An escalation policy exists for one scenario: the first person did not respond. Most policies fail it.
The three ways chains break
Every level targets the same person, so a sleeping primary means the chain runs out. Someone acknowledges without acting, stopping the chain for an incident nobody is working on. The channels are all soft — email and chat do not wake people.
A chain that survives all three
0 min on-call email
5 min on-call SMS
10 min on-call phone call
15 min team lead phone call
The first level is quiet, for someone at a desk. The third assumes they are asleep — only a ringing phone reliably wakes people. The fourth assumes they are unreachable, and it is the level most policies omit and the only one that covers the scenario the policy exists for.
On delay timing
Either each delay is measured from when the incident opened, or from the previous level. Absolute timing is easier to reason about. Whichever your tool uses, make sure everyone knows.
How aggressive should the timings be?
Match them to the consequence. For a payment path, compress it. For an internal tool, stretch it or do not page at all. The mistake is one policy for everything.
Who should the backstop be?
Someone who can fix the problem or make a decision, who is genuinely reachable and knows they are on the list. A backstop who discovers their role when their phone rings at 4am is a surprise, not a backstop.
Testing the chain
Once a quarter, fire a deliberate test outside working hours and do not acknowledge it. You will find phone numbers with missing country codes, an SMS provider that drops messages to one carrier, a do-not-disturb setting that blocks unknown numbers. Have your on-call people add the alerting number to their contacts and allow it through do-not-disturb.
The habit that makes it all work
None of this survives a culture where acknowledgement is casual. Acknowledging means you are working on it now. If you have seen it but cannot act, let it escalate. That is what the chain is for.