Search the Practicum

Type two letters · ↑↓ select · Enter open · Esc close

Type at least two letters.

Where usefulness is moving: from doing to checking

Live Practicum page for Chapter 5 (called Levels Watch in the book). Verified: 2026-08-18. Studies on how people work with AI come out every few months and often disagree. Check the date and the type of data. Do not take our word for it. Check.

Chapter 5 argues that value moves up a ladder: from "I did it myself" to "I checked it," then to "I built the process," "I owned the outcome," and finally "I built a system that runs without me." Here is the evidence that ladder rests on, with the caveat attached to each piece.

What the research shows

Source Date What it shows Caveat
Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (HBS WP 24-013; Organization Science) 2023, journal version 2025 758 BCG consultants. Inside the jagged frontier, AI produced 12.2% more completed tasks, 25.1% more speed, and quality more than 40% higher. Outside it, the AI group was 19 percentage points more likely to land on the wrong answer a field experiment with consultants on a limited task set; the frontier rarely sits where people expect
Randazzo et al., "Cyborgs, Centaurs and Self-Automators" (HBS WP 26-036) 2026 Three ways of working with AI: cyborgs, about 60% (constant back-and-forth), centaurs, about 14% (split the task, then verify), and self-automators, about 27% (hand it over whole). Centaurs produced the most accurate recommendations. Self-automators built neither domain skill nor AI skill the shares are rounded, and retellings add up to 101%; the sample is specific
Yang et al., "How AI Agents Reshape Knowledge Work" (arXiv; Perplexity data) June 2026 An agent works on its own for 26 minutes, against 33 seconds for ordinary AI search. A task that took 269 minutes fits into 36 (87% less time, 94% less cost). Half of agent tasks reach the "create something new" level, against 26% for search data from a single platform, not the whole labor market; this is product usage data, not an experiment
OpenAI GDPval 2025 Models deliver expert-level results on roughly half the tasks in a set of real work deliverables a task benchmark, not an employment forecast. The chapter's working question is "which half?", and the benchmark does not answer it
Microsoft Work Trend Index 2025 The agent boss role: a person who "builds, delegates to, and manages agents." 41% of leaders expect their teams to be training agents within five years a vendor survey with a stake in the topic; these are expectations, not hiring data
Anthropic Economic Index, "Cadences" report June 26, 2026 Augmentation overtook automation. Just over half of use is now collaborative rather than delegated. Two findings map straight onto the ladder: experienced users automate successfully far more often than newcomers, and workers with 15+ years of experience judge that AI could handle about 10 percentage points fewer of their tasks than early-career workers judge of theirs usage and survey data from one vendor's own product, and its users are not a cross-section of the workforce: computer and mathematical roles were about 30% of respondents against 4% of US employment
Lee et al., "The Impact of Generative AI on Critical Thinking" (Microsoft Research with CMU, CHI 2025) 2025 319 knowledge workers and 936 task examples: the more people trusted the AI, the less critical effort they spent. Thinking itself shifts toward checking and oversight a survey with self-reported data. Chapter 8 covers this risk in depth

How to read this

The skill gap is now measurable, and it runs the way the ladder predicts

The most useful 2026 finding for this chapter is that experience with the tool, not access to it, decides what you get out of it. Anthropic's own usage data shows experienced users succeeding at automation far more often than newcomers. Access stopped being the differentiator some time ago; what separates people now is knowing what to hand over and how to check it, which is exactly the move from level 1 to level 2.

There is a second, quieter finding worth sitting with. Long-tenured workers think AI could take over less of their work than early-career workers think of theirs. Read it carefully, because it has two possible readings and the page cannot settle which is right. Experienced people may see the judgment inside their own work that a newcomer cannot see yet. Or they may be underrating a tool they use less. Both readings are live, and the honest position is to hold them both.

"Faster" and "better" split apart at the frontier

The HBS/BCG experiment makes the chapter's practical point: the same tool helps a lot inside its zone and confidently misleads you outside it. Value comes from knowing where that line runs, not from having access to the model. That is what people get paid for at level 2 and above.

How you work matters more than what you use

Only about one in seven people works as a centaur, yet centaurs give the most accurate recommendations. Self-automators outnumber them nearly four to one and build no new skill at all. The difference is not the model or the subscription. It is whether the person keeps the decision.

The agent numbers come from one platform

The Yang data is impressive and honestly collected, but it is usage data from a single product. Take the direction: agents pull work upward, toward creating something new and past the edges of one profession. Do not take the percentages as yours.

What to do with this

Sources

Version: 2026-08-18.

Where usefulness is moving: from doing to checking