Editorial clarification r6 · 8 August 2026: the issue now connects its analysis to precise Practicum worksheets. Facts, sources and the overall conclusion remain unchanged.
AI can rapidly transfer some of the practices of stronger colleagues to a novice. The advantage is not universal: once a task moves beyond a known pattern, a confident model answer can conceal an error.
What actually changed
A study of 5,179 customer-support agents found that an AI assistant increased productivity by 14% on average. The gain reached 34% for novice and lower-skilled workers, while the most experienced workers saw little effect. The authors argue that the tool helped disseminate the practices of high performers. Source: Erik Brynjolfsson, Danielle Li and Lindsey Raymond, "Generative AI at Work", NBER.
In this setting, AI did more than accelerate typing. It lowered the cost of access to accumulated work patterns, and changed the starting position most for the novice.
What became cheaper, and what became more valuable
Advice, a first plan and the application of a known pattern became cheaper. Recognising the boundary of that pattern, checking independently and stopping when a task requires new judgment became more valuable.
The issue's conclusion: AI can narrow the execution gap on familiar work while increasing the premium on knowing whether the task is familiar at all, and on owning the outcome.
Three signals to verify
- +14% productivity on average and +34% for less experienced workers came from a field study of 5,179 support agents. It is evidence from one work setting, not an estimate for every profession. Source: Brynjolfsson, Li and Raymond, "Generative AI at Work", NBER. Give a novice AI assistance only on a repeatable task with a reference answer and source review; measure not the first-draft speed but the share of results accepted without rework.
- In an experiment with 758 Boston Consulting Group consultants, participants using AI inside its capability frontier completed 12.2% more tasks, worked 25.1% faster and produced higher-quality answers. Source: Dell'Acqua et al., "Navigating the Jagged Technological Frontier", Organization Science. Make a short playbook for one familiar task type: inputs, examples of a good result, an automated check and escalation conditions. AI can then reinforce the standard instead of guessing it.
- On a task outside that frontier, participants using AI were 19% less likely to produce the correct answer. The study included one such outside-frontier task, so the number demonstrates a risk rather than a universal effect size. Source: Dell'Acqua et al., "Navigating the Jagged Technological Frontier", Organization Science. For work requiring open judgment, name an independent reviewer and a stop condition before starting; do not delegate the decision when the team cannot say how it will recognize a plausible error.
How to read it
The evidence comes from different methods. NBER observes real customer-support work; the BCG experiment compares performance on selected consulting tasks. A survey of 319 knowledge workers adds another signal: higher confidence in GenAI was associated with less critical thinking, while higher self-confidence was associated with more. This is self-reported correlation, not proof of lower performance. Source: Microsoft Research, CHI 2025.
What it changes for people and work
- Novice: use AI as a learning accelerator, but ask for the basis of an answer and check it against a primary source or an accepted process.
- Experienced professional: your value moves toward framing, recognising exceptions and teaching quality criteria, not only performing the task quickly yourself.
- Manager: access to a model does not replace mentorship. Teams need examples of good work, escalation rules and a person who owns the use of the result.
What it changes for teams and business
Organizations can scale work standards as well as AI access. First separate tasks with a proven playbook from tasks requiring open judgment. AI can be a learning and production lever for the first; the cost of independent verification must be designed into the second from the start.
Working map for the week
| Decision | When it fits | First move | Boundary |
|---|---|---|---|
| Give AI to a novice | The task repeats and a reference exists | Provide a reference answer and source review | Do not measure only first-draft speed |
| Delegate a familiar task | A playbook and standard exist | Record inputs, a good result and an automated check | Do not make the model guess the quality bar |
| Keep the decision with a person | Exceptions or open judgment matter | Name an independent reviewer and a stop condition | Do not automate when a plausible error cannot be recognized |
Continue in the Practicum
To keep an advantage for a novice from becoming a confident error, choose one task and make a Volume 1 workbook: name its outcome, "correct" criterion and a check in a month. If the task is already part of a team process, record in the Volume 2 workbook where AI works alone, where review is needed and where a decision is not delegated.
One sensible next move
Take one task your team plans to give AI and place one of two labels next to it: known playbook or open judgment. For the first, name a reference answer and an automated check. For the second, assign an independent reviewer and a stop condition. If the label is unclear, the task is not ready for automation.
Confidence and sources
The NBER field study is strong evidence for customer support, not every kind of work. The BCG experiment makes the jagged capability frontier visible but tests a limited set of tasks. Microsoft Research reports self-assessed behavior. The rising value of verification is the Practicum's editorial synthesis, not a single estimate from a single study.