In most businesses, an IT failure means people work around it for an hour. In a contact centre, it means a floor of agents sitting silent on full pay while calls go unanswered, and the cost is calculable to the minute.
That property — that infrastructure failure and revenue loss are simultaneous — is what separates contact centre IT from general business IT. It is not that the technology is more advanced. It is that the tolerance for the ordinary is much lower, and decisions that are reasonable elsewhere become expensive here.
This article covers what is genuinely different, and where the standard answers still apply.
Density changes the network conversation
An office of a hundred people generates bursty, tolerant traffic. A floor of a hundred agents generates a hundred simultaneous, continuous, latency-sensitive audio streams plus whatever the applications need — and it does so during precisely the hours the business is measured on.
- Switching capacity must be sized for sustained concurrent load rather than for peak burst, which is a different calculation and usually a different specification.
- Voice needs its own segment, both for quality treatment and because handsets are rarely patched appliances that should not sit alongside servers.
- Quality of service must be configured, not assumed. Voice tolerates far less delay and loss than the file transfers competing with it.
- Cabling and access-layer capacity are real constraints. A floor cabled for general office use will not carry a densely packed agent estate.
- Wireless is not appropriate for agent audio at density. It is fine for supervisors and unsuitable for a floor.
The trunk side has the same shape and is covered in the SIP trunking article: channels must be sized on simultaneous calls including queued callers, and outbound dialling requires a multiple of the agent count rather than a match to it. Sizing on headcount is the single most common capacity error in this sector.
Real-time constraints, and what they rule out
Voice has no retry. A file transfer that loses packets retransmits them; a conversation that loses packets loses words, and the customer hears it. This removes several options that are ordinarily reasonable.
| Decision | Ordinary business | Contact centre |
|---|---|---|
| Single internet circuit | Acceptable with an SLA | A total revenue stop; a second path is not optional |
| Cloud-hosted phone system for a co-located floor | Usually fine | Converts internal calls into internet traffic; the circuit becomes the dependency |
| Best-effort network quality | Users tolerate variability | Directly audible to customers, and measurable in handling time |
| Monitoring service availability | Sufficient | Insufficient — signalling can be healthy while every call is silent |
| Maintenance in a quiet period | Evenings or weekends | There may be no quiet period; windows must be negotiated against operating hours |
The fourth row is the monitoring point that recurs across this series and is sharpest here. A trunk can register successfully — a signalling event — while every call carried on it has no audio, because signalling and media take different paths. Monitoring that watches registration reports a healthy platform during exactly the outage customers are experiencing.
Seat economics
BPO commercial models run on cost per seat, which makes technology decisions unusually sensitive to per-agent pricing and unusually tolerant of one-off infrastructure cost. This is close to the opposite of most businesses and it changes the build-versus-buy calculus.
A per-agent monthly platform fee that looks modest at ten seats becomes a significant line at two hundred, and it scales with exactly the thing the business is trying to grow. Self-hosted platforms invert that: substantial setup and operational commitment, then near-flat cost per additional agent. Which is correct depends on scale, on growth expectations, and on whether the operation has or can retain the engineering capacity to run a platform.
The honest threshold is roughly this: below the point where a business can justify dedicated technical staff, hosted services win on total cost including the operational time that never appears in a comparison. Above it, self-hosting begins to make sense — and the licensing position of the platform becomes a real commercial question rather than a footnote, particularly where an operation intends to host on behalf of clients.
Compliance is not optional here
Contact centres carry obligations that most businesses do not, because they handle payment and personal information at volume, over a medium that records itself.
- Recording retention and access. Recordings are the largest concentration of sensitive data most contact centres hold, and they accumulate faster than anything else in the estate.
- Payment data. Card details spoken aloud end up in recordings unless recording is paused around payment or the digits are kept out of the audio path entirely. Removing them afterwards is far harder than never capturing them.
- Emergency calling. In the United States, Kari's Law and Section 506 of RAY BAUM'S Act apply to multi-line telephone systems, requiring direct 911 dialling without a prefix, notification, and a dispatchable location — which becomes an architectural problem with home-based agents.
- Dialling practices for outbound operations, which vary by jurisdiction and govern abandoned calls, calling hours and consent. These belong in the compliance assessment before a dialler is configured.
- Data residency, where clients impose requirements on where their customers' data may be processed. For a BPO serving multiple markets this can differ per client.
