How to Get a Human Past AI Customer Service Agents
43.9% of Americans have yelled 'human' at an AI service agent. Here are the phrases, menu paths, and channels that actually reach a person fast.


You reach a human fastest by refusing the script: say "representative" and keep saying it, choose the menu path that threatens to cancel or dispute a charge, and switch to a channel the bot doesn't own — app chat, email, or a public post. Staying in the bot's own conversation and asking politely is the slowest option you have.
This is now a mainstream problem, not a niche annoyance. And the numbers behind it are worse than most companies admit.
What the data says about the fight
A "Consumer Patience Index" poll commissioned by Parloa — itself a company that builds customer service AI agents, so read the source with that in mind — surveyed 1,001 US adults and found:
- More than half had actively tried to circumvent a chatbot.
- 43.9% of those resorted to yelling "human" or "person" down the phone line.
- 17% went as far as swearing at it to break free.
- More than half said they would give an automated system only three minutes before walking away.
- The single most frustrating pain point was "talking to a bot that doesn't understand me" (25.9%), ahead of long hold times (22.8%) and being transferred repeatedly (13.4%).
- Only 13.6% trusted an AI to handle a complex request; 30.4% had no trust at all.
- But 85% said they'd happily use automation that resolves their issue nine times out of ten.
That last pair is the whole story. People are not anti-automation — they're anti-non-resolution. The frustration is not that a machine answered. It's that the machine can't finish the job and won't hand it over.
Why the bot can't let you go
It isn't usually malice. It's measurement. Support AI is commonly evaluated on containment — the share of conversations that end without a human — so the system you're fighting was optimized to keep you in it. Meanwhile the escalation trigger a human would use ("this customer is stuck and angry") is exactly the signal the bot is least rewarded for acting on.
That mismatch is why the techniques below work. They don't persuade the bot; they push it past the threshold where the design says "this one is no longer my problem."
Note that a genuinely good setup escalates on detected frustration, complexity, or an explicit request — proactively, not after a failure. If you're on the building side, that's the design to copy. If you're on the receiving end, you're working around a system that skipped it.
The phrases and menu paths that work
- Name the destination, not the problem. "Representative." "Agent." "Human." Say it in the first sentence and repeat it verbatim each time you're re-prompted. Long explanations give the bot more surface to route you back into.
- Refuse to authenticate until it matters. Voice agents often ask for an account number or a code before they've understood the request. You can say: "I'll verify once I'm speaking to a person."
- Take the highest-stakes branch on offer. Menus rank by business risk. "Cancel my account", "dispute a charge", and "report fraud" reach humans far more reliably than "general question."
- Press 0 and hold. On many phone trees a repeated
0still breaks out, even when it isn't announced. So does staying silent. - State one precise, checkable outcome. "I need the £42 charge from August 14 refunded" is escalatable. "My bill seems wrong" is not.
- Never accept "I'll create a ticket" as the end. Ask for a reference number and a timeframe, in writing, on a channel you can reply to.
- Escalate the channel, not the volume. The fastest lever is usually leaving the bot's channel entirely.
Which channel reaches a human fastest
| Channel | Typical speed | Why it works |
|---|---|---|
| Public social post (X, the brand's Facebook page) | Fastest — often minutes | Visible to people whose job is brand reputation, not containment |
| Phone, using a cancellation or dispute branch | Minutes to tens of minutes | Routes to retention or billing humans, who are staffed to close |
| In-app chat, asking explicitly for a human | Minutes to hours | Ticket is already attached to your account and order history |
| Email to a named support address | Hours to days | Written record, but the slowest queue |
| The company's own AI chat, staying polite | Often never | Optimized to contain you; the default state is a loop |
The pattern: the more a channel exposes the company to public or revenue risk, the shorter the path to a person.
When to stop fighting and leave
A separate survey by Acquire BPO found 70% of consumers would consider switching brands after a single frustrating AI support experience. That is the leverage you actually have — and the reason "I'll cancel" is often the most effective sentence in the conversation.
If a company makes reaching a human cost you twenty minutes twice in a row, that's information about the company, not about you. Three minutes is roughly the industry's own estimate of consumer patience. Budget accordingly.
The other side: an agent on your behalf
The mirror image of this problem is delegation. If the job is holding a line, chasing a refund, or drafting a firm written complaint, that is exactly the kind of repetitive, low-judgment work an AI agent handles well on your side of the table — the same category as the outbound calling tools in our guide to agents that make phone calls, and the reason AI voice receptionists exist.
For written escalation the bar is low: a local-first agent like Wolffish can keep the case timeline, draft the complaint email in your words, and remind you when the promised call-back is overdue — with approvals on your phone before anything is sent.
The rule stays the same in both directions: agents should escalate when they're stuck, and say so plainly.
The takeaway
Don't argue with the bot — route around it. Say "representative" and repeat it, take the cancellation or dispute branch, and move to the channel where a human's reputation is on the line. The 43.9% who yell "human" aren't wrong about the goal; they're just using the slowest possible method to get there.
