---
title: "When AI Receptionists Fail: What Business Owners Should Know"
description: Two documented AI customer service failures, Klarna and Air Canada, plus a consumer survey, show what to evaluate before putting AI on your customer calls.
---

[People Answering People](https://amessagecenter.com/people-answering-people)

# [When AI Receptionists Fail: What Business Owners Should Know](https://amessagecenter.com/people-answering-people/when-ai-receptionists-fail)

 Written by [Shaun Grubert](https://amessagecenter.com/people-answering-people/author/shaun-grubert) | Nov 17, 2026, 12:00:01 PM

## Introduction

When evaluating an answering service, most business owners ask one of two questions: "Will it save us money?" or "Will callers accept it?"

Those are the wrong questions.

The right question is simpler and harder: "What happens when this service fails?"

It's a hard question because vendors never lead with their failure modes. They show you their demos. They quote their efficiency gains. They publish satisfaction ratings, often based on people who never used the system for your kind of call.

What they don't show you are the cases where an automated customer service system stopped working, and what the fallout looked like.

But those cases exist. They're documented. And they teach you more about whether an AI receptionist fits your business than any vendor demo will.

This article is built from two documented cases and one consumer survey. Both cases involved text chat, not voice calls, and we say so up front because it matters: the lessons transfer, but phone calls bring failure modes of their own, which we cover below. Each piece of evidence points to a different question you should be asking before you commit to an AI-first model for your after-hours calls.

## Two Documented Cases and One Survey: Real Evidence, Real Lessons

### Pattern 1: The Accuracy Ceiling, Klarna's Rebalance

In early 2024, Klarna, the Swedish payments company, announced an AI assistant to handle customer service chats. The business case was simple: reduce support costs, handle volume at scale, keep satisfaction high.

The early numbers looked good. Klarna reported that the assistant handled about two-thirds of its customer service chats in its first month, and it projected a profit improvement of roughly $40 million for the year. For straightforward questions, such as an order status or a payment due date, the assistant performed well.

Then, in May 2025, Klarna's CEO said publicly that the company had leaned too hard on cost, and that the result was lower quality. Klarna did not abandon AI. It rebalanced: it kept automation for routine questions and began adding human agents again, so that customers could reach a person when they wanted one.

**What this teaches:** When cost is the main design goal, quality is what gives. Every system has a ceiling on the call types it handles well, and the system doesn't always know where its ceiling is. For businesses where a wrong answer carries operational or compliance cost, such as medical practices, property management emergency dispatch, and HVAC emergency triage, that ceiling matters. Remember this was text chat. On a phone call, with a caller who may be stressed, the margin for error is smaller.

### Pattern 2: The Liability Problem, Air Canada's $812.02 Award

In November 2022, Jake Moffatt's grandmother died. He booked flights from Vancouver to Toronto through Air Canada's website, and he asked the site's chatbot about bereavement fares.

The chatbot told him he could book at the regular fare and apply for the bereavement discount afterward, within 90 days of the date the ticket was issued. That was not what the airline's policy said. The published policy did not allow bereavement refunds once travel had already happened.

When Moffatt applied, Air Canada denied the request and pointed to its policy page. He took the dispute to the British Columbia Civil Resolution Tribunal, a small-claims body.

In February 2024, the tribunal ruled against Air Canada. It found the airline liable for negligent misrepresentation. Air Canada had argued, among other things, that the chatbot was a separate entity responsible for its own statements. The tribunal rejected that argument and noted that the airline had not explained why a customer should have to double-check its chatbot against another page on the same website.

The total award, including damages, interest, and tribunal fees, was $812.02 CAD.

The dollar amount is small. The signal is not. A tribunal decision is not binding precedent, but this one is widely cited for a simple idea: a company can be held responsible for what its own chatbot tells customers, even if the correct policy is written somewhere else on its site.

**What this teaches:** A system that sounds confident but is wrong creates liability for you. The customer trusts it because it carries your name. Oversight, testing, and a clear path to a person are not optional on a customer-facing channel.

### Pattern 3: The Frustration Threshold, What a Consumer Survey Found

An answering-service company, in the same industry as us, commissioned a OnePoll survey of 6,000 consumers across the United States, the United Kingdom, and Canada, and released the results in May 2026. It compared an October 2025 wave with an April 2026 wave. That company sells human answering, which is worth knowing as you weigh the findings. We'd rather you hear that from us than discover it later.

Two results stand out. The share of respondents who said they would hang up if they were connected to an AI system rose from 29% to 31% between the waves. And the share who said they would trust a business less if it relied predominantly on AI for customer service rose from 53% to 57%.

This is a survey of attitudes, not a measurement of how any particular AI system performs. It matters anyway, because trust compounds. A customer who has a bad experience with an automated system doesn't just remember the failed call. They carry that skepticism into the next interaction with your company, even if a person answers the second time.

