OpenAI’s own data shows AI automates execution, not decisions

The company tracked how its own researchers use AI tools, comparing activity in early August against late January.The largest jumps in AI-assisted output came in research and infrastructure code, at 198,200 tokens per researcher per day, followed by technical help and review at 158,800 and launching, monitoring, and debugging runs at 133,100.Analysing experiment results also saw substantial automation, at 40,200 tokens per day.

These are tasks with clear, measurable outputs: code either runs or it doesn't, and a bug is either fixed or it isn't.Decision-orientated activities paint a different picture.

Deciding on what to focus on only garnered 2,300 tokens per day for each researcher, computing and staffing decisions 1,500 tokens, and decisions about continuing or stopping work a lowly 200 tokens.Research and experiments planning only hit 5,100 tokens, much lower than the activity-orientated categories.

The trend shows that AI is much better at doing than at deciding.However, coding and monitoring lend themselves to an objective measure, which makes them easier targets to be automated by the same tools.

Prioritising what a team needs to create requires judgement and prioritisation, which is not easily measurable.The description of OpenAI reflects this problem as well: humans still decide about priorities in research, assess whether projects are worthy to pursue, and decide if a system is to be scaled up or down.The more executional tasks get automated, the less likely it is that the most valuable skills will be related to judgement,….

OpenAI's own data shows AI automates execution, not decisions September 26, 2026.