The first question a voice agent company usually asks is which model to use. Ours was which line to own. A session border controller, an SBC, is not the most interesting box in a voice stack until a call fails at 02:14 and the only evidence is a ticket number.
The reference architecture runs its own SBC because the alternative is to rent the one layer that sees the whole call. A resold API can give you a clean dashboard and a fast start. It cannot give you the carrier route, the media leg, the codec renegotiation or the exact millisecond where the audio stopped. For an agent that is judged on a conversation, that is the wrong layer to lose.
What an SBC actually does
An SBC sits at the border between an untrusted network and a controlled one. In a voice stack it is the first and last place where a call is a set of packets you can inspect, rather than an event in someone else's analytics.
It terminates SIP and media, hides internal topology, normalises headers and codecs, applies rate limits and fraud controls, and routes the call to the right application. When a carrier sends a 488 Not Acceptable Here, the SBC is where you find out which leg refused which codec. When a caller hears silence, it is where you compare the RTP timestamps on both sides.
- Signalling
- SIP over TLS, header normalisation and route selection. A malformed
Contactheader is a dropped call, not a cosmetic defect. - Media
- SRTP termination, jitter buffering, echo control and codec negotiation. The media path decides whether the agent sounds present.
- Policy
- Rate limits, allow-lists, toll-fraud checks and spend caps. The line is a financial surface as well as a technical one.
- Failover
- Two edges, two carriers, one route table. A carrier maintenance window should not become a customer incident.
The status page is not a root cause
When a call drops through a resold API, the vendor can tell you that the platform is operational. That statement may be true at the platform level and useless at the call level. The failure could be a carrier route, a certificate, a codec mismatch, a firewall rule or a media server that ran out of ports. Without the SBC, you are asking someone else to investigate a call you own the outcome of.
We do not think every company should run an SBC. We think every company should know which failures it can see. If the answer is "the vendor will tell us", the vendor has become a single point of both failure and knowledge.
The line is not a commodity at the edge. It is where every promise is tested.
What we operate
Our telephony team runs the edge, the media path and the carrier relationships. The work is unglamorous and specific:
- Two edges in separate failure domains. A route that works in one availability zone should not depend on the other one being healthy.
- Number inventory and porting. A number is a regulated asset with a cutover window, a fallback plan and a customer expectation attached.
- Carrier interconnects. Least-cost routing is constrained by quality, coverage and fraud risk. The cheapest route is not always a route we will use.
- Per-call tracing. Signalling, media and application timings share a call ID. A trace is not a screenshot of a graph.
- Toll-fraud controls. Spend caps, destination allow-lists and anomaly alerts. Voice fraud is a financial attack before it is a technical one.
| Responsibility | Failure it prevents | Where it is seen |
|---|---|---|
| Topology hiding | Internal address leakage and direct attacks | Signalling trace |
| Codec policy | One-way audio and silent calls | Media trace |
| Rate limiting | Registration floods and toll fraud | Edge metrics |
| Route failover | Carrier maintenance becoming an outage | Route table and call trace |
The bill for owning it
Owning the layer costs money and attention. There is a pager, a capacity plan, carrier contracts, certificate rotation, codec testing and a person who knows why a particular route is excluded. The SBC itself is not the expensive part. The operational ownership is.
We pay that bill because voice is the product. If the line is a side feature of a larger software business, the calculation may be different. There is no virtue in operating a border controller for its own sake.
When renting is right
If you are adding a single voice workflow, have no telephony team and can tolerate a vendor's failure domain, a managed voice API is a rational choice. It is faster to launch and it moves a class of problems to someone else.
The calculation changes when the call is the business: when a dropped call is a lost customer, when a carrier route is a margin decision, or when the latency of an extra hop is visible in the conversation. At that point, the abstraction stops being leverage and becomes a ceiling.
The line is the truth
A voice agent is not only a model. It is a model attached to a line. The line determines whether the caller is heard, whether the agent is understood, and whether the call ends with a resolution or a dial tone.
We run the SBC because we want to be the company that answers for the line. The model can be replaced. The pager cannot be outsourced.
We do not route emergency calls. Emergency services require capabilities and certifications outside our scope.