01 Facts & context
An index that separates "can do" and "can't do it"
2026-9-30, Anthropic published a study, What work can robots do? The study defines robots as autonomous physical machines capable of perception and action, using O*NET as a base table some 900 occupations and about 19000 mission descriptions, and using Claude to score the physical, cognitive and interpersonal properties of each mission, and to determine the environment in which robots can perform the task today. The study is divided into four levels according to the degree of control of the working environment: E0 cannot be completed, E1 can only be done in a dedicated robotic environment (e.g. plant line), E2 can be done in a structured place where people can operate (e.g. logistics warehouse) and E3 can be done in an unstructured environment (e.g. urban roads).
The study gives several key figures: robots can today complete the physical mission in the United States of America of 74%, which corresponds to a total of 34% working hours, but most are limited to the controlled environment; About one quarter of the entity ' s mandate is E0, E1 about half, E2 about 22% and E3 only about 2%. Combine with LLM, approximately80%Working hours are exposed to robotics or large models, and robots undertake tasks that large models cannot do. Nine of the highest levels of exposure are in the category of vehicle operations, with the taxi driver index 2.2 at the top; The categories of nursing and general maintenance are exposed to low levels of interpersonal and intellectual requirements.
Cost is the second-tier conclusion that has been repeatedly emphasized in the study: robots already cover a large number of tasks in terms of capacity, but are cost-competitive for only 0.3% missions; If prices fall according to historical trends, it will take 40 to reach 10%. The study also states that, according to data feedback from 1977, more exposed occupations have declined even more in wages and employment in subsequent decades, and robots have been able to add about 2% per year to physical jobs that could not have been done before. The sources are reproduced above, and the following sections provide an analysis of the operations of the GSP.
02 GSP analysis
Capability coverage does not amount to substitution: an enterprise has to first perform an “environmental classification” rather than a “vocational list”
What is most valuable in this study is not an occupational ranking, but the forced splitting of “robots” into two variables: “robots can do” and “robots can do it cheaply”. Most of the automation discussions mixed the two, either overestimating the replacement speed or miscalculating the technology when it is blocked by cost. The first strategic option for an enterprise is to recognize that level E reflects the degree to which the environment is controlled, and that the degree to which the environment is controlled is often determined by the enterprise itself — the position of the shelf, the pallet specifications, the layout of the workspace, the consistency of inputs are all variables that can be invested. Moving the task from E2 to E1 already available, sometimes not by robotic upgrades, but by making the site more like a warehouse.
This leads to an anomalous judgement that automated selection should be based not on “the most expensive job” but on “the environments in which the tasks are most easily structured”. A high-paying job with a chaotic work environment and a non-standard job may be more difficult to replace in the short term than a low-paying job with a highly standardized process. The study noted that exposure to workers is more likely to be male, less educated and less paid, which means that pressure to automate falls first on groups with less bargaining power, and that companies, if they move only in a cost order, introduce social and compliance risks.
Example of scenarios: A regional chain of meals divides the after-cooking into three types of tasks - order-based assembly of dishes (E1 potential), on-site handling of shortages of emergency food items (close to E0), and shop count (E2 potential). If the budget is limited, the first move should be to assemble and inventory rather than to appear to be the most labour-intensive on-the-spot disposal. The terms of the judgement are clear: whether the mission has a stable input, whether it can be covered by cameras and sensors, and whether the cost of returning to work after an error is affordable. The three are not satisfied, and even if the robotic demonstration video is again fluid, it should not enter the pilot queue.
03 GSP analysis
The value mechanism is to release manpower from the "structured" period, not to cut jobs.
The implicit value logic of the robotic exposure index is the mission particle size, not the post particle size. A post is composed of multiple tasks weighted according to working hours, and the value of the replacement is equal to the working hours ratio of the replaced tasks multiplied by the difference in the unit working hours, less the total cost of deployment, integration, maintenance and shutdown. The cost competitiveness ratio of the study, 0.3%, shows that this margin is still negative in most scenarios. Thus, the real landable gain in the short term is not to “reduce the number of people” but rather to release staff from low-judged, high-duration periods to study the interpersonal and intelligent work that E0 has been determined to be.
