A December 2025 SurveyMonkey study of 2,017 US adults found that 79% of Americans prefer human customer service over AI. That number has been circulating in developer Slack channels and product team decks ever since, usually as a reason to slow down on AI calling.
It's being misread.
The same study found that 82% of customers would rather talk to an AI than wait for a human agent. These two numbers coexist because preference is deeply contextual. When given the choice between a competent AI now and a human in 20 minutes, most people take the AI. When given the choice between a capable human and a mediocre AI, they take the human.
The 79% figure isn't a ceiling on AI calling. It's a description of where most AI calling is today.
What Taco Bell's disaster actually teaches us
Taco Bell deployed voice AI ordering at 500+ drive-throughs and ran into a cascade of very public failures. The AI struggled with accents. It misheard orders. Customers discovered they could game the system by ordering absurd quantities — 18,000 water cups became a known bypass technique to get a human. The company eventually scaled back, giving individual franchisees the choice of whether to use AI at all.
The tech press covered this as "AI voice fails at fast food." The real lesson is more specific: the failure wasn't the voice model. The failure was scope containment.
A drive-through is adversarial. Customers are in cars with engine noise. Kids are yelling. People order half a menu item and change their mind mid-sentence. The corrections loop — "No, I said two, not one" — requires the system to handle state in a way most voice AI designs don't account for.
McDonald's ended its partnership with IBM for similar reasons. White Castle took a slower, more measured approach — and by most accounts has had fewer public failures.
The difference between those outcomes isn't the underlying model. It's what the model was asked to handle.
Why AI calls fail — and it's mostly not the voice
Why AI Phone Calls Fail
estimated share of call failures by root cause — hover to inspect
The most common reason an AI call fails isn't voice quality or an awkward pause. 34% of AI call failures happen before the call rings — carrier spam filters label the call "Spam Likely" and the recipient either never sees it ring or declines it on sight. No amount of prompt engineering fixes a number that trips Hiya's scoring model.
The second and third biggest failure modes are latency and no escalation path. The voice sounds fine. The AI just takes 1.4 seconds to respond, which feels broken in a phone conversation. Or the caller hits an edge case the agent wasn't designed for, gets stuck in a loop, and hangs up.
Voice quality — the thing most developers obsess over — accounts for a small fraction of actual failures.
The five design principles that change the outcome
1. Contain the scope aggressively
The most effective AI calling agents do one thing. Not "sales and support and scheduling." One thing.
Containment isn't a limitation — it's the design. A narrowly-scoped agent is faster, makes fewer mistakes, and recovers from edge cases more gracefully because the set of edge cases is small. When an agent tries to handle everything, the tail of failure cases grows exponentially.
Before building, write down every reason a caller might go off-script. If that list has more than five items, the scope is too wide for a single agent. Split it.
2. Disclose early, confidently — not apologetically
Most AI call openings sound like an apology: "Hi, this is an automated system and I'm going to try to help you with..."
This is backwards. The disclosure is not the bad news. Lead with what the agent can do:
> "Hi, this is an AI assistant from Acme — I can help you confirm your appointment, reschedule, or transfer you to the team. Which would you like?"
The caller knows immediately what to expect and what's available. Framing the AI disclosure as a feature rather than a warning changes the perception before the first substantive exchange.
This also matters legally. Article 52 of the EU AI Act requires disclosure before or at the start of any real-time AI conversation with a human. TCPA guidance is evolving in the same direction. Getting the disclosure right is simultaneously good design and good compliance.
3. Budget for latency in your conversation design
A 400ms pause in a face-to-face conversation is imperceptible. A 400ms pause on the phone feels like the line dropped.
Phone conversations have a different latency expectation than chat. Callers interpret silence as failure. The practical target for conversational AI phone calls is a response initiation under 800ms — not the full response, just the first word.
Two techniques close this gap without waiting for better infrastructure:
Filler words as placeholders: "Sure, let me check that for you—" buys 600–800ms while the model generates. It sounds natural. An abrupt silence does not.
Optimistic branching: For high-probability response paths, begin generating the most likely response before the caller has fully finished speaking. If the branch is wrong, interrupt gracefully. Most of the time, it isn't wrong — and the perceived latency drops significantly.
4. Design the escalation path before designing the agent
The escalation path is not a fallback — it is the product. Every AI call should have a clear, fast, non-degrading way for the caller to reach a human.
"I'm going to connect you with someone from the team now" is a resolution, not a failure. Callers who escalate and get a good human experience have a better overall impression than callers who stayed in a loop and eventually hung up.
Design the escalation path first. Then build the AI around it. The AI's job is to handle the common cases efficiently. A person on the other side handles the rest. Both are real customer service — only one scales.
5. Trust signals before content
The first ten seconds of a call determine whether the caller engages or disconnects. Most AI calling agents spend those ten seconds on disclaimers and preamble.
Lead instead with a signal that creates trust: the caller's name, a reference to their account, or a specific reason for the call.
> "Hi Sarah — I'm following up on the appointment you booked Tuesday. Does 2pm still work for you?"
Compare that to:
> "This is an automated message from Acme Health regarding your upcoming appointment scheduled for..."
Both deliver the same information. One signals that the system knows who the caller is and has a specific purpose. The other signals a broadcast.
The first version gets more answers. The second gets more hang-ups.
The prerequisite nobody mentions
All five of these principles assume the call is actually reaching the recipient.
Before a caller hears your agent's voice, the call passes through carrier analytics systems — Hiya, TNS/Call Guardian, First Orion — that score the number's reputation based on call volume, answer rate, whether the number receives inbound calls, and acoustic fingerprinting of the audio. Numbers that look like programmatic dialers — high volume, low answer rate, no inbound history — get labeled "Spam Likely" before the first ring.
If 34% of your calls are being killed at the carrier layer, none of the conversation design work matters for those calls.
The infrastructure prerequisite is: a number that has been provisioned correctly for programmatic use, with STIR/SHAKEN A-level attestation, active reputation management, and carrier relationships that don't treat your AI agent traffic as a threat. That's not a feature — it's the foundation the conversation design sits on.
The 79% who prefer humans are reachable. They prefer humans because most AI calls they've received have been bad — robotic, untimely, trapped in useless loops, or outright scams.
Build the call well and you're not competing with human agents on warmth. You're competing with the previous AI call they received, which set the bar at approximately zero.
The quick checklist
If the answer to all six is yes, you're ahead of most AI calling deployments in production today.