The last point is specific to outsourced operations and is worth designing for rather than discovering. A BPO taking work from clients in several jurisdictions may need to demonstrate separation between them, which is a tenancy and access-control question rather than a policy one.
Remote and hybrid agents
Home-based agents are now normal and they are the hardest reliability problem in the sector, because the last mile belongs to someone else. You cannot apply quality of service to domestic broadband, control the wireless, or stop a household video call competing with a customer conversation.
What is controllable is worth doing properly: supply and require wired connections and known-good headsets rather than leaving both to the agent, choose codecs against realistic bandwidth, and monitor per-agent call quality so a persistent problem is attributed to a connection rather than to the person.
Access to internal systems is a remote access problem and should be designed as one, rather than by extending the office network to every home. That includes the emergency calling obligation above, which does not follow the agent automatically.
What actually stops a floor
| Failure | Blast radius | The control |
|---|---|---|
| Carrier or trunk outage | Total | Second carrier, with tested inbound and outbound failover |
| Internet circuit failure | Total for the site | Second circuit on a different provider and path |
| Recording storage filling the platform disk | Total, and entirely avoidable | Separate storage with retention enforced |
| CRM or application unavailable | Agents take calls but cannot work | Availability treated as equal to telephony, not lesser |
| Identity provider failure | Nobody can log in | Treat as critical infrastructure with a documented failure plan |
| Power | Total for the site | UPS sized for graceful shutdown at minimum; generator where justified |
The third row is worth naming because it is so common and so preventable. Recordings accumulate at a predictable rate, on the same disk as the platform, until the platform stops accepting calls mid-morning. It is a capacity failure with a knowable date, which makes it the contact centre equivalent of an expired certificate.
Common mistakes
| Mistake | Why it happens | What to do instead |
|---|---|---|
| Trunk channels sized on agent count | Concurrency and headcount sound equivalent | Size on busiest-minute concurrency including queued callers and dialler ratios |
| Single carrier and single circuit | Redundancy budget went to servers | Second carrier and second circuit; both are total-outage layers |
| Recordings on the platform disk | It was the default location | Separate storage, defined retention, enforced automatically |
| Monitoring watches registration only | It is the obvious signal | Monitor channel utilisation, call volume anomalies and audio quality |
| Handsets on the general network | They were plugged into the nearest port | A voice segment isolated from servers and management |
| Emergency calling treated as configuration | It works when tested from a desk phone | Direct dialling, notification, and a location per endpoint including remote agents |
| Application availability treated as secondary | Telephony is the obvious dependency | An agent who can take calls but not work is still not working |
A representative scenario
A composite of situations we see repeatedly; no client detail is included.
An outbound and inbound operation of around a hundred seats runs a single platform, one SIP trunk, one internet circuit, and recordings written to the platform's own disk. It works for two years.
Three incidents in one quarter force a review. The disk fills and the platform stops accepting calls mid-morning. A carrier fault takes the floor off the air for most of an afternoon with no alternative route. And an audit asks how a 911 call from a home-based agent would be located, for which there is no answer.
The remediation was sequenced by blast radius. Recordings moved to separate storage with enforced retention, which removed the recurring outage entirely. A second carrier was added with tested inbound and outbound failover, converting a total outage into degraded service. A second internet circuit followed. Emergency calling was reconfigured for direct dialling with notification and a location maintained per remote endpoint. Monitoring was extended from registration state to channel utilisation, call volume and audio quality.
No part of that required replacing the platform. What was missing was redundancy at the layers with the largest blast radius, and a compliance position that had never been established.
Implementation checklist
- Trunk concurrency sized on busiest-minute simultaneous calls, including queued callers and dialler ratios, and monitored against capacity.
- A second carrier exists with inbound and outbound failover both tested by removing the primary path.
- A second internet circuit exists, on a different provider and path, or its absence is documented as an accepted risk.
- Voice traffic is on its own segment with quality of service configured rather than assumed.
- Switching and cabling are specified for sustained concurrent load, not office burst patterns.
- Recordings are on separate storage with retention enforced automatically.
- Card details are kept out of the audio path rather than removed from recordings afterwards.
- Emergency calling permits direct dialling without a prefix, provides notification, and carries a location per endpoint including remote agents.
- Outbound dialling practice has been checked against the rules of each jurisdiction called.
- Remote agents have wired connections, supplied headsets, and per-agent call quality monitoring.
