How to build AI voice assistants without code

Amie @ Ben's Bites (00:06)
Alright, and now we are live.

Alright, if anyone is in the audience, if you could drop a quick hello in the chat so we know that you're there, that would be super helpful. there we go. Thanks, Mike. Alright, hello. Hello, hello. Okay, alright, looks like people have joined us.

Tom Osman (00:31)
in chat.

could see it.

Amie @ Ben's Bites (00:36)
Welcome. All right, I know you're probably expecting to see a Ben space here, but I am filling in all on my own this time. If you joined us for past workshops, I've been helping co-hosts over the past few weeks. And now I'm on my own to host an OG Tom Osmond. Tom and I go way back to the MakerPad days, which is Ben's previous company. And Tom's kind of

led the way for me in learning NoCo technology. He also led the way of me learning all about crypto. And today he's going to lead the way and teaching me all about AI voice agents or assistants. So Tom, I'll hand it over to you.

Tom Osman (01:21)
Let me just try and change my camera because for some reason Riverside isn't enjoying it too much, with.

Amie @ Ben's Bites (01:31)
No worries.

Tom Osman (01:35)
Okay, we're back. Yeah, that's weird. Hold on, let me just try and the power off.

Amie @ Ben's Bites (01:36)
All right.

No worries.

Tom Osman (01:42)
Okay, so, I hope you can hear me okay. Is anybody in the chat played around with any voice AI so far? So, quick intro. My name's Tom. I work on the growth slash education team over at Synthflow AI and we basically have a product or platform where you can create AI voice agents without needing to write code. And these assistants, as you'll see in a minute, or agents, they can make or take phone calls.

They can act as widgets and embeddable things on your website or app. And basically, I mean, you can call it what it is, replace the humans so they don't have to make calls anymore. So all those tedious jobs that nobody wants to do, like call centers or these transactional phone calls, they can now be handled with AI amongst other things. But that's the basic use case. Synthflow is a German startup raised...

Quite a bit of money so far, we're a team of about 25, moving quickly. That whole voice space is pretty competitive and a big opportunity, I think, for anybody building in AI. Still really early, so a prime time to get used to building some stuff, I think. Amy, have you played around with any voice AI yet or?

Amie @ Ben's Bites (03:04)
I haven't done voice AI. I've been exploring the AI agent space in terms of text. But I feel like this would be an even more complicated frontier. So I'm curious to see what goes into it.

Tom Osman (03:12)
Okay.

Okay, so if my screen freezes, if my camera freezes by, let me know and can turn on.

So here's what we're going to cover today. We're going to go, the hype? Quickly introduction, Synthflow, some basic functionality, some more advanced stuff, assistant examples, integration options, workflows, HubSpot app, Q and A. That's basically it. We're going to rip through it. So why the hype of the voice AI and why is it not shit basically? So if you followed the Amazon story, everyone's probably used Siri and it was okay, but it was a bit of shit.

And I think they spent around 10 billion, I think on it was the reported number and it just doesn't work very well. So there's a few reasons why it's good. Now we've had the transcription accuracy is way better. The latency is way down. We have human sounding voices with 11 labs amongst other things like function calling combined with these lot.

and you can now create really accurate real-time speech voice agents powered by AI. So that's where we're at. This is why there's a lot of hype. Here's just a taste of some of the people building in this space at the moment. We have lots of vertical people building for different industries. We have the speech to text folks. We have all the nice middleware companies. The horizontal applications is kind of where we come in right down here.

The model providers at bottom like 11 Labs and Deepgram who do all the transcription models. And then the big tech like Apple Meta, Google, OpenAI, et cetera. So big opportunity for anybody building either vertically or horizontally right now. It's pretty vicious space, but lots of, lots of opportunity. So we're going to, I'm just going to share my other screen now, and we're going to run you through the actual

that form. Let me see if it's here. So when you log in to synthflow, let me just actually bring it to the home page so you can see what it is. This is us. Effortless human-like AI phone calls. Build a no-code AI phone call system with our voice agent. Stop missing calls and start converting more leads. So our aim is basically to make it as simple as possible for people to create AI voice agents.

and have it do stuff for your company. So it could be as simple as having an AI receptionist take a phone call for you and take a message, or even forward that call on to a human if it fits like a certain set of parameters. You can basically tell the agent what you want it to do and it can do it. It's basically like a programmatic receptionist is a good way to think of it. So what can you do in Synthflow? We've got some examples set up here.

but if we were to go and maybe create a assistant from scratch, we can do either from scratch, a quick assistant setup or browse a template. I'm just gonna select a template so you can see exactly what this can do. So for this real estate one, we have a lead qualification buyer assistant for real estate. So if we create this one,

We also, straight away we have an agent created, we have a prompt. I'll show you the prompt first and then we'll get into some of the configurations. And this is basically what your system.

