The useful question is not whether documentation can be automated, but how much. Claiming all of it invites distrust; admitting none of it ignores what diff-aware tooling can now do reliably. The answer sits in the middle, and it is measurable.
We looked at a large sample of real merged changes to find the line between what a tool can draft and what still needs a person.
Direct answer
A large share, but not all. In our audit of 29,917 merged pull requests, 35.2% of customer-facing changes were structured enough to draft documentation from the diff alone. The automatable part is the mechanical majority: API changes, config, CLI flags, endpoints. The remainder, pricing decisions, net-new visual workflows, and judgment calls, is where humans still add the value. Tools like DocsKoala automate the first and route everything through human approval.
The automatable share, measured
Across the dataset, a consistent majority of customer-facing changes were mechanically documentable from the pull request.
- 35.2% of customer-facing changes were auto-draftable from the diff.
- Among changes that never got documented, 32.2% were automatable, meaning most of that debt was avoidable.
Full method and figures are in the State of Documentation Debt report.
What automates cleanly, and what does not
| Automates cleanly | Needs a human |
|---|---|
| API endpoints, parameters, responses | Pricing and plan changes |
| CLI flags and config options | Net-new visual workflows |
| Changelog and release-note drafts | Positioning and voice |
| Detecting which article a change affects | Final approval and edge-case judgment |
The realistic model: draft, then approve
The productive setup is not full automation. It is automating drafting and detection, then keeping a human in the loop for approval. That captures the majority of the work while preserving accuracy and voice.
Frequently asked questions
Can 100% of documentation be automated?
No, and it should not be. Structured changes automate well, but pricing decisions, net-new visual workflows, and voice need human judgment. In our audit, 35.2% of customer-facing changes were auto-draftable from the diff.
What documentation is hardest to automate?
Anything that depends on context a diff does not contain: business and policy decisions, brand voice, and net-new workflows that require product judgment or a fresh screenshot the code change does not describe.
Is partial automation actually useful?
Very. Automating the drafting and detection for the structured majority removes most of the recurring labor, while human approval keeps the output accurate, which is the combination that keeps docs current without a dedicated team.
Conclusion
- A clear majority of customer-facing changes are automatable from the diff.
- The remainder needs human judgment, which is why approval stays in the loop.
- The winning model is draft-then-approve, not full automation.
- DocsKoala automates the drafting and detection, you approve.