Anthropic reports that Claude “leads” 26% of its R&D for August 2026, with human supervision. Its index classifies and weights work categories, dividing each person-week equally among that person’s tasks. July supplies the reference basket. This internal result therefore does not directly measure hours saved. The method and appendix raise a practical question: what else must be observed to move from delegation to productivity?

Units determine what a number can mean

Counting tasks seems intuitive. Yet fixing a configuration error and choosing an experiment that distinguishes competing hypotheses are not necessarily comparable units. Before interpreting a percentage, identify its denominator: operations, hours, money or accepted results. A measure can be consistent in its own units and misleading when another unit is silently substituted.

Epoch AI’s proposal for mapping R&D, published on June 17, offers an activity taxonomy. Its authors describe an initial methodological effort with subjective judgments requiring improvement. It helps name what is being observed; it is not an audit of a laboratory’s results.

Consider a thought experiment separate from Anthropic’s data. A person spends a 40-hour week on two tasks: 36 hours on a difficult experiment and four hours on a report. Giving each task half the weight means fully automating the report would raise the automated share to 50%. Even with no supervision, the most time released would be four hours out of 40, or 10%. This does not estimate an actual bias. It demonstrates that the two percentages answer different questions.

The reverse is possible: automating one rare but lengthy task can release substantial time while barely changing a task counter. Weighting everything by hours is not sufficient either. An hour of discussion may prevent weeks of unhelpful experiments. Scientific value requires another observation.

What productivity studies can be compared on

METR’s February 24, 2026 methodological update documents a concrete problem: developers and tasks leave the sample because participants no longer want to work without AI. Its protocol assigned tasks to AI-allowed or AI-disallowed conditions. Selection before random assignment limits generalization.

This work does not evaluate Anthropic’s R&D. It shows why an experiment can also miss part of the phenomenon. The useful comparison concerns methods, rather than competing percentages produced in different settings.

In a survey published on May 11, METR questioned 349 technical workers during February–April 2026. Reported gains distinguish speed from work value, but remain perceptions rather than timed improvements. The survey describes respondents’ experience without observing the counterfactual output they would actually have produced without assistance.

Cunningham and Whitfill’s May 8 note explains that AI can change task selection itself. Measurements on previous activities and on newly chosen activities no longer describe the same basket. That distinction matters when interpreting changes over time.

Calculate the full cycle before claiming savings

Here is a second teaching scenario, not an estimate for a real organization. A team must produce the same experimental dossier to the same acceptance standard. Initially, it requires 100 human hours. The team identifies 30 hours of delegable execution. Assistance removes 24 hours but adds eight hours of supervision and six hours of corrections.

Scenario componentHuman hours
Initial work100
Execution avoided−24
Additional supervision+8
Additional corrections+6
Total after adoption90

Net savings are ten hours, or 10% of initial time. With identical quality and a repeatable cycle, theoretical capacity becomes 100 / 90 = 1.111 times the original capacity, approximately 11.1% more. The 30 delegable hours become neither 30 saved hours nor 30% additional output.

The time break-even point is explicit: supervision and corrections must remain below the 24 execution hours avoided. If they require 26 hours, the dossier takes 102 hours. This threshold is not a recommendation to cut checks. It identifies a use case that has yet to deliver a time saving. Monetary costs would require separate information on compute prices, pay and integration expenses.

To use this calculation, the team must define an accepted dossier before the trial and count abandoned attempts. Selecting only successful runs would bias the result. Human time and elapsed calendar time also need separate columns: several agents working concurrently do not necessarily reduce both measures in the same proportion.

Match the conclusion to the evidence

Anthropic’s figure is an internal delegation signal to interpret in its stated unit. Investment decisions also need accepted outputs, rework and total resource use. None of the studies used here provides an estimate specific to Moroccan teams; the scenario assigns them no expected return.

This framework complements our research commercialization guide. Once a use case is identified, measuring the complete cycle helps determine whether to extend, revise or abandon it.

References