In our prompt, presume most people or anyone listening to this has prompted an AI to do stuff. you've chatted to chat2pt, you've written a prompt and it will do the thing for you. And the better the prompt you give it, the more accurate information or the better result you're to get. So with a voice AI assistant, you're just going to give it more information to get the result you want. For this, we're going to give.

a range of different things. We're going to give it the background info. if you've used ChatGBT, could say you are a business analyst, for instance. Here we say, here's our company info. Here is our target audience. Here is our value proposition. And then here is some agent information. So you're breaking down exactly what you want it to do. We're going to give it the name, the role.

the objective. So in this particular one, this is a lead qualification assistant for a real estate company. So we want this person to collect essential information from essential buyers or renters assisting in pinpointing their ideal home based on specific needs or preferences. We give it some objection handling. So some people might be worried about timing, budget, preferences, working with other agents, and you can basically customize that exactly how you want.

Next up, and probably the most important thing is the call script. The call script basically gives the agent a guide on what it is to do through the call. So we have an introduction, greet and identify the purpose of the call, ascertain if the potential client is looking to rent or buy, then go through the location, the specifications, the amenities, constraints, budget, et cetera, et cetera.

And then at the bottom here, can give it any additional requirements. This is a pretty basic script without any additional actions or variables or anything complicated. Does that make sense to start with?

Amie @ Ben's Bites (09:03)
Yes, so far so good.

Tom Osman (09:06)
Okay, cool. So from here, you can make things a bit more interesting. So let's move to something that we have called actions. So up here we have our custom greeting. You can see if anybody's used things like Zapier or CRM before, you can bring custom variables into this too.

For example, hello, I'm calling from majestic states. I've seen you could be a formal website about buying or renting. Do you have a couple of minutes to finish a request? Now, if we were to test this, we can say, let's say, I've seen you could be a formal website about buying or renting. We can say this. Hey.

and then lead underscore name.

Press save. Or you just click off and it saves. You can see on the left here, on the right here, we have lead name. If I put in Tom, and I put my number in.

seconds. We could get this to call me, I might have to move to a different account because I may have actually just maxed out my assistance in this account.

two, five, six, okay the number might be helpful. If that then calls me we will see that this lead name will come through on the call. So let's just give it a try.

Okay, so the phone call's coming through. Let's answer it.

Okay, so that brings through the variable. Now that is a very basic, that's probably our first assistant that we made a long time ago. To make this better, you can go and do things like upgrade our engine to a version 2 which has just come out. You can then get into the voice, so you can fully customize the voice with 11 Labs, say...

Amie @ Ben's Bites (11:09)
Thanks.

Tom Osman (11:34)
You didn't want straight pull to make the call. You can include maybe a bit more an expressive voice. Maybe you have like a Spanish assistant, female. You can just go through and choose the names. You can listen to them and then choose that voice. If you wanted to have the style exaggerated, for instance, make it less flat. You can do that. You can increase the voice speed, the fade outs, the latency, all the good nitty gritty stuff that can perfect that voice.

Amie @ Ben's Bites (12:35)
Yeah, I'm about to ask, can someone ask, can everyone see Tom or his screen? It's, alright, let's see. It looks like we lost Tom. Give me a second here.

I have a message.

Great.

Alright, welcome back Tom. Yeah, you're back.

Tom Osman (13:20)
Okay, is that back in? Can everyone see me again? Okay, if I freeze, just shout and I'll immediately stop and check it. I don't know. I feel okay.

Amie @ Ben's Bites (13:28)
Yeah, I think it started kind of going sideways for me when you're going through like the different settings you can do for the voice.

Tom Osman (13:40)
Got it, okay. So I can just jump back in and let me just see if we can share my screen again.

you

Okay, we're back.

Amie @ Ben's Bites (13:58)
All right, yeah, I can see your screen. Cool.

Tom Osman (14:00)
Okay, nice. So here are the different actions. You can think of an action basically as a thing that your assistant can do programmatically without you having to do anything. things like sending a simple SMS after the call, containing some information about an appointment or just thanking them for taking the call. A real time booking to say schedule a live appointment during the call. We're to move to another assistant to show you how that works. A call transfer.

This means that you can transfer the call live to a human should the script determine that that condition has been met. So for instance, you could give your assistant a script to say, if the person starts getting annoyed or they specifically say they want to speak to a human, you can then initiate a call transfer. So here is, let me just select a new action and show you how

because a lot of the voice agents are getting really good, but some people just want to talk to a human or if they realize they're talking to an AI, some people still get slightly annoyed. So they can actually request that number to be live transferred. You can actually specify when you want this transfer to take place. You can now add some nice classical background music, a message while that's being transferred, navigate an extension.

