A useful automation candidate has an observable before-state, a repeatable rule and a clear owner. A visible feature such as a chatbot does not establish those conditions by itself.
Follow one real enquiry through the company from arrival to next action, using timestamps from the systems where the events occurred.
The audit: follow one enquiry
Choose a representative customer enquiry. Write down each station it passed: arrival, first review, response, system record and follow-up. Use real timestamps from the inbox and CRM rather than estimates.
Then identify observable gaps: missing owner, missing next action, repeated re-entry, or a wait between two recorded events. These become candidates for a bounded test.
Score the candidates on three axes
Volume: how often does this step run in the measurement window? Use a real count.
Wait time: how long does work sit before this step happens, using defined start and end events?
Structure: can the team describe the rule and its exceptions? If so, automation can prepare or execute the agreed step; pricing and negotiation remain human decisions unless a separate deterministic rule is approved.
When intake is the right first boundary
Intake is a suitable first boundary when source messages can be reconciled against destination records, the required fields are known and ownership currently becomes unclear between those points.
Once that boundary is reliable, follow-up reminders, document preparation or additional routing can be evaluated as separate scopes.
Turn the audit into a bounded first project
Choose a process with repeated input, observable state and a clear owner. Avoid starting with a process whose rules change by person or whose outcome depends on unrecorded judgement.
Define the before-state and acceptance test in operational terms: which events are missed, which fields are retyped, where ownership becomes unclear and what a successful record should contain.
Build the narrowest useful path and keep an escape hatch. The first release should make work more visible even when the automation declines to act.
First-project checklist
- The trigger is observable
- The output schema is agreed
- The owner is named
- Exceptions have a destination
- Success can be checked without vanity metrics