AI Voice Agents - We Just Cracked The Code

Once I dug deeper into Retell, I found it so damn cool. It's the most advanced tool I've found - it offers an MCP, Logic split, call transfer routing and a super cool workflow visualisation. Maybe a bit of an overkill for small businesses who just need a little help, but great for any business who receives a lot of calls often with the same requests.

There are so many awesome features with Retell! You can even setup real time transcription settings or post call data extraction, what a win for sales teams. I also tested the voice and I thought it was really really good!

When I saw that Retell AI offers a debt collector service as well, I thought this is too hilarious and had to try it. Technically you should be able to test the agent and let it call you, but practically I'm still waiting for the call. But I suppose numbers from Europe don't work.

The voice I tested was also incredibly real sounding, I would not have realised she's not real. This is my favorite AI voice agent by far!

By the way, prices are by use.

Hume AI is one of my favorite voice generators, but I'm again not sure if it deserves to be called a proper AI voice Agent, because currently you can't train her on your own data and Hume AI can only access generally accessible data from LLMs. But I did have a real laugh when she answered that adding specific data to her would be like asking a fish to climb a tree. It did say in the system prompts that she should be my funny assistant, so 100 points on that!

Also, Hume AI can read your emotions incredibly accurately, so much so it's almost scary. During your conversation on desktop, you can always see 3 emotion markers and a score for those (check out the screenshot).

Example configurations are customer support, Spanish learning companion, or smart companion, just to give you an idea what you can use Hume AI for. These come with pre-written prompts.

Options for Hume AI include choosing your models (Gemini, GPT, custom, Hume's own). You can also configure Webhooks, whether EVI starts the conversation, inactivity messages, timeouts, and choose the voice for your agent.

ElevenAgent is a sub-product of ElevenLabs, which is the most established AI voice generator. A logical step for this company to venture into AI voice agents too.

It generates system prompts for you based on the simple instructions you give first. Goal, Tone, Environment & Personality are already pre-defined for you, but you can easily adjust them.

You can connect ElevenAgents to your website via webhook, or tools like Cursor, HubSpot, ServiceNow, or Zendesk.

I tried to test it with a proper test environment where it could access my info from the website, but the test didn't work - I kept getting an error message.

I did test the agent under "playground" which worked really well, but of course that one is trained on ElevenLabs data and naturally knows its stuff.

Voice-wise the agent sounded very real, but I wouldn't expect anything else from ElevenLabs, the kingdom of AI voices.

I really like that Voice.AI starts off by asking you a few questions about your business and letting you enter your website. Next, it only takes a few seconds before Voice AI is already creating your agent.

Then you can right away test Voice AI, not on your real data but on how the agent would handle customer calls.

The voice sounds a bit less human and slightly metallic, but the answers are really good and relevant.

It can handle RAG (Retrieval Augmented Generation), which is pretty impressive.

Generally, the setup looks a lot like the one of ElevenAgent from ElevenLabs. There's a knowledge base where you can add your training data, and you can also purchase a phone number directly through them for $12 per month.

Rosie is a simpler AI voice agent, I would call her more of a note taker than an AI agent because of the limited options for input of information.

You can train Rosie based on your Google business profile or your website and add up to 20 FAQs she can use to answer questions, which is to be honest not a lot. I think for a proper call agent, Rosie is too simple.

She mostly connects to your calendar apps like Calendly, Google Calendar, Acuity Scheduling and Zapier.

The voice quality is great and sounds real, there is even background noise as if she was in a cafe somewhere.

Now this is hilarious - Vapi was the only test voice that was male, and it tried taking over the conversation a lot. I swear! It kept asking me how I heard about Vapi.

Basically what I got from the conversation with Vapi is that this is really a tool for developers, and setting it up would definitely require one (according to the Vapi man's opinion).

There was a slight delay with the voice response time.

Funny, he sounded a bit annoyed even at the end, but I think I'm projecting. Maybe it was me who was getting annoyed.

My conversation and onboarding with Bland AI was pretty bumpy. But what was nice is that Bland AI was able to call me, the only AI voice agent that could do that, so that's a plus point.

On the call itself, the agent often said "I don't have the exact numbers at hand," even when I asked about their free plan. I guess the demo isn't trained on their own product data. Then I had to sign up again and receive an SMS code, which didn't come through at first to verify my phone number—which they already had through the call—so that was an unnecessary bump in the road to Bland AI.

Then they led me through another demo, but now the voice was really weird, even though I tried all 4 available voices (check it in the video above).

Pity because otherwise it seems like Bland AI is a cool tool with advanced features like call analytics and even lets you create web widgets (I suppose similar to PayPal buttons). Would have loved to experience the full potential of what seems like a promising product beneath those rough edges.