Build vs. buy: How to decide for your orchestration layer

Five CIOs and CTOs in Twilio's Innovators Report had to decide whether they’d build or buy their orchestration layer. And they all found different answers that worked:

• Juliano Skrzyszowski at Nova Gestões: Don't hand your core business to third parties, build your own orchestration layer, keep your own decision engine.

Juliano Skrzyszowski at Nova Gestões: Don't hand your core business to third parties, build your own orchestration layer, keep your own decision engine.

• Arun Rao at Eltropy: Consolidate the communication layer onto one platform to stop managing disparate vendors.

Arun Rao at Eltropy: Consolidate the communication layer onto one platform to stop managing disparate vendors.

• Jason Johnson at Sweetwater: Moved away from a built in-house orchestration layer when the system hit its limits.

Jason Johnson at Sweetwater: Moved away from a built in-house orchestration layer when the system hit its limits.

All three were right. That’s because "orchestration layer" describes a few different things, and the build-or-buy answer changes depending on which one you mean.

The four layers people call orchestration

Most build-versus-buy debates collapse these layers into one question, but that produces an inaccurate answer for most of them:

• Decision engine. What should happen next, and why. Eligibility rules, risk thresholds, which offer, which treatment, when to escalate. This encodes how your business works.

Decision engine. What should happen next, and why. Eligibility rules, risk thresholds, which offer, which treatment, when to escalate. This encodes how your business works.

• Context layer. Who this customer is and what you know about them, assembled in time to be useful. Identity resolution, profile, history, freshness.

Context layer. Who this customer is and what you know about them, assembled in time to be useful. Identity resolution, profile, history, freshness.

• Channel layer. Getting messages to people across SMS, voice, chat, and RCS, with carrier relationships, deliverability, and regulatory compliance in every market you operate in.

Channel layer. Getting messages to people across SMS, voice, chat, and RCS, with carrier relationships, deliverability, and regulatory compliance in every market you operate in.

• Agent runtime. Speech pipelines, model calls, interruption handling, tool use, guardrails.

Agent runtime. Speech pipelines, model calls, interruption handling, tool use, guardrails.

1. Decision engine: Build (or at least own)

This is the layer Juliano is protecting. Nova Gestões handles millions of sensitive financial negotiations monthly under regulatory oversight, and they used documented APIs to integrate communication into their existing stack precisely so they could keep full control over core business logic and decision engines. If a vendor owns how your business decides things, you've outsourced the part nobody else can rebuild for you.

2. Context layer: Buy the infrastructure and own the model of the customer

The challenging parts with the context layer are identity resolution and freshness, and both are well-solved problems you'd be rebuilding (and probably not as well).

What you should own is the definition:

• Which customer attributes matter

Which customer attributes matter

• How a customer profile is structured

How a customer profile is structured

• What "returning customer" means in your business

What "returning customer" means in your business

48% of companies name connecting data across channels (like voice, email, and messaging) as their top obstacle, but almost none of them are failing for want of a custom-built pipeline.

3. Channel layer: Buy

Carrier relationships, number provisioning, deliverability, and per-market compliance are a full-time business, and it really shouldn’t be yours. That’s why Juliano recommends you prioritize partners who strictly adhere to channel policies, because the path between you and your customer has to stay open. That's a vendor-selection question (not a build question).

4. Agent runtime: Buy the runtime but keep the model choice

Speech pipelines and interruption handling are infrastructure. The model, prompts, and guardrails are yours. Arun Rao recommends multi-tiered input, processing, and model boundaries, plus grounding responses in audited knowledge repositories with source attribution.

5 questions to help you find the right option

These questions will help drive you in the right direction:

• Does this encode how our business makes decisions? If yes, own it. If it's how messages reach people, buy it.

Does this encode how our business makes decisions? If yes, own it. If it's how messages reach people, buy it.

• Would we rebuild this every 18 months anyway? Carrier compliance and channel APIs change constantly. That maintenance is a subscription whether you pay a vendor or your own engineers.

Would we rebuild this every 18 months anyway? Carrier compliance and channel APIs change constantly. That maintenance is a subscription whether you pay a vendor or your own engineers.

• Can we hire for it? Speech infrastructure and deliverability engineering are specialist markets. If you can't staff a team that can be on call for it, you haven't built it, you've just prototyped it.

Can we hire for it? Speech infrastructure and deliverability engineering are specialist markets. If you can't staff a team that can be on call for it, you haven't built it, you've just prototyped it.

• What happens at ten times the volume? This is the question Sweetwater's homegrown systems answered the hard way.

What happens at ten times the volume? This is the question Sweetwater's homegrown systems answered the hard way.

• Can we get our data out? Buying a layer is survivable if you can leave. Ask for the export path before you sign, though.

Can we get our data out? Buying a layer is survivable if you can leave. Ask for the export path before you sign, though.

The build ceiling that’s hard to plan for

Sweetwater's experience is a productive counterweight to Juliano's advice. They had built in-house, and the constraint that forced the change was the technical limitations of those homegrown systems as interaction volumes surged.

What they run now is a bought channel layer wired into the CRM they kept:

• 50,000 texts and 40,000 calls daily

50,000 texts and 40,000 calls daily

• 850 agents

850 agents

• New contact center configurations deployed in under two days

New contact center configurations deployed in under two days

The workflow logic stayed theirs (built in Twilio Studio), but the infrastructure didn't.

In-house systems usually work well right until they don’t. The failure isn’t a gradual decline you watch. It’s when you make a big change in volume (often uncontrollable), you enter a new market with different regulations, or the engineer who wrote everything leaves.

Sequence your orchestration in the right order

Devesh Maheshwari's advice at Lendi Group: Unify the real-time data layer before scaling AI, so every agent works from one source of truth. He calls fragmentation the data fragment tax.

• Settle the context layer first (everything above it depends on what the AI can see)

Settle the context layer first (everything above it depends on what the AI can see)

• Figure out the channel layer (almost always bought)

Figure out the channel layer (almost always bought)

• Decide on agent runtime. Keep the decision engine yours throughout.

Decide on agent runtime. Keep the decision engine yours throughout.

84% of brands say their AI can access customer history while only 40% design personalization around it. That gap is a decision-layer problem, but no vendor closes it for you.

Decide on build vs. buy per layer (not per platform)

You don’t need to build everything, nor do you buy everything. Mix and match for the layers that make the most sense.

Ultimately, you might build a layer, buy the rest, keep the export path, and spend the engineering you saved on the layer that truly differentiates your business.

Read how each of the five executives approached build vs. buy it in the Innovators Report. Ready to test the layers you'd rather not build from scratch? Start with Twilio for free.

Frequently asked questions

Should I build or buy my orchestration layer?

Decide per layer. Own the decision engine that encodes your business logic, and buy channel transport, identity resolution, and agent runtime. Nova Gestões' CTO kept core business logic in-house while integrating communication through APIs.

What is an orchestration layer?

An orchestration layer coordinates what happens next in a customer interaction. It spans four things: decision logic, customer context and identity, channel transport across SMS, voice and chat, and the AI agent runtime.

When does building in-house stop working?

Usually at a step change. Sweetwater moved off homegrown systems when interaction volumes surged past their technical limits. New markets with different regulations and specialist staffing gaps are the other two common triggers.

What should I never outsource?

Your decision engine. Nova Gestões' CTO recommends keeping your own orchestration layer and decision engine instead of handing core business logic to a third party, and integrating through documented APIs.

Which layer should I fix first?

The context layer. Lendi Group's CTO recommends unifying your real-time data layer before scaling AI. Twilio research found 84% of brands say their AI can access customer history while only 40% design personalization around it.