This mechanism has a different meaning for service- and manufacturing-type operations. The controlled environment of the manufacturing industry is naturally present, and the value of robotics is derived mainly from the stability and improvement of beats; The value of logistics and retailing is more derived from peak elasticity, i.e. the use of robotics to absorb peaks than to replace manpower throughout the year. The study referred to the fact that storage tasks were rated E2 and that more than three quarters of the hours spent by recycling sorting workers could be completed in E2, while the wallpaper workers had a break point where “some tasks can be highly automated and some tasks totally impossible”, indicating that bottlenecks determined whether the entire process could not be left unattended. An enterprise that simply replaces non-circumstances may receive a higher degree of complexity in semi-automated rather than a lower cost.
Example of hypotheses: A medium-sized sorting centre calculates that the sorting tasks represent a total of 60% hours of work and can be replaced by E2, but the unusual treatments account for 15% and are irreplaceable. If only robotics are selected without changing the abnormality process, the overall manpower decline is limited and more robotics are deployed. The preferred option at this time is to harmonize the package specification and labelling specifications and to press the anomaly ratio to the manageable area before going on to the equipment. The person responsible should be the operator, not the IT owner, since the standardization of input is a process matter and not a technical one.
04 GSP analysis
User ' s travel perspective: high-visibility jobs, experience risks before unemployment risks
Using robotic exposure as a dark line for staff travels, it is not the employment relationship that is the first to change, but the work content and the way in which it is evaluated. The study shows that drivers are the most exposed, but that autopilots have not changed the number of driving positions on a large scale, indicating that the intermediate situation may be long-term: people are still on duty, but are required to monitor machines, process marginalities that the machines cannot cope with, and assume higher unit time output. The risk of experience during this phase of “human aircraft co-location” is often overlooked in automated discussions.
From the point of view of the employee ' s journey, four paragraphs can be identified: the task has been partially taken over by the machine, the duties have been reclassified to surveillance and abnormal treatment, the assessment indicators have shifted from production to machine utilization and abnormal response and, finally, the number of posts has been adjusted. The E-level framework of the study can be directly mapped to the second paragraph: staff are transformed from implementers to environmental maintainers, whose skills require a shift from physical strength to diagnosis and coordination. If enterprises do not voluntarily rewrite job descriptions and training paths, they treat the machine as an alternative threat, thereby negatively collating with the abnormality process, while pushing up the overall cost of the system.
Example of the hypothesis: After a hospital had a robot on the line, the carer's work was changed from “delivery” to “checking the robot's distribution list and processing errors”. If the test is still based on the pre-existing prescription period, the nurse will tend to bypass the robot to deliver the medicine and the return on the investment will fall short. The solution is to include robotic abnormality rates in the common indicators of the unit and to clarify the role and authority of nurses in an abnormally closed circle. Responsibility for this section is vested in a consortium of human resources and front-line managers, with the input of a task-regulating list, an output of updated job descriptions and training modules, and a trade-off between short-term inefficiencies for long-term processes to be credible.
05 GSP analysis
Organizational division of labour: Automated budgets should be held by the operator, and technical parties should only veto feasibility
The evidence structure of the study points to the conclusion of an organization that, among the decision-making variables for robot substitution, environmental adaptation, task re-allocation and quality tolerance are operational competencies, except for interfaces, integrations and data. If budgetary and decision-making power is placed in the technical sector, two types of failures are likely to occur: first, in order of technical feasibility rather than economicity, and second, when on-line no one is responsible for process modification and robots are isolated from the original process. A more stable division of labour would be for the operator to have a budget and objectives, for the technical party to exercise a viable veto and an integrated responsibility, and for the financial party to independently account for unit mission costs.
