Finding automation candidates

With the map annotated, 'what should we automate?' stops being a brainstorm and becomes a scoring exercise. Three questions per step, and they do almost all the work:

• Frequency — how often does the step run? Automating a 50×/day step pays daily; automating a monthly step rarely repays its build. (Your map's volume annotations answer this — the data habit paying off.)
• Rules vs. judgment — could you write the decision as instructions a new hire could follow verbatim? Rules automate cleanly. Judgment automates badly and degrades silently — the failure isn't a crash, it's a plausible-looking wrong decision at scale.
• Cost of error — a mis-tagged ticket costs a re-tag; a wrongly-issued $400 refund costs $400 and an awkward email; a wrongly-denied refund costs a customer. Error cost sets the supervision level, not whether AI is 'good enough.'

The three seats AI can sit in

• Assist — AI prepares, human does: summarize the ticket thread and surface the order history before the agent reads it. Low risk, immediate value, the right default for judgment steps.
• Draft — AI produces the artifact, human approves: the refund-decision recommendation with reasons, the customer email, the weekly report. The human is a reviewer, which is faster than being an author. Most BA-adjacent wins live here.
• Decide — AI acts, humans audit samples: auto-tag refund reasons, auto-route tickets, auto-approve refunds under $25 on undamaged-return receipt. Only for high-frequency + rules-based + low-error-cost steps, always with an audit trail and an escape hatch to a human.

Harbor Lane's returns map scores itself: refund-reason tagging (high frequency, pure rules, trivial error cost — decide, and it would have surfaced the damage spike months earlier); refund decisions (judgment + real error cost — draft, agent approves); the fragile-flag review (low frequency but the root cause — not an automation candidate at all, an ownership fix with a quarterly checklist). That last one is the lesson: the scoring rubric's most valuable output is the steps it tells you not to automate.