change required to deploy AI at scale.For these reasons,the exposure score represents a raw calculation of the technologys potential.As such,it reflects capability and opportunity rather than inevitability.Throughout the report,the term“exposure score”is a theoretical exposure score.5 New work,new worldSource:Cognizant Figure 1Across the board,AI-driven change is both more extensive and happening more quickly than anticipated.Our 10-year forecast is happening today6 New work,new worldSource:Cognizant Figure 2The percent of jobs with the lowest exposure has shrunk from 31%to 7%,while the percent with the highest exposure has grown from 0%to 30%.More jobs are more highly exposed*Calculated in 2023 *Calculated in 2026Exposure score range7 New work,new worldTo understand which jobs and job families are experiencing the fastest surge in AI exposure,we also calculated a velocity score,which quantifies the difference between the original trajectory of change in exposure scores over time and the new trajectory based on our refreshed analysisSee explainer box for more on the velocity score.)Examples of occupations showing unexpectedly high velocity scores,particularly when compared with their exposure levels in the original research,include:Decision-making roles Managerial and supervisor jobs are now increasingly exposed due to the emergence of agentic AI.Previously,these roles were more insulated from disruption because they involve complex coordination and judgment.Agentic AI alters this dynamic by moving beyond analysis to execution.Where managers once spent significant time allocating resources,monitoring project status or triaging workflows,autonomous agents can now orchestrate these duties.Project managers,for example,could rely on agents to autonomously schedule meetings,reallocate budget based on spend patterns and chase status updates by leveraging tools they are integrated with.Hyper-specialized sectors Such as healthcare,education and law.In these sectors,AI has quickly moved from assisting with low-level tasks to automating more complex tasks that are critical to the role.For instance,AI is revolutionizing healthcare by improving diagnostic accuracy and supporting patient care.In education,it can facilitate student assessment and classroom discussions.In law,it can analyze probable outcomes and assist with contract negotiations.Roles dominated by manual-labor tasksOnce considered a safe haven from AI disruption,many jobs requiring a great degree of physical labor now show significantly higher exposure scores compared with our original research,along with unexpectedly fast velocity.In construction,for example,AI can now help with interpreting blueprints.In transportation,it can inspect shipments or conduct safety reviews.The idea that a car mechanic or plumber can put on a pair of AI-augmented glasses to assist in locating a faulty engine part or leaking pipe is now far from science fiction.8 New work,new worldAbout the velocity score To identify the roles and occupation groups that have seen the most accelerated forms of disruption,we developed an additional measurement:the velocity score.The velocity score represents the difference between the original annual acceleration rate of exposure scores for any given job and the updated rate.This metric reflects how fast the pace of change could be for any given occupation,given the most recent advances in AI.A low score indicates that the latest changes in AI have impacted the role to a relatively minor extent.A high score reveals that the latest innovations will impact the role significantly.Source:Cognizant9 New work,new worldConsider how AI could expand into the physical and operational layers of work 1.2.3.4.Help people adapt as quickly as the systems they use Move toward a more adaptive operating model Build skilling systems that absorb capability shocksIn this report,we identify the biggest AI advancements in the last three years and why this has accelerated job impact.We also highlight the job families that could see the mostand the fastestchange,as well as some job groups in which change may be less dramatic but is still more extensive than originally anticipated.We also provide guidance on how business leaders can navigate the changes ahead.By embracing the following mindset shifts,businesses can better plan for the disruption to their workforce that is happening more quickly than imagined.A$1 trillion productivity storywith a catch本报告来源于三个皮匠报告站（）,由用户Id:619989下载,文档Id:1068218,下载日期:2026-01-2011 New work,new worldSource:CognizantFigure 3More tasks are more automatable by AIThe pace of change in work is now inextricably tied to the acceleration of AI itself.In 2023,most LLMs in use by businesses operated like narrow savants.They could generate text and code with fluency but had little grasp of planning,context or consequence.Today,advances in AI capabilities enable entirely new types of work to be automated or assisted by AI.Consider that in our analysis,one-third of all occupational tasks remain“not automatable.”However,the percent of tasks we classified as“fully automatable”has risen to 10%from 1%three years agoand just two