AI Agents
When an AI agent should escalate to a human
The easiest way to make an AI agent look impressive in a demo is to have it answer everything. The easiest way to make it fail in production is the same thing. The Shopify AI support agent I built is designed around the opposite instinct: knowing when not to answer is the actual product.
The default failure mode: an agent that always answers
Most AI agent failures aren't the model being wrong about something obscure — they're the model answering confidently about something it was never actually equipped to know. An agent with no escalation path doesn't fail loudly. It fails by sounding right.
That's the specific risk a support inbox can't absorb. A wrong answer about a customer's order isn't a minor inconvenience, it's a customer who now trusts the business less than before they reached out. The fix isn't a smarter model, it's a workflow that's allowed to say it isn't confident about this one.
What triggers the handoff in a real system
In the Shopify support agent, the escalation decision isn't a vague vibe check — it's a defined step in the n8n workflow, sitting right after the agent drafts a reply. The workflow fetches the customer's actual order first, so the agent is reasoning from real data rather than guessing, and the decision step then checks whether that reply is actually grounded in what it found.
In practice, the situations worth routing to a person are the ones where the model's confidence and its actual grounding come apart: an order that doesn't match what the customer described, a request that falls outside routine order questions, or a reply the workflow can't verify against the data it fetched.
Escalation has to mean something real, not a flag nobody checks
An escalation path only works if what happens next is a person actually seeing the enquiry with context attached — not a flag buried in a log, and not a generic notification stripped of the details that made it hard in the first place.
That's why the escalation step in this workflow forwards the original enquiry along with what the agent already found, rather than just kicking it back to a queue. Whoever picks it up isn't starting from zero — they're picking up exactly where the agent's confidence ran out.
The uncomfortable part: this makes the agent look less capable
Building in a real escalation path means the agent will visibly not answer things sometimes, and that's an easy feature to feel pressure to cut. But an agent that never escalates isn't more capable, it's just less honest about where its limits actually are — and a business only finds out where those limits were after something's gone wrong with a customer.
The actual value of an agent like this isn't that it replaces a person. It's that it removes the repetitive, answerable load from the inbox while keeping a real person in the loop for exactly the cases that need one — a smaller, more honest claim than most AI support pitches make, and the reason this one is safe to actually run in production.