While things are still evergreen in the AI space (albeit slowing down), the general consensus seems to be that interactive AI voice technology still has a way to go. Our information on this comes from our own testing of the tech as a receptionist for Phonetico customers, along with anecdotal experiences, and some scarce epidemiological stats. We're hoping more people take the time to look into how this technology affects a business's branding, its clientele, and its bottom line, so we can all get a clearer picture of whether AI receptionists are truly worth it.
An AI receptionist is something you may have already interacted with during a customer service call without realizing it. It’s software that can have natural-sounding conversations with customers over the phone. It can greet callers, answer basic questions, handle simple requests, and can even create tickets for more complex issues that need to be handled by a real person.
More and more companies who rely on someone to greet their customers are starting to look into offloading that work from their administrative assistants. Meanwhile, attitudes of businesses are evolving from curious consideration into all-in adoption. Over 60% of small businesses have already started using AI tools to automate customer interactions, including virtual receptionists. And on the enterprise side, Gartner found that 85% of customer service leaders planned to explore or pilot a customer-facing conversational AI solution, with more than 3/4 saying they're feeling real pressure from execs to get something implemented. We're seeing that whether or not the tech is ready, the appetite is most definitely there.
While these models are already being used heavily in the spam calling market, we're going to stick to talking about receptionists for legitimate businesses (but we will definitely be delving into spam calls using this tech in a future article, because it's getting pretty bad).
Since we're testing AI receptionist technology as a feature for our own VoIP clients, we wanted to dig into the pros and cons of putting an AI agent first in line to speak to your customers.
Aahh, finally, the menial task of answering the phone can be automated. Gone are the days of explaining to an unassuming caller information they could've just found on the website, or absorbing the stress of an angry support call. Taking calls from a confused, frustrated, or worse, confused and frustrated, customer can be really taxing on the employee on the other end. It's a well-documented reality that client-facing jobs, whether in person or on the phone, wear human agents down over time. Constantly managing someone else's frustration while keeping your own tone pleasant is its own kind of labour, and that emotional and mental offloading can be a serious game changer for any company, as you can always rely on a robot to customer-service-voice-max. On the other hand, you can't always trust that from your CS employee after the sixth call of someone yelling at them because they can't remember their password.
We should note that our research on the good, outside of our own support team's experience, is a bit limited. The praises being sung about AI receptionists seem to be coming from three sources:
We'd be doing the technology a disservice if we didn't talk about how believable the voice itself has gotten, because it's getting really good in terms of believability. OpenAI, for example, recently came out with a multimodal voice AI that makes most people not second-guess it at all. Chances are, if you've called a company in 2026, you may have spoken to one of these human-like voices without ever realizing it. Creepy, yes, but it seems like this thing is up there in terms of usefulness and effectiveness. It's estimated that about 80% of customers never realize they're talking to an AI agent (via our own tests), a pretty impressive number when you remember those clunky robot voices of yesteryear. But a good business owner should be pretty wary of that remaining 20%. And we haven't even touched on whether these AI agents actually succeed at the tasks that make that human-like voice worth anything in the first place.
They're also getting good at the small performative details that sell the illusion, like fake call-centre background noise and the sound of typing on a keyboard while they "look something up." It's a nice touch, until you realize it exists purely to keep you from clocking that you're talking to a machine.
Despite how human it sounds, the people actually working with this technology are ringing alarm bells about its effectiveness. Cory Doctorow recently wrote about a Mitsubishi customer, Nikhil Suresh, who described his one positive AI receptionist experience: a Mitsubishi bot with a natural, responsive voice that politely took down all the details of his car trouble and promised a callback. That callback never came. Suresh figures Mitsubishi has probably logged the interaction as a success story anyway, even though the whole experience is what convinced him not to buy a Mitsubishi.
And as people become more aware that this tech is out there, they're starting to question who's actually on the other end of the line. That can create a pretty strange experience for a customer who realizes they're talking to an AI after assuming they were speaking to a person. It's understandable that people might feel misled, especially when these tools are purposefully designed to pass as human, right down to the fake background noise.
Another problem is that these agents still lag behind humans on tasks like basic customer identification. If a customer pronounces their business name in a way the AI doesn't recognize, it can get stuck entirely, sometimes telling the customer outright that their account doesn't exist, when a human would almost always figure out what they meant even with an awkward pronunciation. Part of this comes down to the fact that these agents can only answer based on what's in their documentation. A human can improvise, infer, and fill in the gaps. An AI agent can only work with what it's been fed, and when a caller falls outside that script, it shows. So AI is still hit or miss on completing simple verification tasks, and getting it right takes hours of configuration for edge cases. It's still a lot more rigid than a human, and at around 50 cents a minute, or $30/hour (billed on actual call time), it's up to the finance team to decide if that's worth it.
Data sovereignty
There's another issue that matters a lot in Canada: where the data actually goes.
The best-sounding LLMs are currently only available in hosted formats from companies like OpenAI and Anthropic, which deploy servers outside of Canada. For businesses handling sensitive customer information, that creates an additional layer of privacy and compliance considerations under PIPEDA and applicable provincial privacy laws. It's not simply a matter of whether the technology works; businesses also need to know where their data is being processed and stored, and who has access to it.
Background conversations
AI receptionists also have a harder time dealing with the messy reality of a phone call. They can struggle to distinguish between what the caller is saying and what other people or devices are saying in the background. A television, another conversation, or even someone speaking across the room can sometimes be enough to confuse the agent.
For the caller, that can mean having to repeat themselves, correct the AI, or explain something that a human would have understood immediately. And when you're already calling because you need help, that extra friction gets frustrating pretty quickly.
This has real world business consequences. In fact, we would argue that it's unprecedented how a single piece of technology could completely change a customer's view of your business and taint it for good, as many businesses seem to be putting their reputations on the line by making their customers feel duped. They are playing customer-perception roulette after spending years and thousands, if not hundreds of millions of dollars building positive brand sentiment. Now all it takes is 30 seconds for a wary, hard-earned customer to ask "are you AI?" to an AI receptionist, and that's it. Tainted. They'll never see you the same way again. And then they'll tell their friends, maybe even post about it online. Is it worth it? We're not sure yet. As a business VoIP provider, we want to make every one of our clients happy, and we keep working at it until that's the case. But this is a technology that can shift how a customer feels about a brand in a matter of seconds, which is part of why we're being careful about how and whether we roll it out.
So... will we be making this feature available?
This tech is definitely impressive, but it's got a long way to go in terms of effectiveness, and we're still on the fence about whether it's a net positive for our clients, and by extension, their clients too. That's why we're still in the testing phase. It's just not there yet when it comes to reliably completing tasks, and we're looking into guardrails to protect customers and businesses alike.
Our approach to LLMs is deliberately measured. We're not interested in throwing AI at every part of our product just because we can. We want to focus on the areas where the technology actually makes sense, and where we can deploy it responsibly.
That also means keeping the data where we can control it. Where we're using LLM technology, we're focused on private servers hosted in Canada rather than sending everything off to a third-party provider on foreign infrastructure. Examples include things like voicemail transcriptions and our new text-to-speech voice prompt generation. These are areas where AI can truly make the product better without putting an AI receptionist between our clients and their customers.
We'd love to hear about your experiences with this technology. Whether you've had a memorable AI-agent moment as a customer, or you've rolled it into your own workflow, we're curious about the real world side of this. Let us know!
Every feature you need, for only $25/month (psst... that's in CAD)
Get Started
.png){kind=small}
.png){kind=small}
.png){kind=small}
