The Delegation Boundary: Automate What You Can Undo

The article discusses the importance of clearly defining the boundary between automated and human responsibilities in AI-assisted systems, particularly in the context of engineering and business operations. The primary argument is that irreversibility should dictate which tasks are automated and which must remain under human control.

The author cites several case studies, including the EBU/BBC study showing 45% of AI assistant answers about news contain distortions, and two specific public failures involving CNET and Bloomberg. In both instances, the errors reached the public because human approval steps were removed, leading to significant cleanup efforts and damage to the company's reputation.

The author proposes a three-bucket matrix for engineering teams to categorize tasks based on the cost of being wrong. Bucket one includes reversible tasks such as writing triage reports, alert digests, test generation, linting, formatting, drafting pull requests in isolated branches, documentation drafts, and dependency scans. Bucket two consists of automated checks like CI gates, test passes, and output schema validity, which can be reviewed and approved by humans before merge.

Bucket three contains irreversible steps like granting production access and credentials, spending money, deleting or migrating data, sending messages to customers, issuing public statements, and making final decisions during incidents.

The author emphasizes that the delegation boundary should be based on reversibility rather than the nature of the task (routine vs. creative). He argues that the standard advice of "automate the routine, keep the creative work human" fails in engineering reality because dependencies and data migrations can be routine but still irreversibly impact production.

The author also notes that the market has already decided that augmentation (humans working together with AI) is preferred over automation. The Anthropic Economic Index shows 57% of Claude conversation usage is augmentation versus 43% automation. Surveys indicate that audiences demand human oversight in AI-generated content, with only 12% comfortable with AI-generated news and 43% accepting a human-led format.

In summary, the article stresses that engineering teams must establish clear responsibility boundaries between automated and human-controlled tasks, with irreversible steps remaining under human control. This boundary should be enforced within the workflow rather than relying solely on prompt engineering, as it is crucial for maintaining trust and avoiding irreversible damage in AI-assisted systems.