What an air-gapped AI deployment actually requires

In the early days of enterprise AI, the fastest path to production was to find a high-value, well-scoped use case and keep the AI agent's access narrow to manage security. That approach worked well enough to generate early wins, but those wins created an appetite for more sensitive workflows that required deeper integrations with core business systems.

With that expanded scope comes a more demanding conversation about deployment. Scoping down is no longer a viable security strategy, particularly for regulated industries like financial services, healthcare, and travel, which have very strict requirements around infrastructure.

At Decagon, we have worked through this evolution with some of the largest enterprises across these verticals. Earlier this year, we deployed our conversational AI platform entirely inside an F50 financial services company's own VPC to support their voice use cases, one of the most demanding infrastructure configurations we've encountered.

Here is what it took, and what we learned.

Not every enterprise needs the same deployment

Before getting into the specifics, it helps to understand the range of options. Enterprise deployment requirements exist on a spectrum, and the right model depends on your organization's risk tolerance, regulatory environment, and operational capacity.

Decagon offers several deployment options, but generally they fall into three broad categories:

• Standard SaaS is where the majority of Decagon's customers run, including heavily regulated F50 enterprises. Decagon hosts and manages everything, and customers get the fastest path to production. For most enterprises, this model already meets stringent security requirements.
• Single-tenant SaaS gives customers a dedicated VPC inside Decagon's cloud. The infrastructure is isolated at the customer level, with stronger separation than multi-tenant SaaS, while Decagon retains full operational responsibility.
• Cloud-prem / VPC means Decagon's full stack runs inside the customer's own cloud environment. The customer owns the infrastructure, while Decagon manages the software. It is the most demanding option in every respect, and it exists for organizations whose requirements genuinely cannot be met any other way.

These three models cover the vast majority of deployments we operate today, though we also support options such as on-premises datacenter deployments for more specialized requirements.

For a small number of enterprises, that third option is the only one that works. Getting there required us to solve certain deployment problems we had not faced before.

The engineering challenge

A customer VPC deployment means no public internet connectivity beyond the bare-minimum deemed necessary by the customer to reach the environment. Every service Decagon's stack depends on had to be either replaced with a customer-approved alternative or routed entirely through the customer's internal network. No outbound requests leave the environment, and every third-party dependency was evaluated against the customer's approved services list. 

That constraint is table stakes for this class of deployment. In practice, it surfaces three specific challenges that are worth understanding in detail.

Building a telephony layer from scratch

Voice is a core part of what Decagon delivers. In this deployment, the standard telephony vendor was not on the customer's approved services list. Rather than scoping voice out of the deployment, our engineering team built a self-hosted telephony layer from the ground up using telecom industry primitives, running entirely within the customer's environment.

This was the most technically complex part of the project. The team had to solve for SIP signaling, media handling, and connection management without any of the managed infrastructure that telephony vendors abstract away. The result is a production-grade, self-hosted system that delivers full voice capability inside an air-gapped environment. 

To our knowledge, no other AI-native conversational platform has built and deployed something comparable.

Self-hosted models inside the customer perimeter

Decagon’s agents are built upon a coordinated network of specialized models. We use both models from the frontier labs, along with our models developed by Decagon Labs that are purpose-built for CX performance and latency.

In an air-gapped environment, standard external API calls to those providers are not possible. The solution required two things in parallel: private connectivity with frontier model providers into the customer environment, and self-hosted model inference layer for the models Decagon has developed. The full model infrastructure had to be reconstructed within the customer's perimeter without degrading agent quality or response latency.

Deep integration with customer systems

Enterprise customers do not have generic infrastructure. This customer had its own internal tooling, authentication systems, and data systems that Decagon's agent runtime needed to connect with natively rather than through APIs. The integration work required building custom connectors specific to this customer's environment, coordinating closely with their engineering teams, and going through their internal approval processes for each system touch.

This kind of work is operationally intensive. Getting an AI agent to work for specific workflows is one problem, but getting it to work inside a highly specific, tightly controlled environment is a different one.

Why SaaS is still the right call for most

The deployment described above is impressive engineering. It is also expensive, slow to stand up, and operationally demanding for the customer. That is worth being direct about.

Customer VPC deployments require over-provisioned infrastructure by design, high setup costs, and a deployment timeline that is longer than standard SaaS. Because the customer owns the environment, they are also responsible for first-line uptime defense, which adds to the operation burden.

Several enterprises have gone through serious evaluations of the VPC model and concluded that Decagon's standard SaaS meets their actual security requirements. That outcome is not a failure of ambition but reflects a genuine understanding of what the VPC model costs versus what it delivers. For organizations without hard regulatory or contractual requirements that mandate private infrastructure, the SaaS model provides a hardened, enterprise-grade security posture without the overhead.

A new baseline for enterprise AI

The enterprises setting the highest infrastructure bar are not waiting. Decagon now has validated, working deployment packages across a large spectrum and has operational experience to support all of them.

If infrastructure requirements have been the reason your organization has not moved forward with AI, we would love to talk.