And also give messages when you're trying to reach the target, as in, me transfer you to a human, when the call has been answered and if the answer doesn't or the person you're calling doesn't pick up. That's pretty neat. Also, we have human availability. So if you're making AI phone calls through the week, you don't want them to forward them over a weekend. You can set the working hours and the time zone.

when that call should be forwarded and when it shouldn't be. So that's pretty cool. One of the, I'd say probably the most important things with voice AI assistance in general is it's all good to have a conversation. It's all good to get your assistant to make a phone call, but what happens with the information, what happens with the conversation afterwards? So this is where things like information extractors come

If you use an information extractor here, what this will do is you could specify, for instance, is this user looking to buy or sell the house? So let's say in the market, just as like a real estate example, and you say, the user, is the user looking to buy or sell a house?

question mark, then the AI system will keep this in mind when having the conversation. And at the end of the call, it will determine.

Yes or no to this answer. Are they in the market and looking? Are they looking to buy or sell? Yes or no. Simple. And then you can push that automatically into a CRM, into an air table, send an email, update, set a task for somebody to give them a call back. Whatever you want to do with that. Let's say, let's create another one. We can do a single choice. Just to build on that one. We can say, buy.

cell. They're to buy.

So you can say,

So.

You save and attach that. If you ask a conversation, are you looking to buy a house or are you looking to sell a house? You'll pull that information out programmatically again, you can update a CRM. So let's say you're making a hundred phone calls all at the same time to a hundred leads. It could be cold, warm, or otherwise. You could then programmatically pull out every single answer through the script into a CRM or trigger an email, put that wherever you want.

Does that make sense? Like it just a general concept for making a call and extracting info.

Amie @ Ben's Bites (18:25)
Yeah, looks pretty straightforward.

Tom Osman (18:29)
Pre-shake it. Okay.

Amie @ Ben's Bites (18:30)
Which is great. I also love that it's like the interface is so easy to use. I think I have PTSD from anything phone related from the Twilio tutorial days. and that just being like, just like super technical. so I love how easy this is to already like test out, you know, the workflow.

Tom Osman (18:50)
100%. So do you know what what Synthflow is built on, Amy?

Amie @ Ben's Bites (18:55)
I don't, tell me.

Tom Osman (18:57)
bubble.

Amie @ Ben's Bites (18:58)
No way. I guess now that I'm looking at it, it kind of looks familiar. But, nice.

Tom Osman (19:04)
Synthflow is built on Bubble. The back end not, but the front end is. So I think it must be probably one of the most successful ever Bubble apps, I think. It's pretty insane. So.

Amie @ Ben's Bites (19:09)
Thanks.

Yeah, that is something. Great use case.

Tom Osman (19:21)
Let's now look into another assistant use case. So I have a restaurant booking assistant here called Bianca and Bianca works for a restaurant called The Grill and her job is to take reservations. So if we look in the prompt, we can see that this is getting a little bit more structured and there's a few more variables in here. So let's take a look now and see how you can kind of customize them.

tweak your assistant. So we have a lot of information here about the restaurant itself called the grill. The more information you give it, the better it works. The assistant note is our restaurant uses an AI powered voice assistant to handle reservations efficiently. And then we have the call script. So step one, greeting and purpose, reservation request. Step three is date and time collection. And here we give it some specifics. We say ask for a...

preferred date and time for the reservation, let them know that available seating times are 7.30, 9 or 10.30, Monday to Saturday. Save the reservation date in this format. If the user responds with an answer like today, tomorrow or Friday, for instance, convert that into a convert that day into a date. So it's getting a little bit more technical, but this works better for things like for bookings. So if you're pushing a booking into a calendar app, for instance,

These things are little quality of life tweaks that you can make sure your assistant does exactly what you want. Next, we have name collection and seating preference. Now this is where this assistant is a bit unique. So this hypothetical restaurant has an inside and an outside seating area. And one of the things I think makes AI agents or voice agents really powerful is because they can do things faster and

in a way that humans or things also they have functionality that humans can't do. So we have a weather-based recommendation part to this app and we've created an API call, which happens live on the call. show you how that works in a second where according to the date that the people ask for their table, they will ask if they want to sit inside or outside, and then it will do a lookup.

It will hit an API endpoint, see the forecast in that area for that specific day and time. It will bring back the weather and then it will say if it's hot, cold, windy, wet, and according to that result, recommend that they should sit inside or outside. So if they request outside, but it's going to be cold and rainy, the assistant will say, it looks like it's going to be raining that day. I recommend you sitting inside.

