Escalating to a human: how good implementations handle it

The escalation path gets built last and matters most. Everyone spends their attention on the greeting, which nobody remembers, and leaves the transfer as a checkbox in a settings panel.

Then a caller with a real problem asks for a person, gets sent to a phone nobody is holding, and lands in the voicemail the system was bought to replace.

When should a call escalate?

Five triggers, and they should be written into the script explicitly rather than left to the model's judgment.

The caller asks. First request, no negotiation, no "I can help you with that." A system that argues about being replaced is the single most irritating pattern in this category.

Repeated confusion. If the agent has asked the same question twice, or the caller has restated the same thing twice, the conversation is not recovering. Escalate on the second repetition, not the fourth.

Emotional signal. Anger, distress, urgency, or anything a person would recognize as needing a human. A burst pipe caller does not want to be qualified.

Compliance boundary. These escalate instantly, regardless of how casually the question was asked, and they should be hard-coded triggers rather than knowledge base lookups. In a restaurant, the words allergy, allergic, celiac, anaphylaxis, EpiPen and every named major allergen. The FDA Food Code assigns allergen knowledge to a designated person in charge and satisfies its notification duty in writing, which a voice agent is not. In a dental practice, coverage and out of pocket questions, because an AI is unlicensed personnel operating a provider-approved verbatim tree.

Value threshold. If the job or party size is above a number you set, a person should touch it.

What is escalation actually worth?

Run the number, because "escalate more" is only good advice when the arithmetic supports it.

Price the call. Take a home services shop with a $400 average ticket. ServiceTitan's platform data from June 2022 puts the booking rate for a shop with fewer than five technicians at 24%, against 42% for a typical shop and 59% for one with 25 or more. At 24%, an inbound call carries about $96 of expected revenue before anyone answers it.

Price the escalation. A staff member spending three minutes on a transferred call, at a loaded hourly cost of around $35, costs you roughly $1.75.

Now find the break even. Escalation is worth it if it raises the chance of booking by more than $1.75 divided by $96, which is under 2%. In other words, if handing the call to a person improves the outcome by even two percentage points, escalating pays.

That is why the default should be to escalate. The asymmetry is enormous, and it gets more extreme as the ticket rises. At a $2,000 job the break even is a fraction of a percent. The only businesses where the arithmetic reverses are the ones with low ticket values and high call volume, and they are exactly the ones that should not be routing everything to a person anyway.

Rerun it with your own ticket and your own labour cost. If your answer also lands under a few percent, stop treating escalation as a failure state.

What does a good handoff sound like?

Short, and it never makes the caller start over.

The line I would use: "Let me get you to someone who can help with that directly. One moment." No apology, no explanation of limitations, no restating what the agent cannot do.

Then the context travels. Whoever picks up should already have the caller's name, number, and what they said. If your staff member answers with "how can I help you," the handoff failed, because the caller now repeats themselves to a second party and concludes that nobody in your business talks to anybody else.

Warm transfer where you can. The agent stays on, briefs the human in a sentence, then drops. It costs a few seconds and it changes how the whole call is remembered.

Where a live transfer is not possible, the callback promise has to be specific and honored. "Within fifteen minutes" is a commitment. "As soon as possible" is noise, and callers have learned to hear it as no.

Capture the number before you attempt any transfer. Transfers fail. A number captured is a lead recovered.

And watch the clock during the handoff, because hold time is where calls die. Patient Prism scored 11,552,668 dental calls across 8,280 locations in calendar 2025 and found that 31 of every 100 callers hung up before reaching an agent at all. Those callers did not reject a machine. They rejected a wait. An escalation that parks somebody in a queue has reproduced the exact failure you bought the system to fix.

What happens when nobody picks up?

This is the case that separates a designed system from a configured one, and it is the case that will happen most.

Define the ladder before launch. Primary number rings for a set number of seconds, then a second number, then a third, then a defined fallback. The fallback is not voicemail, or at least not only voicemail.

The strongest fallback I know is a text message to the caller within seconds of the failed transfer, from a number that can receive replies, saying a specific person will call back within a specific window. It moves the conversation to a channel the caller can act on while they wait.

The reason it beats a callback is measurable. A TNS survey from July 2022 found that 75% of Americans never answer calls to their wireless phone from a number they do not recognize. Your callback attempt is an unrecognized number to a first time caller. A text is not.

