You run a digital agency. A client asks if you can add AI voice to their stack. Your developer says "easy, we'll spin up an OpenAI Realtime API instance and wire it to Twilio". Two sprints later you're debugging SIP trunking at 11pm and the client still can't make a test call. Sound familiar?

Self-hosting AI voice looks cheap on paper. In practice, the hidden costs-infrastructure, telco compliance, ongoing maintenance-add up fast. Here's what a realistic 12-month total cost of ownership actually looks like, and when a platform makes more sense than rolling your own.

What self-hosting really costs

Start with the obvious: developer time. A competent full-stack engineer billing at A$120/hour will need 80-120 hours to wire OpenAI Realtime (or ElevenLabs, Deepgram, whatever) to a SIP provider, build a basic IVR flow, handle interruptions, and deploy something that won't fall over under load. That's $9,600-$14,400 before the first live call.

Next, infrastructure. You need:

• A WebRTC or SIP gateway (Twilio Elastic SIP Trunking, Vonage, Bandwidth.com).
• Low-latency compute to run the voice agent (AWS us-west-2 or Sydney ap-southeast-2, depending on your LLM provider).
• Storage for call recordings and transcripts.
• A Redis or Postgres instance for session state.

Budget A$300-$600/month for a small production environment. Add another $200-$400/month if you want proper monitoring (Datadog, Sentry) and log aggregation.

Then there's telco compliance. In Australia, if you're making outbound calls to mobiles, you need an approved sender ID or the carrier will block you. That's ACMA registration, DNCR integration, and state-specific dial-hour gating. One agency we spoke to last month spent three weeks just navigating Telstra's sender ID approval process. If you bill your time at $150/hour, that's another $4,500 in hidden cost.

The compliance tax nobody mentions

Australia has three big regulatory traps for AI voice: the Privacy Act, the Spam Act, and the Do Not Call Register. If your self-hosted agent doesn't scrub numbers against DNCR before dialling, you're exposed to fines up to $2.5 million per breach.

Implementing DNCR wash properly isn't hard, but it's fiddly. You need to subscribe to the Wash API (around $300/year for small volumes), write the integration, handle timeouts, and cache results. Then you need to log every wash attempt for audit purposes. That's another 20-30 hours of dev time.

State dial-hour rules are worse. In NSW you can't cold-call a mobile before 9am or after 8pm on weekdays. In Victoria it's 9am-5pm weekdays only. If you're running outbound campaigns across Australia, you need timezone-aware scheduling that respects each state's rules. Most agencies we talk to either ignore this (risky) or spend a week building a cron-based gating system (expensive).

Ongoing maintenance is the silent killer

Let's say you get everything running. Six months later OpenAI changes the Realtime API response format and your agent starts dropping calls. Or Twilio deprecates a webhook endpoint. Or your LLM provider adds a new region with 40ms lower latency and you need to migrate.

Budget at least 10 hours per month for ongoing maintenance, bug fixes, and dependency updates. That's another $1,200/month in developer time, or $14,400/year. If you're managing this for multiple clients, multiply accordingly.

When the platform model makes sense

A platform like VoxReach shifts the cost structure. Setup fee is A$5,500 upfront. Inbound calls cost $0.42/min, outbound to AU mobiles $1.32/min. For a client taking 200 inbound calls per month (average 3 minutes each), that's $252/month in usage. Total year-one cost: $5,500 + ($252 × 12) = $8,524.

Compare that to self-hosting: $12,000 dev + $6,000 infra + $14,400 maintenance = $32,400. Even if you halve the maintenance estimate, you're still at $25,200.

More importantly, the platform handles compliance for you. DNCR wash, state dial-hour gating, ACMA sender ID management-all built in. You also get 30+ native integrations (HubSpot, Pipedrive, ServiceM8, Cliniko, etc.) without writing a line of code.

What to do if you're deciding today

Self-host if: you're building a very high-volume product (10,000+ calls/month), you need tight control over the voice model, or you have deep telco experience in-house.

Use a platform if: you're an agency serving 3-20 clients, you want predictable costs, or you'd rather spend dev time on custom CRM integrations than debugging SIP headers.

The real trap is underestimating maintenance. A self-hosted voice agent isn't a "set and forget" project-it's a service that needs feeding every month.

If you want to see what a platform approach looks like, get started at app.voxreach.com.au/signup. You get calls on your own number to test the full feature set.