**What this teaches:** AI systems can handle routine calls efficiently. But if your callers expect a person, or if they're calling because something is wrong, an automated system starts the conversation with a deficit of trust. The call may be answered quickly, but the caller may leave feeling the company didn't care enough to staff a real person.

## Where Voice Adds Its Own Failure Modes

Both cases above were text chat. Phone calls add problems a chat window doesn't have.

**Timing.** A pause that runs a little too long, or a system that talks over the caller, makes people repeat themselves or hang up.

**Sound.** Background noise, accents, and speakerphones trip up speech recognition, and the errors land on the details that matter most: a name, a callback number, a street address. Ask whether the system reads those details back to the caller to confirm them.

**Handoffs.** A warm transfer to a person can drop the call or lose the context, so the human starts from zero.

**After-hours edge cases.** The panicked caller, the caller who answers a different question than the one asked, the caller who says something no script planned for.

**Hallucination.** A fluent, confident answer that is simply wrong.

None of these show up in a polished demo. All of them show up on a Tuesday at 2 AM. For more on how voice AI works and what it costs, see our guide to voice AI for answering services.

## What the Evidence Reveals: The Evaluator's Framework

If you're considering an AI answering service, or evaluating whether your current hybrid setup is right, these three patterns point to three distinct risks worth measuring.

### Risk 1: Accuracy Ceiling for Complex Call Types

**The question to ask:** What share of my calls fall outside what this system handles well? What happens when it meets a call type it wasn't built for?

For Klarna, the gap was between routine transactional questions and the harder cases. For a medical practice, the gap might be between appointment confirmations and clinical judgment calls. For property management, the gap is between leasing inquiries and after-hours emergency dispatch.

**How to test it:** Ask the vendor to walk you through how the system handles 10-15 realistic scenarios drawn from your business, including your edge cases. Not their highlight reel. What does the system do when it doesn't understand the caller's intent? Does it hand off to a person immediately, or does it try to resolve the call and fail?

If the vendor can't produce edge-case walkthroughs, or tells you "our system handles everything," that's a signal. Good systems have documented limits.

### Risk 2: Liability for Misstatement

**The question to ask:** If your AI system gives a caller information that contradicts your published policy, who is liable?

This is a contract question and a compliance question. You should know the answer before you sign.

**How to test it:** Request the vendor's liability terms in writing. Specifically, ask: (1) Does the vendor carry errors-and-omissions insurance that covers misstatements by the system? (2) If the system states something factually wrong, who bears the liability, you or the vendor? (3) Does the vendor expect you to maintain human oversight of calls that involve policy or compliance-sensitive information?

A vendor that hedges or deflects on these questions hasn't thought them through.

### Risk 3: Caller Trust and Experience

**The question to ask:** For the callers who matter most to my business, such as my high-value accounts, my distressed callers, and my relationship-dependent interactions, am I gaining anything by automating their first interaction?

This is harder to measure, but it's the most important one.

**How to test it:** Start with 10-15 realistic scripted scenarios, not live callers. If you later move to live calls, do it only after a BAA is in place (if any protected health information could be spoken), on a small subset of calls, with a human backstop for every call. Segment the results by caller type: routine versus complex, new prospects versus existing clients, off-peak versus peak. Track what matters most to your business: resolution rate, escalation rate, follow-up time, account retention.

If your highest-value callers show lower resolution rates or higher escalation rates with automation, that's a signal it is creating friction where you can't afford it.

## The Comparison: When AI Answering Fits, and When It Doesn't

| **Call Type** | **AI Strength** | **AI Risk** | **Recommendation** |

|---|---|---|---|

| **Routine intake** (appointment confirmations, FAQ routing, callback capture) | High accuracy, cost-effective at volume | Low risk | Good fit; verify name and number capture |

| **Complex triage** (medical urgency assessment, PM emergency dispatch, HVAC emergency vs. routine) | Can sound fluent while being wrong; limited real judgment | **High risk**, wrong routing costs money | Human-first / human in the loop |

| **Distressed or upset callers** | Limited de-escalation capability | Callers may hang up or ask for a person; account damage | Human-first / human in the loop |

| **Compliance-sensitive calls** (PHI, payment disputes, liability questions) | Hallucination risk, liability exposure | **High risk**, wrong answers = legal liability | Human-first / human in the loop, with human documentation |

| **Long-standing client relationships** | Pattern-matching, not relationship memory | Trust erosion if caller expects a person | Hybrid: AI for intake, a person for complex |

| **High-value account calls** | Efficient routing | Caller frustration if not escalated properly | Hybrid or human-first |

## Five Questions Before You Commit to AI Answering

**When should I not use an AI answering service?**

You should not lead with AI answering when:

1. Your callers are frequently stressed or in urgent need. Patients reporting symptoms, tenants dealing with emergencies, clients with immediate problems: these callers need someone who can assess urgency. A system that sounds fluent but is wrong creates frustration, not efficiency. The safeguards are narrow scope and clear escalation to a person.

