The useful question is rarely 'can AI do this?' — current assistants can attempt almost anything. The useful question is 'what happens if it is wrong, and who would notice?' This guide gives you a simple way to decide, task by task.
Three outcomes, not two
Sorting tasks into 'automate' and 'don't' leads to bad calls. A third outcome — document first — covers most real work: the process exists only in someone's head, so neither a new hire nor an AI tool could follow it yet.
- AI drafts, a person approves: frequent, rule-based work with a quick human check (first-draft replies, meeting summaries, report commentary).
- Document first: work that is repeatable but undocumented. Write the procedure; then revisit.
- Keep human-led: judgement-heavy work, or anything regulated or client-critical where an error is costly (pricing exceptions, HR decisions, legal or tax positions).
Four questions to ask of each task
First, how often does it happen? Rare tasks rarely repay setup time. Second, could someone do it from written instructions? If not, document first. Third, what does an error cost, and would anyone catch it before a client does? Fourth, what data does it touch? Client, staff, financial or health information raises the bar for tools, terms and review.
Work through these in order. A 'no' on the second question means document first, however attractive the task looks.
Design the review step on purpose
'Human in the loop' only works if the loop is defined: who reviews, what they check, and what happens when the output is wrong. Write a short checklist for the reviewer and a stop condition for the pilot, such as an error rate above a level you agree in advance.
Client-facing outputs should be approved before sending, every time, during any pilot. Internal outputs can move to spot checks once the error rate is low and stable.
Written and reviewed by The AI Scale Academy team. Guidance is general and educational, not legal, tax or financial advice.