We’ve been working on a stealth project for Q1 which I’m delighted to unveil – you can now add intelligent AI voice agents to your numbers, be they numbers on Simwood or, via BYoC, numbers on any carrier globally that can route to SIP. Those who listened to the TADSummit Podcast earlier this week will have had early warning of this.
The easiest way to think of them is like ChatGPT over voice, but where you get to define the role the AI is playing and provide resources (web URLs) for it to consume as knowledge. Callers can speak to it and it will respond intelligently, guiding them to a predefined goal such as collecting contact information. Perhaps the easiest way to understand this is to give us a call and speak to our new sales assistant Jay on +44 (0) 330 122 3004.
Jay is configured with a simple directive text such as:
Your name is Jay and you are an enthusiastic ambassador for Simwood. You answer questions about Simwood products, services, and opinions from the various web resources you have access to. You are polite, positive and friendly. If asked, be honest that you are an AI. Do not answer questions that are outside the documents you have access to, other than polite small-talk. Your goal is to take a name and contact details so your human colleagues can get back to the caller.
Jay is also given a list of URLs such as the simwood website, blog and knowledgebase.
Finally, he’s given a URL to which a post-call webhook is sent.
Rather like a human, no two conversations with Jay (or any of the others) will be the same because he’s not scripted. In the same way as ChatGPT is not a search engine, you’re having a real conversation with deep understanding of tone and context, and the extent of the knowledge the agent has is defined by his/her creator. If you speak to Jay and think he’s a bit thick, remember a) he works in sales (I’m joking!) and b) he took about two minutes to create. As we progress, the directive text (called a ‘system prompt’), becomes where you will add the value as it enables you to shape personality, behavior and context. They might end up quite long as you fine tune to create the perfect on-brand and useful agent for your customers.
Don’t think this is confined to information lines either, the agent can take on any role you define. For example, one I was playing with for nuisance calls (when I wear down Compliance enough to allow me to deploy her!) is Nana. Nana loves getting calls, especially from scammers, and can talk for hours about the 1950s. She moves slowly though and doesn’t like impatient people! Her goal is to waste their time! Try and steal her life savings on +44 (0) 330 122 3005 if you’re so inclined. Nana’s directive text is quite a bit longer than Jay’s because she needed telling to not be so quick to leap into talking about the excitement of seeing a steam train, to let the caller work and win her trust slowly, and to wind the caller up subtly. She’s not quite right yet either IMHO.
Imagine the potential of this for your customers? Providing information, collecting information, or directing calls at scale. The roadmap to make this ever more powerful is long but really exciting!
This will be in beta in the Carrier Services API and portal soon for you to build with, and making its way into the Hosted stack next quarter. David Duffett will be demonstrating this at the Cloud Connections 2025 in Florida next week as well.
Now, under the hood we haven’t trained our own LLM model, nor are we professing to have written our own text to/from speech engine, or even our own RAG for those who care – we’ve consumed the best of breed for all. What we have done is build into the Simwood Potato the ability to interface with these for our chosen providers. We wanted to make it as easy as possible for you to consume and deploy. To be honest, this started as a pet project of mine and I got some good way along but had to bring in the big guns who, obviously, smashed it! I reassure myself that that means it was really hard!
You can think of Conversational AI as a service (a layer of foil around the Potato) in its own right, deployable directly on a number or trunk, but there are several underlying technologies which it packages up, and we intend packaging them in different ways for other services. Over time you’ll see these appear as interfaces and capabilities in their own right. This is useful for those of you who like to have more control, perhaps because you have on-premise LLMs, or because you are yourselves an AI company and need to telephony enable your hyper-verticalised agents. As always, please speak to us as our roadmap is shaped by our customers.
We really look forward to your feedback!