This happens like in a split second and it's something that a human likely wouldn't do. So that's pretty cool.

Amie @ Ben's Bites (22:27)
It's cool because you can kind of make the hospitality element kind of consistent that way.

Tom Osman (22:35)
100%. And with this, I can show you how that works. So this is actually called a custom action. So if you see these little things in here, these are called custom actions. So let me just jump out of this in a sec. I'll just do this additional guideline. So at the end of every assistant, we like to give them some additional guidelines and voice rules. So to make this smoother,

we say seamlessly incorporate natural voice inflections like, okay, uh-huh, et cetera, and a bunch of other quality of life things. So for more conversational bots or assistants or agents, we include some human expressions that we want them to use to have like a natural conversation. Just a few additional tips there. Looking at the action, we have a custom action. So this is basically what it looks like.

you can connect any API and do anything in terms of a get, post, put, or patch to bring data inside your live voice assistant. So let's say you wanted to do a patient lookup or an order lookup in a Shopify store, the user will provide maybe an order number, then you can hit the API with that order number, bring the status of the order inside of your response, and then relay that instantly back to the user.

Super good. Let's see if we can initialize this and give this a test. Let's say 11 and we're going to get the forecast for London on the 11th.

Amie @ Ben's Bites (24:11)
I have a feeling it's going to be raining in Overcast.

Tom Osman (24:14)
I think you might be right.

Okay, so that worked and we can see the data that we get back on the forecast. We're using the forecast in this prompt is, let's see the conditions.

you

Yeah, we have the temperature, average temperature of 41 degrees Fahrenheit. That's probably a bit cold to outside, isn't it?

Amie @ Ben's Bites (24:51)
For me.

Tom Osman (24:53)
bit of heat. Yeah, I'm partly cloudy. let's see what day did we say? We said the 20th?

20th. Okay, let's go and test this now and see if works.

Again, let's try this.

Change the number.

and the door call.

I reckon to book a table if possible.

people will be joining you for the reservation? It's going to be six of us.

Table for six.

We have seating times at 7, 30 p.m., 9 p.m., and 10, 30 p.m. from Monday to Saturday. I'd like to book for the 10th, sorry, the 20th of November at 9 p.m., if possible.

No problem. Let's talk in November 20th at 9pm for you. May I have your name for the reservation? Yes, Tom.

Do you have any seating preferences, like indoor or outdoor? Outdoor would be good.

Yeah, let's do that.

seem like the

Okay, let's just quit that demo. So that didn't actually work. Can I have a check? That was just a little bit, but it's almost there. 100%.

Amie @ Ben's Bites (27:09)
You know what? Every single workshop there's always one thing that goes wrong. So it was bound to happen. But I'm impressed though. It handles it like very well, very human-like, probably more courteous than most humans. And it's kind of delightful actually.

Tom Osman (27:23)
Yeah, a hundred percent. It's kind of delightful. Yeah. So that's basically, and you can tie in real time bookings with cal.com. So this could hook directly into a calendar. So it make bookings and check for live availability or book that straight in with no problems. There's also one other thing. Go on. Sure.

Amie @ Ben's Bites (27:28)
Yeah.

We have a, can I ask a quick question? And this might be kind of outside the usual use case. don't know. So when, let's say let's go like clear down this whole reservation train of thought. Like if most restaurants are using like software, like open table or toast, or I don't know what the equivalent is in, in Europe, but those are the same ones. don't know. How would you integrate with those systems? Like, is it an open?

Tom Osman (28:00)
Mm-hmm.

Amie @ Ben's Bites (28:15)
API with.

Tom Osman (28:18)
Yeah, so you can do it in a couple of ways. you, so a really simple one is we get requests from Microsoft teams, for instance. And at the moment you can either do a custom integration yourself using the API, using a custom action, or you can use, think cal.com has a load of integrations in their app store. So if you connect, seems pretty, actually a little bit hacky, but you can connect a

a Microsoft Teams account to cal.com and then we can use our native integration with cal.com and it can check all those calendars here at the same time. So if you have your Google calendar, your 365, et cetera, connected into your cal.com account, you can check that just through our real-time booking on the assistant. With OpenTable, we haven't got the integration yet, but we're working on a lot of new ones. You can...

Amie @ Ben's Bites (29:00)
Okay.

Tom Osman (29:19)
don't know if they have a lookup through Zapier or Make yet, but you probably have to use the custom action, like a brand new custom action here to do that.

Amie @ Ben's Bites (29:25)
Yeah.

