Voice Recognition IVR, Alternative to AI Receptionist?

[Reece Cawthorn]

Hi All,

I'm currently working on a proposal for a large customer that receives approximately 65,000 inbound calls per month.

One of their key requirements is an IVR that allows callers to verbally state the person or department they wish to reach, with the call then being routed accordingly.

My initial thought was to use the AI Receptionist, but based on OpenAI usage costs, this could add somewhere in the region of £6,500–£10,000+ per month for this feature alone, which makes it difficult to justify.

I've also been exploring whether this could be achieved using the CFD and wondered if anyone has implemented something similar. If so, I'd be really interested to hear how you approached it, any limitations you encountered and the associated costs.

Equally, if anyone has any recommendations for a more cost-effective speech recognition solution that integrates well with 3CX, I'd love to hear your thoughts.

Any advice would be greatly appreciated!

Thanks in advance,

Reece

 

[Evolute IT]

Wait for the Grok implementation maybe? It will be much cheaper than OpenAI...

It's technically doable in CFD, but I would never recommend it. The way the voice is recognized is way less reliable than AI. (Miky vs Mikey would not give the same results in CFD.)

 

[Reece Cawthorn]

Thank you for getting back to me and for the information!

The introduction of Grok within the AI Agents would be a major plus for us. Since migrating our AI-enabled 3CX estate from OpenAI to Grok, we've already seen noticeable improvements in transcription accuracy, AI-generated summaries and (most importantly) cost!

Do you happen to have any idea on timescales for when Grok support may be introduced into the AI Agents? This particular customer is targeting a deployment by November 2026, so it would be helpful to understand whether this is likely to align with their implementation timeline.

 

[KyriacosS_3CX]

Hi @Reece Cawthorn, you can discuss this directly with your 3CX representative to get a better idea on the timeframes.

 

[edossantos_sipcaller]

In case you want to do this with the CFD, that's something I personally did some time ago and managed to get it working quite well by using synonyms to cover most of the alternative recognitions returned by the speech recognition engine. However I must admit that it's not a simple project, and having AI it's probably not worth going that way. I would at least wait for Grok to be available to decide if investing the effort on this makes sense.....

 

[VoIPTools]

I believe we can get the AI cost down to a ballpark of $50 per month for 65,000. We would need a very small VM to host the necessary API, and your AI API key, but that's significantly better than your estimated $8,600–$13,200 per month.

We already have the backend API and administrative UI written. We would use essentially the same infrastructure we developed for our VoIPTools in-application AI assistant (coming soon). And then we start asking questions about what else we can do with this infrastructure...

However, let's not forget the cost of recreating the 3CX Company Directory functionality, but that is a one-time cost. You can book a one-on-one with me at https://booking.voiptools.com.

 

[edossantos_sipcaller]

I believe we can get the AI cost down to a ballpark of $50 per month for 65,000. We would need a very small VM to host the necessary API, and your AI API key, but that's significantly better than your estimated $8,600–$13,200 per month.

$50 for 65.000 calls? What AI model are you using to get those numbers? Looks like too cheap to me.

We work with different AI Providers with SIP Caller, Gemini Live 3.8 has been just released, has a much lower price compared to OpenAI (maybe just 25% of their cost), works really well, and the cost is around $0.023 per conversation minute, not considering the tokens used for text to provide instructions. Making a quick calculation, assuming 2 minutes calls in average, that would be around $3.000 per month....

 

[VoIPTools]

I think we need to focus on this specific request:

One of their key requirements is an IVR that allows callers to verbally state the person or department they wish to reach, with the call then being routed accordingly.

This is less about the model used, and more about where you employ AI. The budget I proposed is based on OpenAI numbers. I expect Grok will be even more cost effective. These estimates are based on it taking 5 seconds for the caller to say the person's name, and that's a pretty generous allocation.

I suspect your estimates were based on using a real-time conversational AI to do this. That is not necessary. You just need to capture what the person said, look up the closest matches (with confidence score), and decide whether you have a match. You don't need AI to get the closest matches to your 3CX extension list, there are other ways of getting that information.

Do what you can through traditional means (C#, 3CX APIs, etc.), and only use AI for the portions of the process that require it. That's the key.

 

[edossantos_sipcaller]

Yes, that's a totally different approach, and it's what you can already do with the CFD and the Voice Input component. It's a challenging project, I did it once so I can confirm it's possible, but not an easy one. Also, people are getting used to speaking naturally, and this will feel old school, however a valid approach.

 

[VoIPTools]

Totally agree with you Ernesto. I think we should just replace IVRs entirely with voice-bots. But that wasn't the requirement, and on-going costs was already a concern. I've done it the old way too using the CFD components and frankly I concluded the results were barely passible (if that) AND a lot of work. That's not a reflection on the CFD, it's an acknowledgement that the world has moved on and there are better tools now.

We created our own voicebot and web-based AI agent. They will be showing up in the next version of VoIPTools. Our voicebot simply registers as a local extension on 3CX (or any other phone system). That means it can converse with anyone that reaches that extension. So far, we have not run into any simultaneous call limitation beyond the 3CX licensing limits.

One of the first design constraints we enforced with our voicebot is where it can get information. We didn't want the bot going out to the internet and grabbing who-know-what information and giving garbage responses. To train our bot, we created an ingestion process that scans our website, user manuals, YouTube channel, and other supplementary sources to create an automated, constantly updated, knowledgebase our bot uses to answer questions. It has worked far better than I could have dreamed possible.

We make changes to our manuals and website every day and those changes magically get incorporated, without us even having to think about it, into the knowledgebase. At the same time, we are watching the questions people ask and track the AI's confidence scores. Then we use AI to aggregate those conversations into the top questions that require additional information. It is a feedback loop that helps our voicebot / web-agent constantly improve and ensures the next time someone asks a similar question AI has the information it needs.

That's the world we live in today, and it seems to be getting better with each passing week. It is both an exciting and scary time for programmers like you and me. The biggest challenge is keeping up with all the changes. I'm not keeping up even working 18 hour days including most weekends. My wife is not happy.

 

[KyriacosS_3CX]

Hi All,

Just a mention that with Update 10 the AI in Call Scripts or Programable Extensions should be able to handle this type of scenario while keeping things fairly close to the PBX and the facilities it provides natively.