Whether Your Customer Can Reach a Human Is a Setting

There is a genre of article about saying "agent" or "representative" until a bot gives up, and pressing the hash key to confuse the menu. It is written for people on the outside of the system, and it is entirely fair.

This is the view from the other side. Whether those tricks work is not a property of the AI. It is a setting, in a panel, that somebody at the company you are emailing either configured or did not.

We read four platforms' escalation documentation. Two escalate the moment a customer asks for a human, with no configuration at all. One documents a rich set of handoff rules and does not state what happens if you write none. One is explicit that escalation must be built before launch — and does not say what happens if it never is.

The three questions that matter

Not "can it escalate", because they all can. The useful questions are narrower:

• Does asking for a human work without anyone configuring it?
• Is there a floor the vendor will not let you switch off?
• Is the customer told they have been handed over, or does it happen silently?

This is a reading of published documentation, not a test. We did not deploy four AI agents and try to escape them. Where a vendor's docs do not address something, the row says so rather than guessing — and on this topic the silences are as informative as the settings.

Gorgias publishes a floor, and it is the most interesting thing in this study

Gorgias documents a set of behaviours that cannot be turned off no matter how the AI agent is configured. It always hands over when it detects anger or frustration, when it receives an explicit request for a human, and when it identifies a sensitive topic.

That last category is worth quoting properly, because it is doing real work: mentions of self-harm, threats of violence, threats of legal action, and requests involving financial account details. Those cannot be disabled by a customer of Gorgias, which means a support team cannot accidentally configure a bot that keeps talking to somebody in crisis.

No other platform in this group publishes an equivalent floor. That does not prove the others lack one — it means it is not in the documentation, so a buyer cannot rely on it.

The silent handover

Gorgias also has a setting called "Tell customers when handing over". Switched on, the AI sends a message saying they are being connected to a human. Switched off, in the documentation's own words, the AI "hands over the ticket without notifying the customer".

There are defensible reasons to want that. A silent handover avoids a jarring seam, and the customer arguably does not need a status report on your internal routing. But the person on the other end has been trying to establish whether they are talking to a machine, which is the entire subject of the articles ranking for this query, and the answer has quietly changed without being announced.

It is a small toggle with a large implication, and it is worth someone senior deciding rather than whoever set the agent up.

Fin escalates by default, and there is no button

Fin escalates out of the box in three situations: the customer clearly asks to speak to a human, Fin detects strong frustration or anger, or the customer is stuck in a repetitive loop. That is a good default and it means the magic word genuinely works.

What surprises people is the interface. There is no "Talk to a person" button in Fin over chat — standard quick-reply buttons are not shown, and the handoff is offered conversationally in the text of a reply instead. If you were expecting a visible escape hatch in the widget, there is not one, and you cannot add one.

Whether that is better is genuinely arguable. A button is discoverable and unmistakable. A conversational offer is less intrusive and does not invite everybody to skip the AI immediately. It is a design position rather than an oversight — but it is not the design most buyers assume they are getting.

Zendesk makes it your job, and is honest about that

Zendesk's documentation is unambiguous: escalation is not automatic. You develop an escalation strategy, design the flow, and build the paths using escalation blocks in the dialogue builder, before launching the agent.

It goes further than most in one respect — there is a fallback field for when the escalation itself fails, so you can specify what the customer is told if the handoff cannot complete. Somebody thought carefully about the unhappy path.

What the documentation does not say is what happens if you build no escalation at all. That is the gap. A team under deadline, launching an agent on a Friday, has been told to design a strategy and has not been told what the consequence is of skipping it.

What to check before you launch, and what to check afterwards

• Type "I want to speak to a human" into your own bot, on your live site, today. It is a thirty-second test and it is the only one that matters.
• Then try it angrily, and try it after two failed answers. Frustration and loop detection are separate triggers from an explicit request, and they fail separately.
• Check whether the handover is announced. If your platform can transfer silently, find out which way yours is set.
• Check the out-of-hours path. Several platforms have separate online and offline handoff rules, and the offline one is the one nobody tests.
• Ask the vendor what their unremovable floor is. If the answer is that there isn't one, that is worth knowing before a sensitive conversation reaches your agent.

The tricks articles exist because these paths are often broken. The fastest way to keep your company out of that genre is to spend half an hour being your own difficult customer.

WRITTEN BY AR · UPDATED 2026-08-06

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

How we work · About the author