Live at the M2 AI Summit

Anyone can nail the first call.
We build for the thousandth.

A voice that sounds good for one conversation is the easy part. Staying reliable across thousands of real, messy ones is the actual work.

What breaks at volume

Real calls don't go to script

The interruption. The special request. The background noise. The "actually, make that two." Every call has a moment no script saw coming, and this is how we handle it.

01

Common requests are easy. Trust is won in the exceptions.

Order a Whopper and fries and the drive-thru AI copes fine. Change your mind halfway, add a voucher and ask to split the payment, and it falls over unless somebody built that path on purpose.

Transcript

So the other thing you're gonna see with AI is that the common request is easy, but trust is won in the exceptions.

So, another fast food example. Not sure if anyone's been to the Burger King drive thru at Curletts Road, but they have an AI that takes your order. So if you go there and you know what you want, Whopper, fries, Coke, it'll handle your request, no problem.

But the minute you start hesitating and you say, oh actually, let me get extra tomato. Um, can you make it a large. Uh, this date's not going very well, so let's split the payment. I've also got a voucher as well, so I'll use it, I'll give you the code.

Unless you've architected these edge cases, the likelihood of that system breaking is very high.

Clip: why your 'extra tomato' order breaks most AI

02

The two real bottlenecks have nothing to do with the voice.

Speed fights intelligence, and a phone line carries worse audio than a demo microphone. Neither one can be wished away, so the design has to route around both.

Transcript

Host: That's cool though. I mean, what are some of the things that can go wrong in a case like that? Cause that's really handy if you're like a plumbing company, right? Then you can just have this Chloe on call all the time.

Owen: Yeah. So I guess there's two things at the moment with voice AI that is the big bottlenecks, that I don't think anyone's been able to answer.

One is balancing latency with intelligence. So the faster you want it, the less smart it's gonna get. But the smarter you want it, the less, you know, response time you're gonna get.

And then the other one is, you know, audio quality. So again, if there's interference, between, you know, people talking, there's like technology, it might misinterpret what you say, and then that affects the speech to text, and then, you know, it causes these loops. So those are the two major, I guess, bottlenecks with voice AI.

And I think the models are gonna get better and better, but as of now it's just about live data and seeing how you can streamline the process as much as you can. Like for example, instead of asking me what's your number, and I have to say 021 whatever whatever, and it has to repeat it, it just says, is the number you're calling from the best number? Cause 90% of the time someone calling, that is their best number. And then on the back end in the system, it'll take that deterministic number and use that as the data. So it makes it way more reliable and less likely to break.

Clip: voice AI's two biggest bottlenecks

03

The hard part was never the talking.

You build the common path and it works. Then you hand it to real people, and their accents, their speed and the order they say things in open a new set of edge cases every week.

Transcript

The unscripted edge cases is where you're gonna spend like 80 to 90% of your time.

Cause when you build it, you have the common path in mind, right? You know what outcome you want it to achieve. You build it, you test it, great, it works. Then you try to harden it with edge cases, cause you know the weaknesses of the agent. You build it, cool.

As soon as you give it to someone else, the way they speak, the speed, their accent, the order they say things in, it just opens up a Pandora box of different edge cases.

So for that reason, having live data is really important. But you also need a system that works in the background, that analyses every single call, and then it iterates and improves after each call.

Clip: the hardest part of voice AI

04

You don't get to pick the conditions of a real call.

Job sites, moving cars, noisy kitchens, and half of them on speaker. The agent has to understand the call anyway, which is why a clean demo proves so little.

Transcript

That is exactly why I have a personal vendetta against YouTube demos. When we first started demos, mate, it's like you always have the clearest audio, one clean question, there's no consequences.

But as soon as you start building it, you have accents, you have interruptions, you have the difference between a phone call and a web app. The frequencies they use is completely different, so the audio quality drops on a phone call. It messes with the speech to text.

And then you also have tool failures, if the app that you're using isn't in sync because it's just updated.

Clip: the reality of voice AI demos

A voice is only the front door

Trusting it to handle your line requires clarity on the rules behind it: it quotes what you've set, books only what you can actually deliver, and never tells a customer something you didn't approve.

Interface

How the customer reaches you. A phone call, an email, a website form, a chat or an SMS. Whichever one they use, the same rules run behind it.

Decision layer

The model that reads the request against your rules. What you quote, what you can actually deliver, and what it must never say.

Action layer

The clicking and typing that used to happen at your desk afterwards. The booking written, the job created, the office told.

The three layers of AI agents

How it talks, how it decides, where the work actually runs. The map behind a real agent.

Transcript

The three layers of AI agents. We already talked about the front door, the interface, you know, the voice. Obviously you can have it for chat and SMS.

Um, the decision layer. Your large language model, aka LLM, is like your GPT 5.5, your Mythos, your DeepSeek, Grok. These are all just the brain of the decision layer. These models are all trained on different knowledge, instructions and rulesets. If you want it to be trained on your business, then you would need to use these models, to then give it the data, to then essentially create your own brain.

The action layer is when you're sitting at your desk, you're typing, you're clicking whatever apps you guys use. All of that can be automated.

So when you combine the action layer and the decision layer, as well as an interface, then that's when people actually start, you know, being scared about losing their jobs.

You are the architect, the AI is the technician

With a new AI tool released every other day, the highest-return skill is explaining your own process plainly enough for a five-year-old to follow it.

Transcript

So with that understanding and the difference in mind, the best thing you can do tomorrow when you're sitting at your desk is picture you have an apprentice that never sleeps, never forgets.

Think about the one single workflow or task that you want to automate. How are you gonna explain to that apprentice, step by step, what you want it to do in order to achieve that task?

And then the clearer that you can conceptualize that, the closer you're gonna get to the reality of actually being able to fully automate that task.

Where it started

Before ChatGPT had a voice

Jess spent a whole day ringing plumbers and got voicemail, or a promise that somebody would come and look before quoting anything. That afternoon is the reason this company exists. Read the full story.

Transcript

One day Jess, she's working for this startup, and she needed to call a plumber. Um, so she spent the entire day dialing, and then the calls would either all go to voicemail, or they'd say the same thing, which is basically, well, we need to come out and see it before we give you a quote.

And she told me this, and I was like, man, this is, it's 2025, like, we have AI, we have all this technology. I personally hate being on hold. I hate voicemail. There has to be a more efficient way.

And at the time, I don't think ChatGPT even had voice yet, but we knew it was possible. So we started researching, we started building. Um, and that's how we got started with voice AI.

Built for the thousandth call.

Tired of the phone letting jobs walk? Let's talk about what a reliable agent looks like for you.

Book a quick call