Key answer
Place checkpoints where checking now can change later investment. State what continues after a pass and what actually requires a user decision. Progress reports, automatic checks and approval requests are separate.
Full explanation
After data import, column, unit and missing-value checks can run automatically before analysis. Two equally reasonable sample styles may need a choice. Authorized research, routine editing and verification should not stop after each item to ask whether to continue.
Microsoft's human-AI guidelines recommend timing interventions to the task/environment and showing contextually relevant information. They were evaluated in multiple rounds, including 49 design practitioners testing applicability to 20 AI products. Study; guidelines 3 and 4
A useful checkpoint identifies the judgment, inspectable evidence and ensuing action. “Please review” is vague; “choose beginners or practitioners because terminology density and case depth depend on it” supports a decision.
Verify data yourself and continue after a pass. Pause affected work only when conflicting evidence changes core conclusions or unresolved goals require a tradeoff; explain differences and recommend a choice. Continue unaffected work and use brief progress records elsewhere.
Refine checkpoints over time. Keep those finding consequential downstream problems; merge or automate those receiving only “continue.” Final artifacts meeting requirements matter, not the number of confirmed checkpoints.
References
1. Amershi et al. (2019), Guidelines for Human-AI Interaction, CHI; guideline validation
2. Microsoft Research (2019), Guidelines for human-AI interaction design, items 3 and 4