The Per-Minute Tax on Fintech AI Calling, and the Open Stack That Skips It

2026-11-11 · Cloudonix Team

The market is real, and so is the per-minute tax

If you run KYC follow-ups or collections for a fintech, you've already seen the buyer guides. Eight platforms, all promising natural voices, compliance badges and easy handoffs. The decision you're actually facing is simpler: how much of your margin are you willing to hand over per minute, forever?

The demand side is genuine. Forecasts cited in recent buyer guides put the AI debt collection segment at roughly USD 2.8 billion in new spending between 2024 and 2029, growing at about 15% a year. KYC chasing and payment reminders are repetitive, scripted and high volume, which makes them close to ideal for automation. Nobody serious disputes that.

Important: Industry forecasts cited in 2026 buyer guides project roughly USD 2.8 billion in new AI debt collection spending between 2024 and 2029, at a 15% CAGR. The buyers paying that bill are the same fintechs comparing platform rates of $0.04 to $0.14 per minute.

What deserves scrutiny is the pricing model every vendor on the list shares. Per-minute metering made sense when minutes were the scarce resource. In AI calling, the scarce resource is your own workflow, and the minutes are a commodity you can buy at wholesale.

What the per-minute math looks like at collection scale

Here is the published pricing from a recent comparison of eight platforms, with one open alternative added:

Run a mid-size collections operation at 250,000 minutes a month. Plivo's bundled rate costs you $10,000. Bland starts at $35,000. Vapi's platform fee alone is $12,500 before telephony, transcription, LLM and voice costs land on top. These are platform fees, the toll you pay before a single word is spoken.

The open stack in the last row charges nothing at that layer. OPBX is MIT licensed. Dograh OSS runs on your own server. Cloudonix carries the calls on session-based trunks rather than per-minute metering, so a four-minute hardship conversation costs the same session as a forty-second reminder. You still pay your speech and language providers, but you pay them directly, at their rates, with no margin stacked on top.

The cost gap is structural. A hosted platform bundles telephony, speech models and orchestration into one per-minute price and keeps the difference between that price and its own costs. At small volumes that convenience is cheap. At collection volumes it becomes one of the largest line items in the operation.

Self hosting flips the equation. Your cost curve follows your servers and your AI provider invoices, both of which you can negotiate, swap or optimize. The platform layer stops scaling with your success.

The open stack alternative: OPBX, Cloudonix and Dograh

The stack has three parts, and each one earns its place.

OPBX is the open source business PBX. It holds your numbers, extensions, IVR menus, queues, business hours and, critically for this use case, an auto dialer with campaign scheduling, timezone awareness, CAC and CPS pacing, ML-based answering machine detection and caller ID pooling. It routes calls to AI providers over SIP or WebSocket, and its provider registry lists 48 entries, from Retell and VAPI to ElevenLabs, Synthflow and Dograh.

Dograh OSS is the voice agent itself, the open source variant you run on your own infrastructure. OPBX connects to it over a WebSocket with your endpoint and the agent's UUID. Your prompts, your conversation logic and your customer audio stay on servers you control, which matters a great deal when the audio contains identity documents and card details.

Cloudonix is the carrier layer underneath: agentic voice trunking, elastic SIP with TLS, global number coverage, and pricing that is session-based rather than per-minute.

flowchart LR
  subgraph SelfHosted["Your infrastructure"]
    OPBX["OPBX PBX"]
    Dialer["Go dialer worker"]
    AMD["AMD worker, Java 21 Vert.x 5 + ONNX"]
    Redis["Redis 7"]
    Dograh["Dograh OSS agent"]
  end
  CX["Cloudonix voice platform"]
  PSTN["Customer phones"]
  CRM["Your CRM and KYC systems"]
  Dialer --> OPBX
  AMD --> OPBX
  OPBX --> Redis
  OPBX -->|"WebSocket audio"| Dograh
  OPBX -->|"SIP trunks"| CX
  CX --> PSTN
  OPBX -->|"HMAC signed webhooks"| CRM

The dialer, AMD and the AI agent run on your servers; Cloudonix carries the calls and your systems receive signed webhooks for every outcome.

What a collection campaign looks like on this stack

A reminder campaign starts in the OPBX dialer. You load contacts, set the calling window per timezone, cap the call attempts and calls per second so carriers don't flag you, and pick a caller ID strategy. The Go dialer worker paces the calls, and the AMD worker, a Java service running ONNX models, decides within the first moments whether a human or a voicemail box picked up.