Specific responsibilities can be broken down into three layers. The Head of Operations defines mission boundaries and acceptance indicators, including ceilings on the anomaly rate, rules on human switching, and staff redistribution programmes; (b) The technical officer evaluates level E accessibility, integration cycles and downtime risks and has the right to call off the non-achievement scenario; The FOS is responsible for calculating the dynamic ratio of robotic unit costs to human unit costs for each task, setting a rolling observation threshold based on the starting point of the study of 0.3%. There is a need for a common cost caliber between the three, otherwise “cost-for-money” becomes a rhetorical exercise.
Example of hypothesis: Which car parts plant chooses first between welding and material handling. Welding tasks can reach E1 in a dedicated environment, but have frequent changes; Material handling is E2 but complex. If the technology sector is in the lead, it may choose a more scalable welding; If it is operationally owned and the replacement shutdown loss is accounted for, it may be moved up to stabilize the beats. The trade-offs should include order fluctuations and replacement frequencies, the responsible person should be the director of the plant and the technical party should retain a veto on safety and integration.
06 GSP analysis
Inputs and alternative explanations: In addition to the cost curve, there are three gates of preference, regulation and capacity
The study clearly states that there are triple barriers to capacity, preferences and regulation in addition to prices, and that most robots need a highly structured environment. This means that “robot prices decline” is not a sufficient condition for automation. When an enterprise makes input calculations, the cost of integration, maintenance, site adaptation and retraining is systematically underestimated when the unit cost of the equipment is only included in the model, thereby overestimating the return. The evolution of cost competitiveness in the study from 40 to 10% is in itself a rebuttal of the perception that “a fall in prices necessarily leads to a rapid replacement”.
An alternative explanation that needs to be taken seriously is the historical decline in wages and employment, which may come mainly from trade, industrial transfers or changes in demand structures, rather than robotics themselves. The study states that feedback on controlled industry trends is confusing with other factors, but that external readers are not able to review it independently, and that enterprises should not use this conclusion as a basis for their own staff reduction decisions, but rather as one of the scenarios for stress testing. What is really useful to enterprises is that the same E-level methodology uses sorting results on its own field data rather than national statistics.
Example of hypothesis: A textile enterprise evaluates the automatic fabric and offers the equipment are acceptable, but the plant area humidity control and light modification costs more than the equipment itself, and the time of change for small bulk orders makes the machine underutilized. Logically, this is a dual failure of capacity and environment, which should be classified as a suspension rather than a rejection, and conditions change once the order structure shifts to a large volume equivalent. The terms of the judgement should be clear: order volume thresholds, change frequency ceilings, and re-engineering of the investment recovery period are not necessary.
07 Suggested actions
Implementation order: standardized input, then controlled environmental robots, and finally unstructured
In conjunction with the E-level framework, a conservative but adversarial order of advancement may be given. The first step is not to buy equipment, but to count the full process tasks using a single calibre, indicating existing and potential levels by environmental control and identifying bottlenecks. The second step is to standardize the input for E2 and E1 missions: to harmonize packaging, labelling, position layout and input traffic, and to compress the uncontrollable factors to the equipment ' s affordability range. The third step is to deploy on a small scale in a single place or in a single class to validate the anomaly rate and the frequency of manual intervention rather than the peak capacity. The fourth step is to consider cross-worker replication and E3 scenes.
The key to this sequence is that there are clear exit conditions for each step. If a standardized anomaly rate continues to exceed expectations, the procurement of equipment should be discontinued instead of an additional integrated budget; If the frequency of manual intervention in the pilot is higher than the threshold set, indicating that the actual level E of the mission is lower than the paper rating, it needs to be re-rated rather than simply expanded. Most of the deployments mentioned in the study are concentrated in factories and logistics warehouses, which are indicative of the fact that the controlled environment is a real border of current commercialization and that firms are moving forward against this border, with a significant increase in the probability of failure.