percentage points short of the 12%originally forecast for 2032.Even more revealing,nearly 40%of all tasks can now be classified as being“partially”or“mostly”assistable by AI vs.just 15%previously.This also exceeds the 2032 forecast of 31%.This middle category is where change is most intense,as agentic systems become operational(see Figure 3).With that in mind,here are the three key AI capabilities we considered when updating our job exposure scores:Three short years;three big changes in AI capabilities12 New work,new worldMultimodal models provide AI with the eyes and ears that help connect digital systems with the physical world.These models can parse images,diagrams and video,recognize spatial relationships,and cross-reference visual data with text or numerical inputs.Where traditional AI could only describe the world,multimodal AI can interpret it.This new digital-physical connection has real occupational consequences.Jobs involving design review,product testing and quality control were previously beyond AIs reach because they relied on visual comprehension.Now,true multimodal models can evaluate design layouts,identify defects in manufacturing lines and assess the completeness of building construction from site photographs.Combined with sensor data and robotic integration,multimodality extends automation into the tactile and perceptual fabric of work.As a result,these types of jobs have climbed the exposure scale sharply.AI can now understand images,diagrams video and spatial relationshipsJobs involving design review,product testing,maintenance and quality control are now more highly exposed1.Multimodal AI:creating systems that see Three short years;three big changes in AI capabilities13 New work,new worldReasoning was once the missing ingredient in AIs cognitive repertoire.Early generative models produced fluent language but faltered on multistep logic or long-term coherence.The breakthrough came with structured reasoning frameworks and reinforcement-style fine-tuning.This culminated in models that demonstrate consistent,transparent chains of thought,enabling them to test hypotheses,deconstruct problems and evaluate alternative strategies.This reasoning ability has reclassified entire clusters of cognitive work.Analytic tasks,such as those found in consulting,finance and law,have shifted from being partially to mostly assistable by AI.For instance,a market analyst can now prompt an AI to not only summarize market data but also identify outliers,construct scenario models and justify recommendations with evidence.Audit and compliance tasks could now be entirely executed by reasoning agents that understand both numerical logic and procedural context.Planning,forecasting and diagnostic problem-solving are now within the operational domain of AI systems.New reasoning models can tackle complex cognitive activitiesExposure levels have risen sharply for people who do planning,forecasting and diagnostic problem-solving 2.Expanded AI reasoning:creating systems that think Three short years;three big changes in AI capabilities14 New work,new worldThe defining feature of AI in the post-2024 landscape is its agentic capability.If multimodality gives AI eyes and ears,and reasoning expands AIs mental map,agentic capabilities give it hands.This is because AI systems no longer stop at generation;they can take meaningful action.Consider that new technologies such as Model Context Protocol servers,intelligent function-calling systems and secure tool integration now allow AI to work directly with core business platforms.AI agents can now work together to fetch live data,execute commands in third-party software and monitor results for feedback loops.For example,multiple marketing agents can plan a campaign,query databases for segmentation,create advertising assets,schedule social media posts and report performance,all through connected tools.This agentic capability has pulled many administrative and coordination tasks further into the high exposure score zone.Schedulers,office administrators and project assistants once saw limited exposure because AI could not manipulate enterprise software directly.Now,the boundary between“knowledge work”and“process work”is fading as systems handle execution as well as instruction.Agentic capabilities have also changed management itself.Supervisory tasks,such as allocating work,checking progress and escalating issues,can increasingly be mediated by autonomous systems.In hybrid environments,teams of human and machine agents already collaborate across shared dashboards,with AI handling workflow triage and exception management.Supported by new technologies,AI systems can now take meaningful action Schedulers,office administrators,project assistants and supervisory roles have moved from limited to high exposure levels3.Agentic AI:creating systems that act Three short years;three big changes in AI capabilities15 New work,new worldEach of these capabilities is powerful in and of itself,but taken together,their power is compounded.Multimodality provides richer feedback,reasoning improves an agents decision quality and agency gives it control over the environment.Th