Four organizations,
one incident desk.

Amazon describes this organisation in its own public job postings as “over 17K Amazonians … spanning 56 global locations, and handling over 50M contacts a year.” It protects the marketplace: account health, payment risk, policy enforcement, brand protection, catalogue integrity. When its tooling breaks, the checks that keep buyers and sellers safe break with it.

One part of that organisation already ran a command center. I had spent the previous decade inside it. The command center's readiness work for Prime Day and the 2021 peak shopping event, which I helped run, is where the gap showed: outside the command center's home organisation, most operations teams had no on-call schedule, no escalation path, no distributed coverage and no worldwide notification mechanism. The teams that took the shared service had mostly been carrying high-severity events on improvisation and goodwill.

Improvisation is expensive in the first ten minutes of an incident, and the first ten minutes are most of the damage.

The teams could not each hire an incident manager, and a central desk could not take uptime accountability away from teams that already owned it. So the service had to be shared, thin and voluntary: six incident managers covering every participating team around the clock, run as a six-month pilot at no cost to the participating teams.

That constraint decided the design. A desk with no authority and no spare capacity cannot be the place where incidents are debated. It has to be the place where an incident is classified, routed and announced: fast, and identically every time. Everything below follows from that one sentence.

It also decided the staffing. The plan began as temporary junior contracts, filled from inside the operations production population. We changed it to permanent program-manager roles, because those contracts could not reliably attract the skill set, judgement and motivation the job needed, and could not be offered on anything but an hourly basis. The reversal went into the plan as a two-way door: if the pilot stopped, the staff would be absorbed back into the parent team. Hiring hubs were placed next to existing command center counterparts in each of the three regions, so new hires could learn by sitting near people who already did the work.

Five mechanisms carried the service. I led the design work; where the command center I came from already had a mechanism, we adapted it rather than rebuilt it. The test for each was the same: could someone on shift at 03:00 in another region execute it without calling anyone?

Two numbers were measured on every escalation: first correspondence (how long until a human acknowledged it) and time to resolve (how long until it was closed or handed on with next steps). Both were reported weekly as the percentage inside the service level, against a 95% goal and a 90% floor.

The month-over-month view is the one I'd put in front of an operator. December 2022: 75% first correspondence and 73% time to resolve, on 71 alerts. January 2023: 88% and 89%, on 283 alerts. Four times the load, and both numbers up: thirteen points on first correspondence, sixteen on time to resolve. That is the signature of a mechanism working rather than a team trying harder.

The Americas and Europe were already close to the line. the third region was at 36%, a new team, a new process and the largest share of the growth in volume. It was not a discipline problem; it was an onboarding problem. It reached 68% in January and 100% by week 5 of 2023, and worldwide attainment moved with it. The adoption work that followed through February and March (the roadshows, the escalation-process reviews, the first drill) was aimed at the same gap.

Across 2023 and 2024, every month, every missed target was reviewed ticket by ticket and given a verdict: a genuine miss, a one-time exception, a mis-attributed resolver, or an allowance for a large-scale event or a still-ramping new hire. Roughly 430 ticket-level decisions across 2023 and 2024. It is unglamorous work, and it is the only reason the percentages above mean anything: the team could not quietly redefine a miss, and a real miss produced a change rather than an argument.

The same discipline shows up in how the wider organisation ran its peak trading events. For Black Friday and Cyber Monday 2022 (deals running in 20 countries), the control center held a round-the-clock conference bridge with named call leaders from the Friday through the following Wednesday, a daily roll call, hourly red/amber/green status pulled from each operations team at quarter past the hour, one master incident ticket carrying every open issue, and a 30-minute update cadence on anything live, against a fixed status template: issue, customer impact, impacted teams, actions taken, expected resolution, available mitigation levers. Missed check-ins escalated up a named ladder to vice-president level. I was one of six named escalation contacts for the whole control center in the run-up to the event, and the single point of contact for my team in the 41-question readiness assessment (capacity, hiring, operations, readiness, resilience, risk) in both 2022 and 2024.