That text carries obligations. Under the FCC's revocation rules, in force since 11 April 2025, a consumer may revoke consent by any reasonable method, and you must treat any wording a reasonable person would read as a stop request as valid, not just the word STOP. The engineering consequence most deployments miss is that a spoken opt-out on the voice call has to write into the same suppression list your texting platform reads. One table, one phone number key, shared across voice and SMS. Ask your vendor whether that is wired, get the answer in writing, and have counsel review the program.

The escalation ladder also needs hours attached. A transfer rule that makes sense at 2pm is a promise you cannot keep at 9pm. After hours, the honest move is usually to book rather than transfer, which only works if the agent can write into the calendar you already use instead of leaving a message.

And someone has to own the failed transfer queue. Not "the team." A name.

How much escalation is too much?

Both directions are a problem, and the number tells you which one you have.

Escalation running very high means the script does not cover your actual calls. The fix is script coverage, not a stricter transfer rule. Read the transcripts of the escalated calls and you will usually find three or four question types that account for most of them.

Escalation running near zero means the agent is answering things it should be handing off. That is worse, and it is invisible until a customer complains. Zero escalation on a phone that takes real calls is a red flag, not an achievement.

Which brings us to containment rate, the most gameable metric in this category. Containment is the share of calls the system resolved without a human. It is the number vendors lead with, and it improves whenever escalation gets harder. A system that refuses to transfer looks excellent on a containment dashboard and is failing every caller who wanted a person. Containment should never be read on its own. Read it beside escalation rate, transfer success rate and caller satisfaction.

Transfer success rate is the one almost nobody measures. Not how many calls were escalated, but how many escalations actually reached a human who handled the call. Ask your vendor for that number. If they cannot produce it, they are not measuring the part that matters.

The strongest ground truth signal is callback within 24 hours, and it is not a native feature on any platform. It has to be built by correlating call logs against your CRM or practice management records. Take every call the system marked as successfully contained, then count how many of those numbers dialed you again within a day. If 30% of your contained callers called back the next day, your containment rate is a lie. That single report is worth more than every other dashboard the vendor ships.

Review the escalated calls weekly for the first month. They are the highest signal transcripts you have, because every one of them is a place your system reached its edge.

Escalation rate is also a leading indicator of trouble elsewhere. A rise usually shows up before the signs your business is losing customers on the phone become obvious in your booking numbers.

Does escalating make the AI look bad?

Less than you think, and the largest public deployment in the industry is the proof.

Wendy's is the drive-thru voice AI programme still expanding, past 500 company-operated and franchise locations by late 2025, and it reports that 86% of transactions complete without employee intervention. Those figures are company-reported and unaudited, so label them. But the interesting number is the other one. 14% of orders are deliberately routed to a human. Wendy's is not claiming full automation, and that is precisely why its deployment outlasted two larger ones that did.

Callers work out they are talking to software fairly quickly, and what they do next depends almost entirely on how the system behaves after that. A system that hands off cleanly when asked reads as competent. A system that insists it can help while failing to help reads as a company trying to avoid you.

The same principle governs chat. Whether live chat or a chatbot helps a small business turns on the same question: is there a human behind it when it matters.

It also matters competitively. If you are up against national lead aggregators in your own market, the thing you have that they do not is a local person who will pick up. Hiding that behind software gives away your only structural advantage.

What to do this week

Write the five triggers into your script as explicit rules, with the exact escalation line. Hard-code the compliance words rather than leaving them to interpretation.

Run the break even arithmetic above with your own ticket value and labour cost, so you know what an escalation is actually worth before you tune the threshold.

Build the ladder: primary, secondary, tertiary, fallback, with ring durations and hours. Put a name against each step.

Set up the text fallback for failed transfers, with wording that names a person and a time window, and confirm the spoken opt-out writes to your suppression list.

Ask your vendor for transfer success rate, not just containment rate.

Then test it the only way that counts. Call your own number, ask for a human, and have every person on the ladder deliberately not answer. Watch what the caller experiences. Fix whatever you saw.

Questions about your escalation path?

Email me at eric@seod.com with your current transfer and after hours rules, written however you have them, even if it is "it rings my cell." I will tell you exactly where a caller gets stranded and what I would put in the gap, including the fallback wording.

I do this myself and it takes about ten minutes to spot the hole. There is almost always one, and it is almost always the same one.

More on the surrounding pieces sits in the AI phone and lead response library.