- Access to internal systems for remote agents is designed as remote access, not as an extension of the office network.
- Monitoring covers channel utilisation, call volume anomalies and audio quality, not only service availability.
- Application availability is tiered equally with telephony.
- Where multiple clients are served, separation between their data is enforced technically rather than by policy.
Frequently asked questions
What makes contact centre IT different from general business IT?
Three properties. Density: a hundred agents in one room generate a hundred simultaneous continuous audio streams, which is a sustained-load calculation rather than the bursty pattern an office produces. Real-time constraints: voice cannot be retried, so packet loss is audible to the customer rather than invisibly corrected. And seat economics: BPO commercial models run on cost per agent, which makes per-seat platform pricing unusually significant and one-off infrastructure cost unusually tolerable. Underlying all three is that infrastructure failure and revenue loss happen in the same minute, with agents on full pay unable to work.
How many SIP channels does a call centre need?
Sized on simultaneous calls at the busiest realistic minute, which is not the agent count. Queued callers occupy a channel for the whole time they wait, transfers and conferences can hold more than one channel per conversation, and outbound predictive dialling places more calls than there are available agents by design, so its requirement is a multiple of headcount. Sizing on the number of desks is the most common capacity error in the sector, and the failure is invisible internally — callers receive a busy tone from the carrier and nothing appears in the platform logs, because the calls never arrived.
Should a call centre host its own platform or use a hosted service?
It depends on scale and on whether the operation can retain engineering capacity. Hosted services charge per agent, which scales with exactly the thing the business is trying to grow; self-hosted platforms carry substantial setup and operational commitment then near-flat cost per additional agent. The practical threshold is whether the business can justify dedicated technical staff: below it, hosted usually wins on total cost once the operational time that never appears in comparisons is counted. Above it, self-hosting becomes economically sensible, and platform licensing becomes a genuine commercial question rather than a footnote.
What compliance obligations apply specifically to contact centres?
Several that most businesses do not carry. Recording retention and access control, since recordings are usually the largest concentration of sensitive data in the estate. Payment data, because card details spoken aloud end up in recordings unless recording is paused around payment or the digits are kept out of the audio path. Emergency calling: in the United States, Kari's Law and Section 506 of RAY BAUM'S Act require direct 911 dialling without a prefix, notification, and a dispatchable location, which becomes architectural with home-based agents. Dialling practice rules for outbound operations, which vary by jurisdiction. And data residency, which for a BPO can differ per client.
What actually stops a call centre floor?
In rough order of frequency: a carrier or trunk outage, an internet circuit failure, recordings filling the platform's own disk, the CRM or application becoming unavailable, identity provider failure preventing anyone logging in, and power. Four of those are total-outage layers and only one of them is the phone platform, which is where redundancy budget usually goes first. The recording storage failure deserves particular attention because it is entirely avoidable — recordings accumulate at a predictable rate, so it is a capacity failure with a knowable date rather than an incident.
How should home-based agents be supported?
By controlling what can be controlled and monitoring the rest. The last mile belongs to someone else, so quality of service cannot be applied to domestic broadband and household traffic cannot be prevented from competing. What is controllable is worth doing properly: supply and require wired connections and known-good headsets rather than leaving both to the agent, choose codecs against realistic bandwidth, and monitor per-agent call quality so persistent problems are attributed to a connection rather than to the person. Access to internal systems should be designed as remote access rather than by extending the office network, and the emergency calling location obligation does not follow the agent automatically.
Conclusion
Contact centre IT is not more sophisticated than general business IT. It is less forgiving, because the gap between a technical failure and a commercial one is measured in seconds rather than hours, and because density concentrates every ordinary constraint into the same room at the same time.
The decisions that matter are consequently unglamorous: redundancy at the layers with the largest blast radius rather than at the most visible one, capacity sized on concurrency rather than headcount, recordings treated as a storage and compliance problem rather than a switch, and monitoring that watches calls rather than services.
Sources and further reading
- Building reliable call centre infrastructure— the full architecture, layer by layer
- SIP trunking explained— channel sizing and carrier failover in detail
- VoIP security for business— toll fraud, and capping what a compromise costs
- FreePBX CRM integration— screen pop, click to call, and writing outcomes back
- Infrastructure monitoring that works— why service availability is not call availability
Services This Relates To
Written by KYCONNECTS Engineering. Client names are withheld under confidentiality.