The wrongly framed question: "AI or humans?"
Framed that way, the question pushes you to pick a side, and no serious support business runs on a single extreme. The useful question is different: for this specific conversation, at this specific point, who is more likely to resolve it well — the automatic assistant or a person? That decision can be made with explicit rules instead of leaving it to case-by-case improvisation.
What an AI assistant handles well
Questions that repeat, that have a clear answer in the knowledge base, and that don't require judgment about a customer's particular situation are natural territory for automation: hours, policies, how to use a product feature, order status. There, the assistant isn't just faster than a human — it's more consistent.
The signs that it's time to escalate
There are fairly reliable signals: the customer is asking for an exception to a policy, the conversation carries emotional weight, the question requires information that isn't in any knowledge base article, or the customer explicitly asks to talk to someone. A good system recognizes these before trying to force an automatic answer that's probably going to go wrong.
Why escalation has to be explicit, not a silent collapse
The worst version of "when to escalate" is when there's no real decision at all: the assistant keeps trying to answer with partial information until the customer gets fed up and writes "I want to talk to a person" out of frustration. Explicit escalation actively recognizes that it's reached the limit of what it can confidently resolve and communicates that upfront, instead of pretending it can keep helping.
The role of the critic pass in this decision
A fidelity-check mechanism that audits whether the answer the assistant is about to send is actually backed by the knowledge base ends up being, in practice, the most reliable trigger for deciding when to escalate: if there isn't enough support to answer confidently, that's the signal that a human needs to take the case.
Automating more isn't always the goal
It's tempting to measure an assistant's success by the percentage of conversations it resolves without human intervention, and that number does matter — but chasing it without limit leads to forcing automation where it doesn't belong. A good system doesn't maximize the percentage of automated cases — it maximizes the percentage of cases resolved well, whether the AI or a person resolves them.