In a report published on October 16, 2026, Anthropic says surveyed staff expect AI could perform, on average, 75% of their current work hours by 2030. Employed US respondents put the figure at 28%. For workers and employers planning for change, these are expectations—not measured automation or company plans.
Neither percentage is a personal probability of losing a job. A task becoming feasible for AI leaves unanswered whether an employer will use it, how much supervision remains and what happens to demand for the work.
What the comparison can tell us
The August surveys covered 10,980 US adults and 1,334 staff. The public sample was online and opt-in, then demographically weighted; staff responses were unweighted. Demographic weighting cannot rule out other selection biases. The own-work comparison uses only employed public respondents asked about 2030, not all 10,980 adults. Staff used a slider; public respondents selected ranges, converted to midpoints.
The groups were estimating different jobs, too. The study does not establish which group forecasts better or how much occupation, experience and question format explain the gap. Proximity to AI development is not proof of forecasting accuracy.
Capability, adoption and time savings
Anthropic’s economic scenario explorer treats capability, adoption, autonomy, productivity and workers’ adjustment as separate inputs. A useful planning framework keeps capability, adoption and time savings separate:
Capability: Can the tool complete a particular task to the required standard? Identify the work involved, rather than treating a job title as one indivisible activity.
Adoption: Will the organization actually use it? A technically possible handoff still needs to fit the workplace’s rules, customer needs and available resources.
Time savings: After preparation, checking and correction, how much human effort remains? A fast first answer is not necessarily a fast completed task.
Consider a hypothetical 40-hour working week containing ten hours of tasks suitable for AI. Suppose AI is used for half of those task-hours and halves the time they require, including review. The saving is 10 × 0.5 × 0.5 = 2.5 hours, or 6.25% of the week—not 50%.
Those inputs are illustrative, not survey findings. The calculation holds work volume and quality constant and excludes transition costs and new tasks. It shows why a task-level speedup cannot simply become a whole-job saving, let alone a staffing target.
The same caution applies to economic scenarios. The explorer’s published limitations say it omits policy responses, business cycles and possible financial-market or demand disruptions. Its outputs describe consequences under assumptions; they do not establish which future will happen.
Why expectations need an outcome check
There is a concrete historical example of perceived and measured savings diverging. In METR’s early-2025 randomized study, 16 experienced open-source developers worked on 246 tasks in familiar projects. Allowing AI increased completion time by 19%, although participants later estimated that it had made them 20% faster.
That result concerns specific tools and work in early 2025, not today’s workforce. In its February 2026 update, METR said selection and measurement problems made its later experiment unreliable for estimating current effects. It believed benefits had probably improved, but could not reliably establish the size of that improvement.
The practical lesson is to ask for evidence from comparable work, not to assume either that enthusiasm proves a gain or that an old slowdown settles the matter. RohitAI’s analysis of adoption versus productivity explains the related distinction between growing usage and better results.
Prepare for change without choosing a forecast
Workers can ask for training that covers limits as well as useful applications. In a separate Elon University survey published on October 13, 52% of employed respondents rated guidance on AI’s limitations and when human judgment should take priority as extremely or very helpful. SSRS surveyed 1,158 US adults in August, including a worker oversample. This documents demand for guidance, not proof that a particular course works.
For a worker, a useful next step is to identify one recurring task and ask which parts an approved tool can assist with—and which decisions remain yours. For an employer, the corresponding commitment is to provide time, training and a clear standard for acceptable work before expecting savings.
A forecast can justify preparing for a possibility. A workflow change needs evidence about that workflow. A staffing decision needs more again: what work remains, what customers need and how responsibilities will change. The survey percentages cannot supply those answers.