Human-in-the-loop automation: where AI should draft and people should decide
“Fully autonomous” sounds impressive but is rarely what a business needs. A practical framework for deciding which steps to automate, assist or keep human.
Modern AI models are excellent at drafting, summarising, classifying and extracting. They are also capable of being confidently wrong. Good automation design accepts both facts and puts people exactly where their judgement matters most.
Three levels of automation
- Automate: low-impact, easily reversible, well-defined tasks — copying data between tools, sending an acknowledgement, creating a reminder.
- Assist: AI drafts or recommends, a person reviews and decides — customer replies, proposals, reports, summaries.
- Human: people decide, possibly with AI research support — pricing exceptions, complaints, legal, financial, health-related and hiring decisions.
Four questions to place a step
- If this step is wrong, who is affected and how badly?
- Can the mistake be noticed and reversed easily?
- Does the step require context or empathy the system cannot see?
- Would the customer expect a person to be involved?
If any answer points to serious or irreversible harm, the step belongs in the Assist or Human level.
Design the review step to be fast
Human review only works if it is easy. Show the draft next to the source information, highlight uncertainty, make editing one click away, and log who approved what. A good review step takes seconds, not minutes.
Log everything that matters
Every automated action with an external effect should be logged with what happened, why and when. Logs turn mysterious failures into fixable bugs, and they build trust with customers and regulators.
Automation with clarity, control and accountability — that's the standard we hold ourselves to.