Skip to content
digitice.io
AI-first How we work Cases AI Radar Team
DEEN
Book a conversation ↗
digitice.io / AI Radar / All studies

AI RADAR · STUDY · 21 Oct 2026

People and organisation

At Anthropic, more Claude sessions focus on instructions and oversight

This includes plans for AI agents, reviews of their results and monitoring dashboards. Anthropic examines what employees use Claude for.

This includes plans for AI agents, reviews of their results and monitoring dashboards. Anthropic examines what employees use Claude for.

Source

AI also creates instruction and oversight work

Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory · The Anthropic Institute · 16 Oct 2026

Source finding

At Anthropic, the average employee’s share of interactive Claude sessions for instructions, documentation and oversight rose from 10% in autumn 2025 to 22% in spring 2026. This includes plans for AI agents, reviews of their results and monitoring dashboards.

Our interpretation

AI can create additional preparation and oversight work. Businesses should make these tasks visible. The finding describes sessions at the vendor itself, not a share of working time or a general productivity gain.

Method and evidence

Observational analysis of internal employment, usage and work records plus 17 interviews. For the session-content analysis, 2,000 short session summaries were classified automatically. This analysis is distinct from the separate engineering-team and hiring analyses.

Limitations and interests

Anthropic studies its own rapidly growing company and sells the AI under study. No random assignment, no independent replication, limited generalisability. Automated classification can be wrong. Tokens and estimated reading time do not measure actual human review time.

What businesses can do

Track time spent on instructions, documentation and oversight separately from execution. Assign responsibility for these tasks and specify which results need review.

Where to start

Start with the section on work tasks and Figure 7 in the original report, then the session-classification methods. Published on 16 October 2026.

Open original study ↗

← Previous studyFew respondents report rules for stopping AI initiatives
Next study →Using AI does not yet mean embedding it in daily work

More from this area: People and organisation

  • 4 Nov 2026Mixed learning groups broaden teachers’ AI use
  • 3 Nov 2026AI narrows the performance gap between education groups
  • 2 Nov 2026AI improves performance on more demanding IT tasks

All studies in the AI Radar → · ← Back to the AI Radar

digitice.io

Building AI-first companies.

FAQImprintPrivacy
© 2026 digitice GbR