Cool. And then we have another question about real-time booking. Someone's kind of wanting to explore that more and what the differences are between like an inbound, outbound call agent.

Tom Osman (29:43)
So real-time bookings, we have two native integrations at the moment. Cal.com is you basically put in your event ID and it happens out of the box for you. So you don't have to worry about any configuration. Same with Go High Level. The difference between inbound and outbound assistance, you can just see on the left here, is let's just take this inbound one. So this...

inbound assistant is deployed to a phone number. So we have a Twilio account where we bring in a phone number and then we deploy the assistant to this phone number. This means that at any time, anybody can dial this phone number and speak to this specific assistant. So if you want a set inbound receptionist, for instance, or a salesperson, or even an assistant that just routes calls to different parts of your business, you're to want to use an inbound assistant set to one.

particular phone number. With the outbound assistant, it will also just tweak the, like how you, how you prompt it, right? So an outbound assistant is going to be different to an inbound one. With the outbound, you can deploy this to something called a batch campaign. Now we can touch on this. We have a lot of lessons in the academy of how to do these, but with a batch campaign, you can upload.

a list of contacts, or let's say if you've already uploaded some, say you wanted to add three people to a batch campaign, can say I want to call these people starting at 6.45 tomorrow and ending at 9.15. Time zone is, let's pick around a mile, there you go, Fort Nelson. Select the phone number that you want to use for your outbound campaign.

and then you can press create that campaign and then your outbound assistant will make those calls and bring all that information back into Synthflow or push it into the CRM view. So slightly different use cases. There is actually a third use case that I wanted to show you today and that is our new widget. Now our widget is powered by OpenAI's real time voice assistant.

So you can see here.

This is like a fully embeddable assistant that you can put on your website or app that you can call at any time powered by OpenAIR's real-time voice or Synthflow that can do all of the exact same things that an inbound or outbound caller could do. So people don't have to call the number. If you wanted to maybe convert people or engage people on your site, or even have them make reservations or book appointments, you can do that with our widget here. I don't know if you can...

hear that but I can potentially share an assistant with you. Yeah I don't think we do, does does Riverside share system audio? Pretty sure it doesn't right?

Amie @ Ben's Bites (33:04)
Let's see.

I don't know.

Tom Osman (33:08)
Now this is going to be, that's okay. This is going to be pretty much the same configuration as your general assistance. can come complete with actions, with real-time bookings, with custom actions to do lookups, but this can sit on your site that people can just click it and have a call within the system pretty much. So perfect for things like customer support or bookings or reservations. Insurance leads are always a popular one too. So yeah, that's our...

That's our widget.

Amie @ Ben's Bites (33:40)
Marco asked if your HubSpot integration includes integration to HubSpot meetings.

Tom Osman (33:48)
No, not yet, not yet, but it will be. So the guys have just launched the second version of our HubSpot app. Let me just actually, I think I can.

Amie @ Ben's Bites (33:50)
Okay.

I think you're frozen again. Your video is anyway, for me.

Tom Osman (34:07)
turn it off for minute and see.

So, let me just reshare.

Amie @ Ben's Bites (34:20)
I

We'll wait for Tom to come back.

There he is. And he's back.

Tom Osman (34:33)
God, what is going on with Riverside? Okay. Sorry, folks. Let me share this with you. So this is our HubSpot app. This is probably maybe for the slightly more technical amongst us. But this, I think is where Hubs or where you kind of get into the big business small team sort of thing.

Amie @ Ben's Bites (34:35)
It's okay, I know.

Peace. There we go.

Tom Osman (35:02)
Sam Altman was saying, we're saying that the one person billion dollar company, isn't quite that, but you can pretty much build out whole teams making calls to your entire CRM with this app now. So for instance, if you set up a trigger inside of HubSpot, this is HubSpot workflows, by the way, if anybody hasn't seen this, I can go back so you can see you have workflows inside of HubSpot and you have.

sorry, automations and workflows, created a workflow. You have a trigger. This can be something like a, when a web hook is received from another app, when an event occurs, like somebody's filled in a form or when a filter criteria is met. And this filter, I think is going to be one really powerful one for anybody using HubSpot. It could be when somebody hasn't been spoken to in 30 days, let's say, sticking on the real estate use case.

