Audit the automations that already exist: for each active flow and intent, how many
tickets it touched in the period, a labeled estimate of time saved, and the noise it
created — then a keep / kill / tune verdict. This skill never disables, edits, or
creates flows/intents.
1. Confirm the period (default: last full quarter). Inventory the active automations —
list the live flows and intents, and for each open it and read what it does:
trigger, actions, and whether it routes, gathers, or resolves. (Reading the
flow/intent inventory is admin-only; if you can't read it on this login, report that
the inventory cannot be read and stop.)
2. TOUCH COUNTS — per automation, search for tickets showing its footprint in
the period (matching category/board/trigger pattern, automation-applied tags or
statuses, its notes). State which footprint signal was used; where attribution is
fuzzy, give a RANGE, not a point count. If you cannot tell whether the automation or
a human did the work, say so and bound the count rather than crediting the
automation. Record caps; never present capped searches as exact touch counts.
3. TIME-SAVED ESTIMATE (labeled) — per automation: touches x minutes saved per touch,
where the per-touch figure comes from what the automation replaces (triage-and-route
~2-4 min; info-gathering ~5-10 min; full resolution ~ the category's median handle
time from a sample). Every such number carries the word "estimate" and shows its
per-touch assumption — NEVER present modeled savings as measured hours.
4. NOISE SIDE OF THE LEDGER — per automation: tickets it routed wrong (bounced after
auto-routing), duplicate/unnecessary notes or messages, reopened auto-resolutions,
tech-visible clutter. Count reopened auto-resolutions against the automation
honestly — a fast wrong answer is negative ROI.
5. VERDICTS — keep (clear positive ledger), tune (works but misroutes/over-fires — say
what to change), kill (near-zero touches, or noise exceeds value). Zero-touch
automations from a past era are kill candidates, not trophies. Verdicts are
recommendations only; changes go through an admin after human review.
6. Output: one table (automation / touches / est. time saved / noise observed / verdict
/ one-line rationale), the top tune recommendations in enough detail to act on,
total estimated hours saved clearly marked as an estimate with assumptions listed,
and a methodology note (period, attribution signals per automation, per-touch
assumptions, searches, caps).