Example of the hypothesis: An electric warehouse first harmonizes the swing-box specifications and labels in this order, and then deploys robotics on a sorting line, using four weeks to measure only the anomaly rate and the number of manual interventions, before the target is met to the second line. The responsible person should be the head of warehouse operations, enter the list of tasks and standardized norms, export the pilot conclusions and conditions for replication, and trade-off at the expense of short-term on-line speed in exchange for replicable certainty.
08 Suggested actions
Indicators and experiments: measuring “mission replacement rates” rather than “decreasing”, using cost ratios as rolling gates
Most automation projects use cutbacks or productivity upgrades as the sole indicator, which could mislead decision-making at an early stage. A more appropriate core indicator is the mission replacement rate: the proportion of the total working hours of the post that the robot stabilizes, which is complemented by three complementary indicators of the frequency of manual intervention, the length of abnormally closed rings and the unit cost ratio. The ratio of mission substitution answers “How much is technically running”, the cost ratio answers “The economic value does not increase”, and the frequency of intervention answers “Stableness supports replication”. Three have met the criteria for expansion.
In the experimental design, it is recommended to use task-to-task rather than post-to-post comparisons: to compare unit cost and quality fluctuations in manual and machine operations under the same tasks, the same shift structure, and to avoid the conclusion that post differences are contaminated. The threshold is the starting point for the study - to document where the current cost margin of the enterprise is and then set a quarterly observation point, rather than setting an optimistic catch-up. The study says that robots can be added every year.2%The physical work that could not have been done previously means that the boundaries of capacity are expanding slowly and that the experimental mechanism should periodically re-run the rating rather than the one-off conclusion.
Example scenario: A food processing plant sets the rule that a copy to the second production line is authorized only if the unit cost of a mission is less than eight percent of the human capacity and the frequency of eight consecutive weeks of intervention is less than the given value. The effect of this gate is not to predict accurately, but to prevent overexploitation at a single point. The responsible persons are joint teams for business analysis and finance, with weekly running data entered and recommendations for expansion or closure exported.
09 Suggested actions
Applicable conditions and adjustments: which enterprises are to be expected and which should move early
The framework is not equally applicable to all enterprises. Even earlier, high levels of duplication of tasks, self-engineering of the environment, large and stable orders, high manpower costs and compliance pressures are typical of enterprises such as warehousing sorting, partial manufacturing of workspaces and vehicle operating scenes. What is more important is to wait for businesses whose tasks are based on interpersonal interaction and fine handling, where there is a high volume and change of style, and where there are many uncontrollable factors on the ground, such as nursing, maintenance and customization services. The study ' s diagnosis of nursing and general maintenance is supporting this division.
The adjustment judgement should bind several observable signals: whether the level of alternative environments for similar tasks has been upgraded, whether unit cost of equipment has been subject to a trend decline, whether regulations have liberalized the particular scenario and whether the anomaly rate of the enterprise has declined with standardized input. Continuous improvement of any signal can increase input weights; If the cost ratio has stagnated for a long time and the frequency of interventions has not decreased, it should reduce expectations or shift to a mixed model of partial automation and labour. The failure of the “mixed model” is the root cause of the additional input in the wrong direction of many projects.
The modesty that needs to be retained is that national indicators are no substitute for on-site measurement, that historical relevance does not guarantee future causes and consequences, and that the costs of research are derived from hypothetical price trends. The firm ' s judgement should be set out in its own data, and the mechanism for overturning the conclusions should be maintained on a regular basis. A truly robust automation strategy is not to predict when robots will arrive, but to always know at what level each of their missions is currently, whether it is worth it or not, and what the conditions for giving up are.
Research sources
- Can we predict the jobs robots will do? - Anthropic ↗www.anthropic.com ·
Prepared from the cited public sources with AI-assisted research. Reported facts and editorial interpretation are presented separately.
CONTINUE THE RESEARCH
Explore your question further.
Request the full research report and discuss your market, brand questions and business context with GSP.
Request the full research report ↗