80%. That is how far an agentic data analysis system built on Claude Code can get toward doing the job of a human data analyst, according to a practitioner who spent time building exactly that.
The project, documented by a former data analyst, uses Claude Code as the backbone of an automated analysis pipeline. Feed it a dataset and a question, and the system can clean data, run queries, generate visualizations, and produce web-based reports. The practical findings are more interesting than the concept.
What Works and What Does Not
The system performs well on structured, repeatable tasks. Generating reports works significantly better when the agent has explicit templates and pre-defined components to work with. Give it a clear output format and it will fill it reliably. This matches what most people using AI coding tools have found: agents are strongest when the problem is well-constrained.
The 20% gap is where things get interesting. The system struggles with hypothesis generation, the part of data analysis where a human looks at a dataset and asks "I wonder if X is driving Y." That kind of data intuition, knowing which questions to ask before you have the answers, remains firmly in human territory. The model can execute an analysis plan, but it is not great at creating one from scratch.
This is a useful benchmark for anyone evaluating AI agents for knowledge work. The pattern holds across domains: AI handles execution well but falls short on the strategic and creative framing that makes execution valuable. A data analyst who uses this system to skip the mechanical 80% and focus on the hypothesis-driven 20% is probably more productive. An organization that tries to replace the analyst entirely will get polished reports that answer the wrong questions.
For teams already using Claude Code for development work, extending it into data analysis workflows is a natural next step. Just do not expect it to know which questions matter.