80% Done: The Open-Source Playbook for Replacing BPO with Conversational Voice AI

I recently heard the new philosophy for building enterprise software, and it’s basically:

“Start at 80% done — and figure out the 20% that actually matters.”

Couldn’t agree more when it comes to customer service.

There are 4 ways to handle enterprise support:

1. Sierra/Decagon - SaaS that "just works." Actually service companies in disguise. You own nothing & pay for platform & per resolutin..

2. Build from scratch - 18 months on plumbing before your first business-specific eval.

3. BPO + "AI services" - Teleperformance promising tech. Their revenue is agent hours. They'll never automate themselves.

4. Open-core - Start with 80% done (infrastructure). Own the 20% that matters (evals, improvement loops, customer intelligence). Pay just for the LLM costs, without any margin. Total freedom.

And we'll talk about why open-core is a way to go, enjoy!

The False Choice in Conversational AI Contact Centers

Enterprise contact centers are stuck in the same trap as enterprise software - a false choice between:

Option A: Rent from a SaaS

• Upload your knowledge base

• Connect your tools

• Write your prompts

• And hope it all just works out

Spoiler alert: it doesn’t. These are service companies in disguise. Take a look at their job boards. They’re hiring Agent Engineers and Agent Product Managers. They’re writing custom code for every client behind a shiny SaaS facade.

And don’t even get me started on how closed they are. Your customer data, conversation info, and improvement loops - all trapped in their platform.

Option B: Build from Scratch

• Set up your voice infrastructure

• Build a conversation storage system

• Create escalation paths

• Design eval frameworks

• Write integrations

You’ll be stuck on the plumbing for months before you get around to writing your first business-specific eval.

Modern conversational voice AI solutions can integrate with your existing tech stack and tools, such as CRM and support systems, to enhance operational efficiency. This seamless integration enables unified communication and streamlines workflows across platforms - but all this takes time.

Option C: Ask your BPO to “add AI”

Sounds safe, right? Teleperformance, Concentrix, TTEC - they’re all promising AI now. You get to keep your existing relationship and just tack on some AI.

Here’s the problem:

• Their business model is based on agent hours. They’ll never automate themselves out of existence. Classic innovator’s dilemma.

• The AI is just a feature to win RFPs - it’s never going to be able to resolve tickets on its own.

• They’re operations companies bolting on tech. No eval frameworks, no continuous improvement loops. It’s just QA teams doing spot checks the same way they always have.

• And you still don’t own a thing.

All three options are terrible. Here’s why:

The 80/20 of Contact Center Operations

The 80% (table stakes):\

The 80% is just infrastructure. Everyone needs it. Nobody should be building it from scratch. Custom conversational voice AI solutions are optimized for performance and can handle high volumes of concurrent communication, such as calls and messages.

The 20% is what makes your support better. It’s your IP and what your domain experts know that no SaaS vendor can replicate.

Why Evals Are the Real IP in Contact Centers

A lot of people get this wrong about AI systems.

Over the last 30 years, business processes moved from paper to software. Every workflow became a database schema, UI, or API contract. A mirror image of the business.

As we move to conversational interfaces, these same processes have to be translated again - into evals.

Evals tell you how conversations should go. They're executable tests.

Every business process becomes an eval. A mature operation will have thousands of these.

This is your competitive advantage. Not the prompts, not the voice provider - the evals.

When you use a BPO like Sierra, they own your evals. When you build on open infrastructure, you own them.

The Pricing Arbitrage

With open-core, the economics work differently:

-picture screen

What the 80% of Phone Calls Looks Like in Real Life

We’re building this thing and here’s what the architecture looks like:

┌─────────────────────────────────────────────────────────────┐\
│                      Agent Plane                            │\
│

┌─────────────────────────────────────────────────────────────┐\
│                      Agent Plane                            │\
│

┌─────────────────────────────────────────────────────────────┐\
│                      Agent Plane                            │\
│

This system is designed as an AI platform that empowers developers to build, customize, and deploy conversational voice AI solutions across a range of devices and environments. The platform supports the creation of a custom voice for your AI agent, allowing you to fine-tune the voice, tone, and prompts to match your brand identity.

Custom Conversational Voice AI Solutions can be built on dedicated infrastructure to ensure high performance and reliability. Data security is a critical aspect, with options for secure data encryption to protect sensitive information.

Agent-agnostic, self-hosted or cloud - you get to own the data.

The Actual Improvement Loop You Want

Here's how things go with BPO (that's Business Process Outsourcing)

1. Customer calls

2. Some offshore agent handles it - but here is the thing - its not very consistent

3. Maybe they even log a bit of the conversation

4. Any knowledge that was gleaned just stays in the agents head - which is of course useless when they leave

5. And then that agent chews

6. And suddenly you are back where you started

Now here's what happens when you use your own infrastructure :

1. customer calls

2. AI handles it

3. You log the full transcript along with the tool calls

4. The eval catches any failure patterns that happen

5. Domain experts go and write new evals

6. The system gets better

7. And the better it gets the faster it improves.

The second loop is what evals are all about. Every failure now becomes a test case. Every test case makes the system better - from now on.

This happens with our open-core model.

It can't happen with BPO though.

It's locked in to the SaaS

Who We're Talking To

This isn't for everyone.

If you are just starting out and you only handle 100 support tickets a month, you should just use Intercom. Next!

But if you are an enterprise dealing with 100,000 contacts a month - across voice and email, and you have workflows specific to your domain that no SaaS will ever get right then you need to own your infrastructure.

You need people on board who understand your business, building evals that capture what makes your business special.

You need a platform that lets you go off the rails if needs be

You need to be able to start with 80% done and then build the 20% that makes all the difference.

What's Going to Happen Next

We are open sourcing the control plane. That includes the unified inbox, the conversation store, the eval framework, the observability.

The 80% will be yours. The 20% is what we will help you build.If you're running a contact center and fed up with being forced to choose between the vague promises of BPOs and the unwanted shackles of SaaS lock-in , lets have a chat

ModelGuide is building a contact centre infrastructure that you control, a fully open source solution with voice, email, and chat functionality all working together seamlessly, regardless of the agents you use. And the best part? You get to host it all yourself, on your own servers.