SpaceX Starlink support deployed Grok Bot as persistent job owners instead of reactive ticket handlers. Two core patterns emerged:
IncidentSweep: Engineer defines criteria + exact reply text → bot filters 204 matching tickets in 10 minutes → human approves batch → sends in controlled waves. Previously took hours of manual replies. Bot never adds tickets post-approval, halts on any failure, logs everything.
IncidentWatch: Runs every 5 minutes, scans last 15 minutes for ticket clusters around same root cause. When spike detected without known incident, alerts on-call with issue summary + count + examples, holds matching tickets. Once confirmed, hands off to Support Bot for status updates.
Economics: After ops consolidation, ticket volume +175% with flat headcount. Resolution cost dropped to $0.20–$0.30 per ticket vs $1–$4 industry standard for AI support tools (runs inside existing plan, no per-usage fees).
Architecture pattern: Bot maintains state across turns, runs on schedule/event triggers, operates in team's existing toolchain. Human stays in loop for commitment decisions, bot owns continuous monitoring + execution within defined bounds. Stops when criteria unclear.
Same pattern applies to:
• CRM scanning for stalled deals → draft follow-ups → batch approval
• CI failure clustering → auto-reproduce in test env → file tickets only on confirmed repro
• Invoice matching → flag PO mismatches → stage payment batch for single approval
Key constraint: Quality of criteria and message templates supplied by team. Bot executes within those bounds or halts. Teams typically expand job set rapidly once first continuous job runs cleanly.
This is job-oriented AI vs ticket-oriented AI. The bot is a teammate with standing responsibilities, not a per-query assistant.
IncidentSweep: Engineer defines criteria + exact reply text → bot filters 204 matching tickets in 10 minutes → human approves batch → sends in controlled waves. Previously took hours of manual replies. Bot never adds tickets post-approval, halts on any failure, logs everything.
IncidentWatch: Runs every 5 minutes, scans last 15 minutes for ticket clusters around same root cause. When spike detected without known incident, alerts on-call with issue summary + count + examples, holds matching tickets. Once confirmed, hands off to Support Bot for status updates.
Economics: After ops consolidation, ticket volume +175% with flat headcount. Resolution cost dropped to $0.20–$0.30 per ticket vs $1–$4 industry standard for AI support tools (runs inside existing plan, no per-usage fees).
Architecture pattern: Bot maintains state across turns, runs on schedule/event triggers, operates in team's existing toolchain. Human stays in loop for commitment decisions, bot owns continuous monitoring + execution within defined bounds. Stops when criteria unclear.
Same pattern applies to:
• CRM scanning for stalled deals → draft follow-ups → batch approval
• CI failure clustering → auto-reproduce in test env → file tickets only on confirmed repro
• Invoice matching → flag PO mismatches → stage payment batch for single approval
Key constraint: Quality of criteria and message templates supplied by team. Bot executes within those bounds or halts. Teams typically expand job set rapidly once first continuous job runs cleanly.
This is job-oriented AI vs ticket-oriented AI. The bot is a teammate with standing responsibilities, not a per-query assistant.