If a lead or an old lead hasn't been contacted in 30 days, goes into this filter, goes into this trigger, which you can trigger a phone call with a specific model. Let's say Alan AI specializes in engaging homeowners. Again, you just create these in Synthflow and bring them across. You bring in the contact's phone number, the time zone and the prompt. This is actually...

a really interesting one because if you bring the prompt from Synthflow into HubSpot, let me just see, I don't have it open, but the prompt that we wrote in Synthflow where it has the script and the information about the system, if you copy that and bring that in here and delete it in Synthflow, this one will override it. And this means that you can use all of your variables

from HubSpot that anything you know about this person, obviously there's a lot of data points in here, inside the script. So you can use, for instance, the current employer, notice period, date of birth, desired benefits, previous conversations with anybody. You can use this programmatically inside of the script for your agent. So it makes a phone call. It could say, just giving you a call today, notice we haven't spoken to you in a while. I know you spoke with Jake.

20 days ago, I just wanted to check in, if you're still buying or buying, selling at home.

Same with the greeting. can use any data that you have inside a HubSpot inside your greeting. And then the good part is you can do branches. So it's all well and good making a single phone call. But if you have a lead list, you're going to want to follow up. You're to want to do check-ins. You might want to send emails or SMS's, WhatsApp's, add them to Facebook audiences, all that crazy stuff. You can do that direct from HubSpot. So.

For instance, have a trigger gets triggered, make a phone call, delay for five minutes to make sure the phone call is finished, and then branch based off the call status. So if the call is completed, if it was busy, if it was failed, hung up, you might want to do another action. So for instance, you could do this. You could say if it was busy, you go delay.

You could say set amount of time, say 30 minutes.

And then you could copy this one.

place there. So if they didn't pick up, wait 30 minutes, make another phone call. And then you can again, carry on branching or if it was completed, you might do something like send an email or send an internal notification to a team member with the details of that conversation. And you can see how you can start building out these big internal workflows inside the HubSpot to make programmatic voice calls without the need to have

any like enterprise grade software pretty much.

Amie @ Ben's Bites (39:20)
So is, does it, mean, I assume maybe you have to set up an action for it, but how does it work with like, does it put a transcript attached to the contact? Does it put a summary of like any interaction? Do you know like what that looks like?

Tom Osman (39:35)
I Let me see. This is a test account, so I use some rogue conversations on here. But they are, let me see if I can find myself.

Amie @ Ben's Bites (39:36)
Mm-hmm.

Yeah.

We can't show, that's fine if you just tell.

Tom Osman (39:50)
And.

Yeah, so what happens is...

When the call is completed or when any call is made, all of that activity gets logged inside of your app. So you can see the conversation with the script. You can do the call type. So it comes out, the call outcome was connected. The call direction was inbound. This was the person you contacted. And this was the script. It will update the route. This might be an old call.

They updates the recording so you can actually play the recording directly in HubSpot. And back to the conversation, we're actually also working on the live integration too with HubSpot meetings. So all that can happen inside HubSpot. This on the left-hand side, actually, this is an important one. You can sync your information extractors with HubSpot fields.

Amie @ Ben's Bites (40:40)
Well.

Tom Osman (40:56)
maybe a bit more like down the rabbit hole for HubSpot users. But if you created these fields inside of Synthflow as information extractors, so this is a recruitment call. want to see them. Are you in the market? What are you looking for? What's your notice period? You want to extract those information and we can sync them with these fields. So when the call gets made, the information gets extracted and it automatically updates the CRM without having to do anything.

Amie @ Ben's Bites (41:02)
Pretty cool.

Tom Osman (41:31)
So you can kind of do the job of 200 recruiters pretty easily with this.

Amie @ Ben's Bites (41:36)
Yeah, this is pretty crazy. Like I'm just thinking of like your business development team suddenly became way smaller and way more cost effective with this tool.

Tom Osman (41:47)
100%. We're seeing a lot of these, like AI powered SDR teams, think, 11 X. don't know if you're following them. They're a big team. think they've just raised 50 million to do these AI powered lead generation bots, which combine voice and emailing and texts and followups and stuff. So yeah, it's getting pretty crazy what you can do.

Amie @ Ben's Bites (42:10)
Yeah, absolutely. And you said there's a whole academy, right, that you worked on building out the videos yourself, right?

Tom Osman (42:13)
So what else would you like to see?

I did. So we actually have the Synthflow Academy. Let me see if I can bring it up. It's academy.synthflow.ai. So...

We have three courses. me just see if I can share again. Yes, info.academy.ai is...

It's here, we have our basics tutorials. We've got, I think, 1100 students through the basics. We have the intermediate, advanced, full course on integrations and a new agency course dropping soon. That's actually one thing that I didn't touch on is if you have any, do you have any like marketers or digital marketing agency people in Ben's Bytes? So I guess probably so.

Amie @ Ben's Bites (43:21)
Yes.

