# Evidence note: AI-assisted independent knowledge work

Last checked: **18 August 2026**. [Version française](independent-knowledge-work-cases.fr.md).

This note explains the evidence behind the
[synthetic independent-work case](../examples/en/independent-client-follow-up.md).
The people, clients, volumes, costs, and pilot results in that case are fictional.
They are demonstration values, not an additional field study.

## Useful anchors

- [Dell'Acqua et al.](https://doi.org/10.1287/orsc.2025.21838) studied 758
  consultants. For tasks inside the model's capability boundary, participants
  completed 12.2% more tasks and worked 25.1% faster on average. On a complex
  task outside that boundary, they were 19% less likely to be correct.
- [Noy and Zhang](https://doi.org/10.1126/science.adh2586) found faster work and
  higher average quality on controlled professional-writing tasks. The study
  did not measure rare errors or a continuing client relationship.
- The [OECD SME report](https://www.oecd.org/en/publications/generative-ai-and-the-sme-workforce_2d08b99d-en/full-report/component-4.html)
  shows that simple, occasional, and peripheral uses remain more common than
  core-business uses. Its survey data describe use, not causal productivity.
- [Dillon et al.](https://www.nber.org/papers/w33795) found roughly two fewer
  email hours per week among active users later in a six-month experiment, but
  no aggregate change in task quantity or composition.

These studies can inform a starting range when the task, expected finish,
human experience, and review process are close enough. They do not predict an
independent professional's income or guarantee a time saving.

The worked case therefore keeps pricing, commitments, recommendations, and
final decisions with the human. Its 23% saving is a labelled synthetic result
for a copilot task, not a lower bound for a business agent or an orchestrated
agency. See the [work-mode evidence note](agentic-integration-levels.md) for
those distinctions.