OpenAI report reveals exact workplace skills AI cannot automate — Check details

The report showed that AI is barely utilised for tasks requiring contextual understanding, risk assessment, and complex judgment.

Artificial intelligence is transforming corporate jobs, but a new internal report from OpenAI highlighted exactly which skills remain safe from automation.

According to a Business Insider report, there is a clear divide in how OpenAI’s researchers use AI tools — AI excels at execution and measurable tasks but struggles with strategic decision-making and judgment.

Tasks AI is automating

According to OpenAI's internal usage data, the biggest automation gains are in routine, execution-focused tasks with clear success metrics, such as coding or bug fixing.

• Coding and infrastructure: Researchers used AI the most for writing code, hitting 198,200 tokens per researcher per day.
• Technical review: Assistance with technical help and review accounted for 158,800 tokens.
• Debugging: Launching, monitoring, and debugging computer runs hit 133,100 tokens daily.
• Data analysis: Analysis of experimental results showed substantial AI involvement at 40,200 tokens.

Skills AI cannot replace

The report showed that AI is barely utilised for tasks requiring contextual understanding, risk assessment, and complex judgment.

• Planning and strategy: Planning research and experiments generated just 5,100 tokens per day.
• Project focus: Deciding which projects to prioritise accounted for only 2,300 tokens.
• Resource allocation: Computing and staffing decisions recorded a mere 1,500 tokens.
• Critical choices: The decision to continue or halt a project registered the lowest AI usage at just 200 tokens.

“People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems,” the ChatGPT maker noted in its findings.

Recent updates to ChatGPT, Claude, and Gemini

Anthropic's Claude: Anthropic released Claude Opus 5.5 on 22 September 2026. This updated model is approximately 40 per cent cheaper to run than its precursor.