Tom Osman (43:23)
Okay, nice. So we can, you can white label the whole of the platform. So everything that you saw me demo this time around, you can put your logo on it and just brand it yourself. So we handle all the payments, all the usage charges, all the assistance. So you can actually just build and assign assistance to clients and then give them custom pricing plans so they can sign up.

all fully branded to your business without any mention of SIDs. So digital agencies can buy and create and resell these pretty damn easily.

Amie @ Ben's Bites (43:57)
Thanks. Is there any, and if you can't share that's fine. there any like use case that you're seeing that like you're surprised by?

Tom Osman (44:11)
probably insurance actually has probably been the biggest one, the biggest in demand because there's a, there's a lot of like most of the entry level jobs in insurance is just taking phone calls and giving quotes. And most of the quotes fall between a certain set of parameters based off of like age range, location, health issues. It's usually the same sort of script that's being read out every single time.

Amie @ Ben's Bites (44:16)
Yeah, that means yeah.

Tom Osman (44:38)
And you don't have to be like a highly qualified individual to be able to have that conversation. So we're seeing big insurance companies suddenly twig on that they can just do this first call programmatically, update their CRM, automatically filter by the results, and then have maybe a more senior member of the team follow up if that person fits like a certain criteria, basically. And obviously that brings their costs way down.

improves all the margins, et cetera.

Amie @ Ben's Bites (45:09)
Yeah, incredible. We have two questions. What's the difference between a synth flow workflow and a list within the prompt?

Tom Osman (45:22)
Synthflow workflow. Do you mean... Let me just see if I can bring it back up. So the workflow actually just being like the automation, I guess. I will show you.

So anything that's in the prompt is going to be like a set of fixed, it's going to be like a fixed piece of information. It's going to be like a bit of a system prompt. your, your script, your additional information, your guidelines of how you want the conversation to be held is going to be things like your company information and anything static.

Amie @ Ben's Bites (46:06)
video. We lost you, but we can hear you.

Tom Osman (46:10)
Okay, I'll leave it off, I was trying to rejoin the set. So the workflows are going to be what happens either before the or after the call. So you could create a workflow that either triggers the call to be made or happens after the call is completed. So for instance, if you wanted to email a team member after the call has been done or update an Airtable record or update HubSpot or send a Slack message, you can do that with a workflow.

Amie @ Ben's Bites (46:12)
Okay.

Tom Osman (46:40)
If you want to trigger a call from an external app, you can use a workflow for that. So when a field changes, for instance, make a call, you use workflows. Anything static as in the prompt, anything dynamic, you build workflows and custom actions.

Amie @ Ben's Bites (46:58)
Got it. And are there any specific company? we lost them.

I will ask if there's any examples anyone can test out live. I'm sure they have like, here it is.

Tom Osman (47:11)
Sorry again.

Amie @ Ben's Bites (47:12)
No worries. I'm sure you have demos available within an account for free, but are there any specific companies that the audience members could call to test out the voice bots live?

Tom Osman (47:26)
So you can sign up for free account with Synthflow and it gives you, I think 15 minutes free.

the entry level account is, just to check it hasn't changed, it's the last checked. But also we can get you, I think the starter plan is $29 a month and it gets you 50 minutes, concurrent calls, unlimited assistance, the actions and the integrations. And then it kind of goes up from there. But I can get, if anybody's interested, a 50 % off coupon for anybody on today. So just email me.

Tom at synthflow.ai and I'll get that too so you can get that even cheaper.

Amie @ Ben's Bites (48:12)
Great. Awesome.

I just lost my train of thought.

Tom Osman (48:22)
You can build some pretty cool stuff. has additional things it has. You can do multi-language assistance. So any voice that you can find on 11 Labs, like Spanish, Portuguese, Greek, Russian, I think as well. You can do multi-language assistance. They can change language as speaking while on the call. So if somebody can't speak English, I know living in Spain now, one of the most annoying things when you bring up and with my Spanglish.

Amie @ Ben's Bites (48:22)
Welcome back.

Yeah.

Yeah.

Tom Osman (48:50)
You try and speak to someone in English, but this assistant can speak multiple languages at the same time. So if you're using like 11 labs, multilingual V2 voice, for instance, and you import it, then that assistant can just swap voices mid prompt and also, or mid conversation, sorry. And the prompt can just stay in English. So you don't have to write the prompt in Spanish to have your assistant speak Spanish. You just write everything in English and the assistant can keep changing the language.

is needed.

Amie @ Ben's Bites (49:20)
Thanks. And is there like one model that you're using or can you kind of

Can you select your own model for like how the AI responds?

Tom Osman (49:28)
You can't select your own model at the moment. We have like a Synthflow LLM, which is just optimized for like real-time conversations. And we've just started rolling out and we have our GVT 4.0 is probably the most reliable music at the moment. And we are slowly rolling out real-time voice. So the real-time voice API is now available on widget based assistance, but we're slowly rolling out to inbound and outbound too. So that should be live pretty soon.