1. Your calls require real-time judgment about escalation. "Is this a habitability emergency or routine maintenance?" "Is this a chest pain or a refill question?" These are judgment calls, and a fluent system can still get them wrong. A trained human operator, working from your own protocol, makes them in real time.

1. Your business depends on relationship continuity. If your customers expect to be known by name, or your model relies on account retention and relationship depth, an automated first interaction creates friction. The caller may not notice at first. They'll notice when they have to re-explain themselves.

1. A wrong answer carries liability risk. Policy misstatements, compliance errors, false warranty claims: if your system can state these wrongly, you're exposed. Legal liability doesn't care whether the mistake was made by a person or a machine. You own what your system says.

1. You can't verify accuracy on your specific call types. If the vendor can't show you edge-case performance on calls that look like yours, you're flying blind. Demos always work. Real calls are messier.

## Where This Leaves You

The three patterns, accuracy ceilings, liability exposure, and trust erosion, aren't reasons to reject AI answering entirely. They're reasons to reject AI-only answering for high-stakes call types.

Many businesses choose a hybrid model: automation handles high-volume, low-variance, low-urgency calls, and people handle anything involving judgment, compliance, escalation, or relationship value.

The vendor question worth asking is not "Can your system handle everything?" It's "What does your system explicitly not handle, and how do we make sure those calls get to a person?"

An honest vendor has a clear answer.

## Frequently Asked Questions

**What do the Klarna and Air Canada cases teach me about choosing an AI answering service?**

Two lessons. First, when cost drives the design, quality can slip: Klarna kept AI for routine questions but added human agents back after leaning too hard on savings. Second, you are responsible for what your system tells callers: the Air Canada tribunal ruling is widely cited for that point. Both cases were text chat, but the lessons transfer. Test accuracy on your specific call types and keep a clear path to a person.

**Can an AI answering service handle medical practice calls?**

Some AI services market HIPAA compliance and BAA coverage. The question to verify is whether the BAA covers every layer the audio touches: telephony, speech-to-text, the language model, text-to-speech, and recording storage. A gap at any layer is a gap in your coverage. For a medical practice, a human answering service is usually a simpler chain to verify. Get the BAA's scope in writing before you sign.

**Is AI answering cheaper than live answering?**

It depends on call volume and call type. AI services are typically priced per-minute or per-interaction; live services typically charge a monthly base plus usage. At high volumes of routine calls, AI can be cheaper. At low volumes, or for any call requiring live judgment, a monthly-base live service is often more cost-effective. Run your last three months of call volume through both pricing models before deciding.

**What does the consumer survey say about AI customer service?**

A OnePoll survey of 6,000 consumers across the US, UK, and Canada, commissioned by an answering-service company in our industry, compared an October 2025 wave with an April 2026 wave. The share who would hang up on an AI system rose from 29% to 31%, and the share who would trust a business less for relying on AI rose from 53% to 57%. It measures attitudes, not system performance, but an AI system that saves money on call handling and costs you customer trust is a bad trade. Test the impact on your own callers before scaling.

**How do I know if my AI answering service is failing before it becomes a public problem?**

Set clear success metrics before you deploy: resolution rate, escalation rate, customer satisfaction, retention impact on your highest-value accounts. Review the data monthly, segmented by call type and caller segment. If your most important callers show declining resolution rates or higher escalation rates, that's a signal to scale back automation for those calls and bring people back in.

## What We Believe

A Message Center is a live human answering service. We've been answering calls since 1962. We watch the AI answering space closely because our clients ask about it, and we have a responsibility to give them honest guidance.

What this evidence teaches us is this: AI answering services are legitimate tools for some call types and some businesses. They are not appropriate for every call. The business owners who get this right aren't using AI because it's trendy; they're using it for the specific call types where it works, and they're keeping people in the moments that define customer relationships.

If you've worked through the framework above and want to talk through where your call profile falls, we're available.

*A Message Center is a US-based, family-owned live telephone answering service, answering since 1962. Human operators, HIPAA-compliant, BAA available for review. 24/7/365. People Answering People.*

[View full post](https://amessagecenter.com/people-answering-people/when-ai-receptionists-fail)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Shaun Grubert"
  },
  "dateModified" : "2026-11-17T12:00:01.376Z",
  "datePublished" : "2026-11-17T12:00:01Z",
  "headline" : "When AI Receptionists Fail: What Every Business Owner Evaluating Automated Answering Should Know",
  "image" : {
    "@type" : "ImageObject",
    "height" : 1024,
    "url" : "https://43961547.fs1.hubspotusercontent-na1.net/hubfs/43961547/AI-Generated%20Media/Images/Stressed%20Business%20Owner%20With%20Broken%20Ai%20Bots.png",
    "width" : 1536
  },
  "mainEntityOfPage" : "https://amessagecenter.com/people-answering-people/when-ai-receptionists-fail",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60,
      "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
      "width" : 60
    },
    "name" : "People Answering People"
  }
}
```