Live answers stream to your Dograh agent. The agent handles the scripted part: identity confirmation, balance, payment intent. The moment the conversation turns into a dispute or a hardship case, OPBX's configurable fallback moves the caller to an extension, a ring group or a queue, with the context preserved.

sequenceDiagram
  participant D as Dialer worker
  participant O as OPBX
  participant C as Cloudonix session
  participant P as Customer phone
  participant A as AMD worker
  participant G as Dograh agent
  participant H as Human agent
  D->>O: contact due, timezone and CPS pacing applied
  O->>C: originate outbound call
  C->>P: ring
  P-->>A: answered
  A-->>O: human detected
  O->>G: stream audio over WebSocket
  G-->>O: payment intent captured
  alt customer disputes
    G-->>O: handoff requested
    O->>H: fallback to queue or extension
  end

Answering machine detection filters voicemails, the Dograh agent handles the scripted conversation and disputes fall back to humans automatically.

Triggering a single call programmatically is one REST request against Cloudonix:

curl -X POST "https://{API_HOST}/calls/{YOUR_DOMAIN}/application" \
  -H "Authorization: Bearer {API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "destination": "+12025550142",
    "caller-id": "+12025550199",
    "application": "collections-reminder"
  }'

When this runs, Cloudonix dials the destination as if a <Dial> verb had fired, waits for the answer, then executes the voice application you named, in this case your collections reminder flow. An optional callback URL receives session status changes, which is how your CRM learns the outcome without polling.

Everything except the trunks lives on infrastructure you own. The stack is a Docker Compose deployment: Laravel and PHP for the PBX core, Go for the dialer, Java for detection, MySQL, Redis and SeaweedFS for storage. Call detail records archive across hot, warm and cold tiers for up to seven years, which covers most audit windows for regulated lending.

That seven-year trail, plus HMAC-signed webhooks and per-call audit logs, is what your compliance team will ask about first. It's there by default.

The case for staying on a bundle

Steelman time. The hosted platforms have a real argument, and it's convenience plus accountability. One invoice, one SLA, one support queue when calls degrade at 9am on a Monday. Their SOC 2, HIPAA and PCI certifications transfer to you contractually, which shortcuts months of vendor review with your risk team. And someone else's engineers absorb the 2am incident.

If you're a 30-person fintech running 20,000 minutes a month with no platform engineers, pay the $800 to $2,800 and ship. The bundle is the right answer at that scale. Open source only wins when the savings exceed the cost of the people running it.

Warning: Vendor certifications do not transfer to a self-hosted stack. The moment you run the PBX and the agent yourself, SOC 2 and PCI DSS become your audits, your controls and your evidence. Budget for that before you budget the savings.

The verdict

Adopt the open stack now if you're spending five figures a month on per-minute platform fees, you employ at least one engineer comfortable with Docker and SIP, and your compliance team prefers data residency over vendor attestations. That's a growing slice of fintech.

Wait if you're pre-product, sub-50,000 minutes or under a regulatory deadline that leaves no room to stand up infrastructure. Rent the bundle, revisit in a year.

Tip: Next two quarters: deploy OPBX and a Dograh OSS agent in a sandbox, mirror one live reminder campaign in parallel for two weeks, compare connect rates and cost per resolved call against your hosted platform. Parallel pilots kill most of the migration risk.

Where this analysis breaks

Three honest gaps. First, Cloudonix doesn't publish list prices, so I can't give you a precise session-based total cost of ownership; you need a quote to finish the math. Second, Dograh OSS is younger than the commercial agent platforms, and its ecosystem of templates and integrations is thinner. Third, my 250,000-minute scenario assumes your AI provider costs are similar across both options, which holds only if you negotiate comparable speech and LLM rates. If your hosted platform bundles voice models at genuinely better wholesale prices than you can get alone, the gap narrows.

The direction is clear either way. Per-minute platform fees are a tax on scale, and scale is the entire point of collections automation.

If you want to test the claim, the code is public: OPBX is on GitHub under an MIT license, and the Cloudonix DevZone has the trunking and AI connector guides to wire it up in an afternoon.

Frequently asked questions

How much do AI calling platforms charge for debt collection calls?

Published rates range from $0.04 per minute (Plivo) to $0.14 per minute (Bland AI), with Vapi at $0.05 plus separate telephony and AI provider costs. At 250,000 minutes a month, the platform layer alone runs $10,000 to $35,000.

Can an open source stack really handle KYC and collection workflows?

Yes. OPBX ships an auto dialer with timezone scheduling and CAC/CPS pacing, ML-based answering machine detection, caller ID pooling and configurable human fallback. Dograh OSS provides the voice agent itself, connected over WebSocket.

Who handles compliance certifications on a self-hosted stack?

You do. Platform certifications like SOC 2 and PCI DSS belong to the vendor, so self-hosting means your own organization must carry those audits. Cloudonix is GDPR and HIPAA oriented on the trunk side, but the certification burden for the full stack is yours.

What does Cloudonix charge for voice trunks?

Cloudonix uses session-based pricing rather than per-minute metering, so cost stays flat as call duration grows. AI provider costs for speech and language models are paid directly to the providers at their rates.