Amie @ Ben's Bites (49:58)
Is there any... Okay. You want like an uppity assistant versus a nice one. Nice gatekeeper mode.

Tom Osman (49:58)
And that brings in all the emotion and everything.

Yeah, exactly. 100%. Yeah, so for language assistant, would you write the prompt in English or localize? You just write, you can write everything in English. If you did want to do it in Spanish, for instance, you can do, but it works just as well in English and then it would just hold the conversation in Spanish.

Amie @ Ben's Bites (50:26)
Nice, very cool. Is there anything we should get excited about that you guys are working on that you can share?

Tom Osman (50:34)
yeah, the real time voice is a big one. real time voice with custom actions. at the moment, open eyes, open eyes, real time voice API, the one that sounds really good, sounds really human-like has no latency at the moment. can't take actions. So where you saw the weather being looked up live on the call, or if you were to look up like an order ID from Spotify or Shopify, for instance.

You at the moment can't do that using that model because it just doesn't allow for function calling, but it will do soon. So we're getting ready for that. So when that comes, we can flick everything live and then you're going to have like a real human like sounding voice assistant, powerbo, open AI that does amazing function calling. So get all your API endpoints ready and get all your integrations made. And then once that's live, we can just flick over, flick over the model and update every assist.

Amie @ Ben's Bites (51:33)
And when you're building out these prompts for the assistants, are you finding that like the best prompts are quite long versus short? I just did a prompting workshop last week and I think the most, like the biggest takeaway for people is that like you need to put time into writing a decent prompt if you want decent results. Do you find that to be a use case or the same, do you find that to be true with the prompts in Senseflow?

Tom Osman (51:51)
Hmm.

Yeah, a hundred percent. think using the best way still is to use, I think Claude to write really in depth prompts for you. So you know that there was that meme like make it better, make it better, make it better. And it just gets better or more exaggerated. Just fine doing that. Just be as thorough as possible. I think there's a really long, like multi hundred thousand token context length in Synthflow at the moment. So put all of your, all your company info in the prompt.

Amie @ Ben's Bites (52:13)
Yeah.

Tom Osman (52:29)
put all of your additional information. It's really good. It's really fast. So the more you can give it, better and more accurate it's going to be.

Amie @ Ben's Bites (52:37)
Well, thank you so much for your time. And I'm now bullish on voice agents. So I didn't realize like they were, they've come a long way from the beginning days of AI, that's for sure.

Tom Osman (52:54)
Yeah, they've come a way. think when people realize that you can do like live API calls in a fraction or like in a couple of milliseconds live on the call, then that kind of opens up like a whole, a whole new world. So yeah, it's a pretty good time to be learning and people are paying a lot of money because if you, if you are in AI to build valuable things for a company, then

Amie @ Ben's Bites (53:05)
Totally.

Yeah.

Tom Osman (53:20)
These assistants can make 50, 100, a thousand, 2000 concurrent calls, multi-language anywhere in the world, 24 hours a day, seven days a week. So if you just compare the cost of this being a couple of dollars versus a person, depending on which side of the fence you fall on for that, it could either be the worst thing in the world or a big opportunity.

Amie @ Ben's Bites (53:45)
Yeah, well, in my experience with cold calling, it's like consistency and like it's a numbers game. So if you can get both those things dialed in, then...

Tom Osman (53:54)
Hmm. But actually, yeah, I don't think actually cold calling personally is going to be the good one for this. I think this is going to be great for all the extra touch point calls that people or businesses don't have time to make, but would be like a really valuable thing for a customer. So it could be, say if you had a delivery company or you had

tens of thousands of deliveries being made, people want to know where the delivery is from. So if you had a quick call just to say, Hey Tom, just want to let you know that your package is just left our facility and it will arrive at the shop in at this time. So you don't have to like log in or like check an update or check a text to see where it is. Just a quick courtesy call or a quick message. It happens all the time. I could courtesy call before like a check-in to a hotel, maybe all those things are working really well.

Amie @ Ben's Bites (54:21)
Right. Yeah.

Yeah, that's an interesting take that, you know, with all the fear of AI taking over human jobs, this actually is a good use case of adding in more human touch to processes. Yeah. Awesome. Cool. Well, if you're in the audience still, this video will be available in the recordings after it gets uploaded. And Tom, thanks so much. This was really interesting.

Tom Osman (54:59)
100 % That's it

Anytime. See you folks. Bye.

Amie @ Ben's Bites (55:16)
Alright, see ya. Bye everyone.

‍

‍